Speaker: Ashrujit Ghoshal [CSE Bytes Talk]

Date: Oct 28, 2026

Time: 3:30 PM – 4:30 PM

Venue: CSB-25 (Aryabhatta Hall)

Abstract:

The McEliece cryptosystem, introduced in 1978 and based on binary Goppa codes, is the oldest public-key encryption scheme still considered secure against quantum attacks. Its remarkable longevity rests in part on the apparent difficulty of distinguishing a McEliece public key from a random linear code.
In this talk, I will describe a simple classical quasipolynomial-time distinguisher for Goppa-McEliece in the asymptotic regime underlying Classic McEliece. For code length n, the algorithm runs in time n^{O(log n)} and distinguishes a McEliece public key from a uniformly random binary linear code with advantage 1-o(1). The attack is not purely asymptotic: it applies to all Classic McEliece parameter sets considered in the NIST process and gives improved, although still far from practical, concrete attack estimates.
Interestingly, the attack arose from an unsuccessful attempt to construct doubly efficient private information retrieval (PIR) schemes from algebraic locally decodable codes. I will explain the distinguisher through this PIR perspective, which gives a relatively intuitive view of the structure being exploited. Finally, I will discuss heuristic extensions of the distinguisher to quasipolynomial-time ciphertext decryption and key recovery. The key-recovery attack, in particular, remains structurally close to the distinguisher and may have implications for concrete security estimates of Classic McEliece.
This is joint work with Yuval Ishai, Aayush Jain and Nuozhou Sun.

Speaker: Bhabya Deep Rai (CS21D200) [tMeet Talk]

Date: Oct 27, 2026

Time: 4 PM – 5 PM

Venue: SSB 334 (A M Turing Hall)

Abstract:

Suppose a fixed binary word (w) of length n is hidden from us. We can query a word u and learn how many times u occurs as a subword of w. In this setting, how many queries are needed to reconstruct w? Richomme and Rosenfeld (STACS 2023) showed that O(sqrt(n log n)) queries are sufficient to exactly reconstruct any word of length n. Fleischmann, Lejeune, Manea, Nowotka, and Rigo (DLT 2020) showed that if the queries asked are restricted to be of the form 0^k 1, then min(a,b) queries are sufficient, where a and b are the number of 0s and 1s in w, respectively.

We explore the query complexity of the problem when the queries are allowed to be only Lyndon words i.e. words that are lexicographically least among all its non-trivial cyclic rotations of the word. Lyndon words are a rich and well-studied family of words that capture subword information particularly well, hence are a natural restriction for word reconstruction. In this work, we present two results:

1. We show a deterministic query algorithm that reconstructs a binary word w using at most O(k log n) Lyndon queries, where k is the number of 01-factors in w. This improves the above algorithm for an appropriate range of values of k.

2. We also show that there is a randomized query algorithm that, with high probability, reconstructs w using at most 2^{O(sqrt(log n)/epsilon)} queries, correct up to a fraction of (1-epsilon) of the entire word. This not only significantly improves upon the min(a,b) Lyndon subword query bound established by Fleischmann et al, but also outperforms the deterministic O(sqrt{n log n}) general subword counting query complexity proved by Richomme and Rosenfeld (which explicitly uses subword queries that are not Lyndon words).

The main technical ingredient we establish and use is a correspondence between subword counting queries and counting bipartite order-induced subgraph queries in a specialized class of bipartite graphs called Parikh graphs. The structural properties of the bipartite adjacency matrix of Parikh graphs enable us to yield a better upper bound in terms of k. We associate a monotone Boolean function with the word using this connection, and use the PAC learning algorithm for monotone functions by Bshouty and Tamon (JACM 1996) to approximately reconstruct our word, where we simulate queries for the Boolean function using Lyndon-subword queries.

This talk is based on joint work with Jayalal Sarma (to appear at FSTTCS 2026).

Speaker: Nishad Kothari [tMeet, aRtCS talk]

Date: Oct 23, 2026

Time: 3 PM – 5 PM

Venue: SSB 134 (Ground Floor)

Abstract:

Motivations:
The (human) pursuit of mathematics (including that of theoretical computer science) is facing a very difficult time due to the irresponsible and unethical (in my honest opinion) behaviour of certain large AI corporations (and of the individuals supporting their destruction of the aforementioned human pursuit). While individual mathematicians (including theoretical computer scientists) may hold different opinions regarding whatever is going on, one thing is certain: each of us is deeply affected, and is facing existential as well as philosophical questions (many of which we haven’t bothered to think about in the past, or at least not to this extent). It is high time that we use this as an opportunity to inculcate more solidarity (despite our many differences and varied opinions); this is the first motivation.

Secondly, it is very common amongst us to just talk about the results and the technicalities of our work, while often neglecting the journey and the more human aspects (of that journey); the latter are at least as important (if not more). Many students and postdocs have requested me to give talks about my own PhD journey; certainly I am not the only one receiving such requests. Ergo, I have been thinking about giving such a talk (or a series of them) for a while; it seems that, in light of the preceding paragraph, there can’t be a better time than now!

Finally: while there was no AI (at least in the sense that is being discussed here) during my PhD days, there was something far less powerful: sheer computational power that would allow one to exhaustively generate and check all mathematical objects up to a certain finite size. During the second half of my PhD, if I may borrow from Shakespeare (for a change), I was constantly facing the existential question: to compute, or not to compute? This is somewhat reminiscent of the burning question of our times: to AI, or not to AI?
-----------------------------------------------------------------------------
Abstract:
I was cruising during the first couple years of my PhD due to an early breakthrough while being almost oblivious to the existence of the aforementioned computational power (in the context of my research); however, soon after this, the problems (I was working on) became much harder. Around the midpoint of my PhD journey, I gave in to the temptation of using computations (after resisting their use for a long time) to verify/disprove some of my conjectures; neither did I use computational power to prove anything nor did I know/learn how to do so. However, before I could realize, I had become so dependent on computations, and addicted to them, that it adversely affected my ability to do mathematics (that is, the ability to conjecture, to prove, and most importantly, to patiently struggle in order to gain a deeper understanding through that struggle). At some point, it felt as if I was drowning. Luckily, I had excellent PhD advisors (Joseph Cheriyan and the late U. S. R. Murty) who were genuinely invested in my growth, and they stood by my side while I learnt to stay afloat. Two others (apart from my friends and family) played a key role: Marcelo H. de Carvalho and Cláudio L. Lucchesi (both from Brazil); little did they know that they were being roped into a rescue operation!
------------------------------------------------------------------------------

There will be a solidarity break around the midpoint; snacks and refreshments will be served.

Speaker: Kaushal Santosh Bhogale (CS22D006)

Date: Oct 22, 2026

Time: 3 PM – 4 PM

Venue: SSB 334 (A M Turing Hall)

Abstract:

Early ASR systems for Indian languages were constrained by limited training data, restricting evaluation to one or two academic benchmarks and making generalizable claims of improvement difficult to establish. As these systems moved into deployment and were subjected to diverse, uncontrolled acoustic conditions, single-benchmark evaluation proved insufficient to characterize model performance, motivating a systematic collation of existing benchmarks. We present Vistaar, a collation of 59 benchmarks spanning 12 Indian languages and diverse domains, on which we evaluate five ASR systems and observe substantial, inconsistent variance across benchmarks and models. We further fine-tune Whisper on 10,700 hours of aggregated training data to build IndicWhisper, which attains the lowest WER on 39 of the 59 benchmarks, establishing a broad, overfitting-resistant evaluation suite.
As ASR systems improved, a deeper flaw emerged in the evaluation metric itself: WER computed against a single reference transcript penalizes permissible spelling variation and non-standardized code-mixed words, inflating error rates beyond what is perceptually meaningful. We address this by proposing a new framework, OIWER, that leverages large language models to generate exhaustive per-word spelling variants and incorporates them into scoring. OIWER reduces pessimistic error rates by 6.3 points on average, narrows artificially inflated performance gaps between models, and aligns 4.9 points more closely with human judgment than prior normalization-based methods.
With an orthographically fair metric established, we construct two complementary benchmarks to characterize the state of Indian language ASR under realistic conditions and guide model selection for deployment. Voice of India (VOI) comprises 536 hours of unscripted telephonic speech across 15 languages and 139 regional clusters, with orthographically-aware transcripts enabling analysis across geography, audio quality, speaking rate, and demographic factors. Vimarsha extends this to all 22 scheduled languages over 100 hours, combining demographically diverse on-field recordings with acoustically difficult in-the-wild audio to correct the optimistic bias of clean benchmarks, and a lattice of variations framework encoding multiple valid transcriptions per utterance to correct the pessimistic bias of rigid scoring. Evaluated jointly, VOI and Vimarsha reveal substantial shifts in model rankings and systematic geographic and demographic failure modes obscured by prior evaluation practices, providing a comprehensive account of the current state of Indian language ASR.
Together, these contributions establish a benchmark and metric ecosystem for Indian language ASR that is simultaneously broad in coverage, acoustically realistic, and orthographically faithful, providing a reliable foundation for tracking and driving further progress in the field.

Speaker: Prakhar Krishna Kumar (CS19D002)

Date: Oct 22, 2026

Time: 11 AM – 12 PM

Venue: Online GMeet

Abstract:

Graphs provide a natural way to represent complex real-world systems through entities and their relationships. The structured format of graphs is intuitive for humans and is computationally efficient for machines, making them well suited to applications such as social networks, drug interactions, and molecular systems. In this work, we study the problems of node classification and graph classification, with a focus on their practical challenges such as limited supervision, extreme label sparsity, unknown homophily or heterophily, neural network’s inductive bias, and architecture selection.

Our first focus is on large-scale node classification, where the labels data is very sparsely available. While the classical label-propagation methods typically rely on a compatibility matrix to capture the relationships between classes, this information is generally unavailable and is supplied by domain experts. We propose distant compatibility estimation, a method to learn these compatibilities directly from very sparsely labeled graphs, without requiring any substantial increase in computation than the subsequent label propagation.

We then examine the representational power of graph neural networks and practical learnability. While theoretical analysis is often related to GNN’s expressiveness in terms of the Weisfeiler Lehman test, such characterization does not capture the optimization behavior properly. We study the inductive bias of graph neural networks and show that the global pooling can influence if a GNN model favors discriminative subgraphs or more dispersed informative nodes throughout the graph. This provides us a useful perspective to incorporate domain knowledge by GNN architecture selection.

Finally, we use these graph-representation principles to the problem of coronary artery disease detection from CT images. We develop a graph-based method that uses GNNs to model spatial relationships in coronary vessel anatomy and address the limitations of conventional machine-learning and convolutional approaches.

Broadly, our work advances the utility of label propagation and deep neural models for learning over graph-structured data via practical and scalable methods for relational learning.

Date: Oct 19, 2026

Time: 3 PM – 4 PM

Venue: CS Office

Abstract:

Saraswathi/Ayudha Pooja celebration at CS Office - Kindly make it convenient to attend

Speaker: Nishad Kothari [tMeet, aRtCS talk]

Date: Oct 16, 2026

Time: 3 PM – 5 PM

Venue: SSB 134 (Ground Floor)

Abstract:

Dedicated to the human spirit in these dark times!
-----------------------------------------------------------------------------
Motivations:
The (human) pursuit of mathematics (including that of theoretical computer science) is facing a very difficult time due to the irresponsible and unethical (in my honest opinion) behaviour of certain large AI corporations (and of the individuals supporting their destruction of the aforementioned human pursuit). While individual mathematicians (including theoretical computer scientists) may hold different opinions regarding whatever is going on, one thing is certain: each of us is deeply affected, and is facing existential as well as philosophical questions (many of which we haven’t bothered to think about in the past, or at least not to this extent). It is high time that we use this as an opportunity to inculcate more solidarity (despite our many differences and varied opinions); this is the first motivation.

Secondly, it is very common amongst us to just talk about the results and the technicalities of our work, while often neglecting the journey and the more human aspects (of that journey); the latter are at least as important (if not more). Many students and postdocs have requested me to give talks about my own PhD journey; certainly I am not the only one receiving such requests. Ergo, I have been thinking about giving such a talk (or a series of them) for a while; it seems that, in light of the preceding paragraph, there can’t be a better time than now!

Finally: while there was no AI (at least in the sense that is being discussed here) during my PhD days, there was something far less powerful: sheer computational power that would allow one to exhaustively generate and check all mathematical objects up to a certain finite size. During the second half of my PhD, if I may borrow from Shakespeare (for a change), I was constantly facing the existential question: to compute, or not to compute? This is somewhat reminiscent of the burning question of our times: to AI, or not to AI?
-----------------------------------------------------------------------------
Abstract:
I was cruising during the first couple years of my PhD due to an early breakthrough while being almost oblivious to the existence of the aforementioned computational power (in the context of my research); however, soon after this, the problems (I was working on) became much harder. Around the midpoint of my PhD journey, I gave in to the temptation of using computations (after resisting their use for a long time) to verify/disprove some of my conjectures; neither did I use computational power to prove anything nor did I know/learn how to do so. However, before I could realize, I had become so dependent on computations, and addicted to them, that it adversely affected my ability to do mathematics (that is, the ability to conjecture, to prove, and most importantly, to patiently struggle in order to gain a deeper understanding through that struggle). At some point, it felt as if I was drowning. Luckily, I had excellent PhD advisors (Joseph Cheriyan and the late U. S. R. Murty) who were genuinely invested in my growth, and they stood by my side while I learnt to stay afloat. Two others (apart from my friends and family) played a key role: Marcelo H. de Carvalho and Cláudio L. Lucchesi (both from Brazil); little did they know that they were being roped into a rescue operation!
------------------------------------------------------------------------------

There will be a solidarity break around the midpoint; snacks and refreshments will be served.

Speaker: Ramesh S, Senior Tech Fellow, General Motors R&D, Warren, MI USA

Date: Oct 16, 2026

Time: 11:30 AM – 12:30 PM

Venue: SSB 334 (A M Turing Hall)

Abstract:

The talk will highlight the recent and emerging trends in the use of AI in automotive features and their development. AI is an important enabler for automated driving features and systems. Besides being part of the end system, they also support the development of these systems. Automotive safety is a well-developed area involving a variety of elaborate processes and methods, e.g., Hazard Analysis and Risk Assessments (HARA), Fault Tree Analysis (FTA), and there are several ISO standards (ISO 26262, ISO 21448) of the development of safe automotive system. AI systems, being probabilistic and data-dependent, bring in new challenges to the safety of automotive systems. New standards and guidelines are emerging, and the talk will focus on some of these as well.

Bio: Dr. Ramesh S, Senior Technical Fellow, has been with General Motors R&D for more than 20 years conducting and leading advanced research projects in the areas of model-based development of embedded systems and software, rigorous verification & validation and more recently AI/ML based systems. Prior to joining GM R&D, he was a full professor in the department of Computer Science & Engineering at Indian Institute of Technology Bombay, India where he co-founded a Centre for Formal Design and Verification of Software. He has published several research papers in International Journals & Conferences, and author of
more than 50 patents in the areas of modeling, analysis and verification of embedded systems and software. He has been on the program committees of several international research conferences and on the editorial boards of journals. He led an USCAR committee on guidelines for V&V of AI/ML components and is serving as an expert in ISO and SAE committees for developing guidelines for AI/ML based systems.

Speaker: Anuja Modi (CS21D405)

Date: Oct 14, 2026

Time: 11:30 AM – 12:30 PM

Venue: SSB 334 (A M Turing Hall)

Abstract:

Constructing Indistinguishability Obfuscation (iO) from simple, safe assumptions in the post-quantum regime is an important problem with wide-ranging applications. In this work, we make progress towards this goal. We formulate a new, arguably simple assumption called LWE with Structured Preimage that suffices to build iO. Our assumption is instance-independent, falsifiable and non-interactive. Our assumption posits that a set of LWE (Learning With Errors) samples is indistinguishable from random given a short Gaussian matrix which is carefully chosen to satisfy the required functionality while avoiding problematic leakage.

We provide a thorough cryptanalysis of our assumption, using the extensive cryptanalytic effort against the related evasive LWE assumption (Wee, Eurocrypt 2022 and Tsabary, Crypto 2022) to gain confidence. We emphasize that our assumption is superficially similar but fundamentally different from evasive LWE. Thus, we obtain iO from a simple lattice assumption whose hardness is somewhat better understood as compared to brand new assumptions.

Speaker: Vaibhav Krishan, PostDoc IITM [tMeet talk]

Date: Oct 13, 2026

Time: 4 PM – 5 PM

Venue: CS 36 (CSB)

Abstract:

Abstract: Quantified Boolean Formulas (QBFs) can be viewed as two-player games: the players alternately assign Boolean values to variables, with one player trying to satisfy a given propositional formula and the other trying to falsify it. The QBF problem asks whether the first player has a winning strategy. In this talk, we ask a more quantitative question: how many winning strategies are there?

This question introduces challenges that do not arise in ordinary solution counting. A strategy may depend on the opponent's previous moves, the space of strategies can be enormous, and the number of winning strategies itself can be doubly exponential in the size of the formula. At the same time, QBF and its variants provide a natural language for reasoning about adversarial choices and unknown environments, with applications including formal verification, planning, reactive synthesis, and model checking.

I will discuss both theoretical and practical approaches to counting winning strategies in QBFs. I will describe a solver we have developed that exploits the structure of the underlying strategy space to avoid explicitly exploring or representing it in full. I will then discuss the theoretical foundations behind these techniques and how they suggest stronger methods for future solvers. More broadly, the talk will illustrate an interplay between theory and practice: theoretical structure leads to practical algorithms, while computational considerations motivate new theoretical questions.

This talk is based on joint work with Sravanthi Chede and Anil Shukla (FSTTCS 2026, to appear), and with Sravanthi Chede, Leroy Chew, and Anil Shukla (SAT 2026).
-----------------------
Speaker Bio: Vaibhav is a postdoctoral researcher in the CSE department at IIT Madras. Previously, he was a postdoctoral fellow in the TCS group at IMSc, Chennai. He received his PhD in CSE from IIT Bombay, where he was advised by Prof. Nutan Limaye and Prof. Sundar Vishwanathan.

His research interests lie broadly in complexity theory and automated reasoning, with a focus on understanding computational hardness from both theoretical and practical perspectives. Apart from his academic career, he has spent more than five years in industry as a quantitative trader.

Date: Oct 09, 2026

Time: 5 PM – 5:30 PM

Venue: SSB 133

Abstract:

Snack place in SSB ground floor (behind SSB 133)

Speaker: Akshay Kumar E (CS22D014)

Date: Oct 09, 2026

Time: 4 PM – 5 PM

Venue: SSB 334 (A M Turing Hall)

Abstract:

Developing cryptographic implementations that are resistant to side-channel attacks is challenging in modern CPUs. Data-dependent switching activity induces subtle transitions deep in the microarchitecture that can cause sensitive information to leak through the CPU's power consumption. Existing leakage detection tools largely rely on simulation-based testing, which is not only time-consuming but is also limited to the explored execution scenarios.

We present microSCALE, a fully automated framework for detecting power side-channel leakages that occur due to transitions in the CPU's microarchitecture. By statically analyzing the CPU's RTL, microSCALE enables systematic exploration of microarchitectural leakage sources within seconds. It precisely localizes the origin of leakage in the hardware and traces it back to the corresponding software instructions. We evaluate microSCALE on several protected crypto-implementations, including GIFT, ASCON, PRESENT, and AES, running on an open-source RISC-V processor, and identify several leakage sources. For instance, we identify that a protected implementation of the GIFT cipher has about 80% of its instructions that leak on the Rocket Core. We show how microSCALE can help in hardening this implementation, reducing the side-channel leakage around 90.3%.

Speaker: Nishad Kothari [tMeet, aRtCS talk]

Date: Oct 09, 2026

Time: 3 PM – 5 PM

Venue: SSB 134 (Ground Floor)

Abstract:

Motivations:
The (human) pursuit of mathematics (including that of theoretical computer science) is facing a very difficult time due to the irresponsible and unethical (in my honest opinion) behaviour of certain large AI corporations (and of the individuals supporting their destruction of the aforementioned human pursuit). While individual mathematicians (including theoretical computer scientists) may hold different opinions regarding whatever is going on, one thing is certain: each of us is deeply affected, and is facing existential as well as philosophical questions (many of which we haven’t bothered to think about in the past, or at least not to this extent). It is high time that we use this as an opportunity to inculcate more solidarity (despite our many differences and varied opinions); this is the first motivation.

Secondly, it is very common amongst us to just talk about the results and the technicalities of our work, while often neglecting the journey and the more human aspects (of that journey); the latter are at least as important (if not more). Many students and postdocs have requested me to give talks about my own PhD journey; certainly I am not the only one receiving such requests. Ergo, I have been thinking about giving such a talk (or a series of them) for a while; it seems that, in light of the preceding paragraph, there can’t be a better time than now!

Finally: while there was no AI (at least in the sense that is being discussed here) during my PhD days, there was something far less powerful: sheer computational power that would allow one to exhaustively generate and check all mathematical objects up to a certain finite size. During the second half of my PhD, if I may borrow from Shakespeare (for a change), I was constantly facing the existential question: to compute, or not to compute? This is somewhat reminiscent of the burning question of our times: to AI, or not to AI?
-----------------------------------------------------------------------------
Abstract:
I was cruising during the first couple years of my PhD due to an early breakthrough while being almost oblivious to the existence of the aforementioned computational power (in the context of my research); however, soon after this, the problems (I was working on) became much harder. Around the midpoint of my PhD journey, I gave in to the temptation of using computations (after resisting their use for a long time) to verify/disprove some of my conjectures; neither did I use computational power to prove anything nor did I know/learn how to do so. However, before I could realize, I had become so dependent on computations, and addicted to them, that it adversely affected my ability to do mathematics (that is, the ability to conjecture, to prove, and most importantly, to patiently struggle in order to gain a deeper understanding through that struggle). At some point, it felt as if I was drowning. Luckily, I had excellent PhD advisors (Joseph Cheriyan and the late U. S. R. Murty) who were genuinely invested in my growth, and they stood by my side while I learnt to stay afloat. Two others (apart from my friends and family) played a key role: Marcelo H. de Carvalho and Cláudio L. Lucchesi (both from Brazil); little did they know that they were being roped into a rescue operation!
------------------------------------------------------------------------------

There will be a solidarity break around the midpoint; snacks and refreshments will be served.

Speaker: Prof Aalok Thakkar (Ashoka University) & TVH Prathamesh (Krea U)

Date: Oct 09, 2026

Time: 2 PM – 3 PM

Venue: MR1 (SSB 233, First Floor)

Abstract:

This talk introduces proof assistants and surveys their use in mathematics, theoretical computer science, and formal methods. We compare Rocq, Lean, and Isabelle, with particular attention to the type systems of Rocq and Lean and their implications for formalization. We also explore how theoretical computer scientists can use AI alongside proof assistants in their research, discussing the opportunities, limitations, and potential pitfalls of AI-assisted formalization.
As a case study, we present DiagonaLean, our project on formalizing undecidability proofs in Lean. We describe the project’s goals and the results formalized, discuss key design choices and differences from formalization in Rocq, and reflect on our experience with an AI-assisted workflow. In particular, we examine which aspects of the formalization process AI changed and which continued to require substantial human guidance and mathematical judgment.
The talk is intended as an introduction to proof assistants and their role in research, rather than a tutorial on Rocq or Lean.

Speaker: Ingale Kranti Bhimrao (CS22D003)

Date: Oct 08, 2026

Time: 12 PM – 1 PM

Venue: SSB 334 (A M Turing Hall)

Abstract:

The shared L2 translation lookaside buffer (L2-TLB) is the convergence point of modern GPU address translation. All compute units (CUs) on a device route their L1-TLB misses to this single structure, and under single-instruction, multiple-thread execution, tens of thousands of threads issue translation requests concurrently. The resulting traffic contends at the L2-TLB and stalls the CUs waiting on a pending trans- lation. Reducing the L1-to-L2 traffic that drives this contention is a direct way to recover GPU translation throughput. Much of this traffic is avoidable. Workgroups of a GPU kernel are distributed across neighbor-ing CUs and translate overlapping virtual pages within a short temporal window. Each L1-TLB’s MSHRs merge in-flight duplicates only within the same CU; they do not coordinate with peer L1-TLBs, so iden-tical requests from neighboring CUs arrive at the L2-TLB independently. Per-walk locality compounds this: the page-walker fetches a 64-byte cache line containing eight contiguous page-table entries (PTEs), yet conventional L1-TLBs install only the requested PTE and discard the rest, so each discarded sibling may become a future L1-to-L2 request for a translation the walker had already produced. The two effects produce the same symptom: redundant L1-to-L2 traffic that drives L2-TLB contention. The state-of-the-art Intra-Group Sharing (IGS) design is a de-duplication mechanism: it consolidates per-CU L1-TLBs into a single shared L1-TLB per group, accessed through a 4×4 crossbar, and eliminates duplicate entries within the group. However, nearly every CU request now traverses the crossbar, and IGS’s only filter against this traffic is per-CU coalescing of concurrent in-flight requests, which removes only 32.4 % of it in our simu-lations. We propose PATCH, a Persistent And Translation-Coalescing Hierarchy, which redistributes the same number of L1-TLB entries into a per-CU translation filter and a coalesced shared L1-TLB. A small per-CU Persistent Translation Filter exploits temporal reuse: it filters out a CU’s repeated requests, keep-ing them off the crossbar. The shared L1-TLB holds one copy of each translation for the shader array and, adopting CoLT-style coalescing, keeps the contiguous translations among the eight a walk returns; an age-weighted replacement policy protects densely packed entries from being evicted before sparser ones. PATCH stays entirely on-die and changes no page-table format, allocator, or page-table walker. Our evalua- tion using 19 GPGPU benchmarks shows that PATCH improves instructions per cycle by 1.14× (1.30× on the translation-sensitive benchmarks) and lowers the L1-TLB miss rate by 32 % over the baseline. It keeps 69.1 % of CU-side translation requests off the crossbar and reduces dynamic translation energy by 42 %
relative to IGS.

Speaker: Prof. Venkatesh Raman (IMSc)

Date: Oct 06, 2026

Time: 4 PM – 5 PM

Venue: SSB 334 (A M Turing Hall)

Abstract:

We trace the possible history of fine-grained complexity starting with a nice algorithm, due to Williams, beating 2^n for Max2SAT.
We end with the new kid on the block, NSETH (non-deterministic SETH) and discuss its consequences for multiple hypotheses in fine-grained complexity. En route, we will see an interesting sub-quadratic non-deterministic algorithm for the complement of 3-SUM.

Speaker Profile:
Prof. Venkatesh Raman retired recently as Professor from the Institute of Mathematical Sciences, Chennai. He is a well known researcher in the areas of Data Structures and Parameterized Complexity . He has several foundational and landmark results in Succinct Data Structures and Parameterized Algorithms to his name.

Speaker: Prof Aalok Thakkar (Ashoka University)

Date: Oct 06, 2026

Time: 4 PM – 6:30 PM

Venue: MR1 (SSB 233, First Floor)

Abstract:

The purpose of this short course is to convince you that different functional programming languages cultivate different models of thought. We will explore two distinct perspectives on computation through Haskell and OCaml.

The two languages make significantly different choices about how programs are defined, evaluated, and structured. By studying them in contrast, we will see how these choices influence the way we reason about recursion, infinite structures, abstraction, and state. We will use concrete examples to connect programming practice with the underlying ideas of semantics, giving us a clearer view of why these languages encourage different styles of reasoning.


We will assume some basic familiarity with OCaml. We will start with a short introduction to Haskell syntax and write some simple programs. Please install Haskell GHC or use an online compiler

Speaker: Prof Aalok Thakkar (Ashoka University)

Date: Oct 05, 2026

Time: 12 PM – 1 PM

Venue: MR1 (SSB 233, First Floor)

Abstract:

Algebraic datatypes form a semiring under sums and products, and a recursive type is a solution of a polynomial equation. Taken literally, this lets one symbolically differentiate an algebraic datatype, and it turns out that the derivative of a datatype is itself, with a hole in it.

This talk explains the two sentences above. We begin with definitions of sum, product, and recursive types (least fixed points). We then differentiate the formal expression for lists to recover Huet's zipper, a data structure for moving through a list and editing it locally in constant time. Applied to binary trees, the same rules yield the tree zipper.

Much of this talk, including its title, borrows from McBride's 2008 functional pearl, which refines the derivative into a dissection. Midway through a left-to-right traversal, the elements to the left of the current position have been processed (the clowns) and those to the right have not (the jokers). Giving the two sides different types yields a two-variable analogue of the derivative that obeys the rules of Newton's divided differences. From it one derives, for every polynomial datatype, a tail-recursive fold (catamorphism). The algebra of datatypes thus describes not only the shape of data, but the course of a computation over it. This talk will be broadly accessible; familiarity with functional programming would be nice.

Date: Oct 01, 2026

Time: 4:30 PM – 6 PM

Venue: CS 36 (CSB)

Abstract:

Please see the link for more details.

Looking forward to seeing you all there!
Best wishes,
Nishad (on behalf of Chennai Combinatorics Network)

Speaker: Manikandan Narayanan [CSE Bytes Talk]

Date: Sep 23, 2026

Time: 3:30 PM – 4:30 PM

Venue: CSB-25 (Aryabhatta Hall)

Abstract:

In the 1990s, when the Human Genome Project was in its early years, eminent computer scientist Donald Knuth commented in an interview that "Biology easily has 500 years of exciting problems to work on, it's at that level". The incredible opportunity for computer scientists today is that many of these open biological questions can actually be tackled using the power of computation and the wealth of data generated by modern biology. In this talk, I will pose some biological questions around the reconstruction and analysis of molecular networks connecting genetic factors, genes, or clinical outcomes such as disease states; and walk you through how these questions can be formulated as precise computational problems. These problems at the interface of computer and life sciences, and our approaches to solving them using multilayer/ensemble graph algorithms and transfer/causal learning models, are what drive our BIRDS (Bioinformatics and Integrative Data Science) group. We hope to share our enthusiasm for these problems in computational biology, the challenges we've faced when solving them, and the opportunities that lie ahead. In the spirit of CSE Bytes, we will pose more open problems than solutions, with the hope of spurring those of you who are interested to engage further with these problems.

Speaker: Narayanan aka Naru

Date: Sep 22, 2026

Time: 3:30 PM – 4:30 PM

Venue: CS 36 (CSB)

Abstract:

List-colouring, introduced independently in the 1970s by Vizing and by Erdős, Rubin, and Taylor, generalizes ordinary vertex colouring by assigning each vertex a personalized palette of admissible colours. A recurring theme in this field is chromatic-choosability, the condition where a graph's list chromatic number equals its ordinary chromatic number. This talk provides a survey of the progress made in identifying chromatic-choosable graph classes and characterizing the ‘jump’ — the gap that can exist between these two parameters.

We delve into the principal methods that have shaped the field, providing technical insights into the Alon–Tarsi method, which utilizes graph polynomials and orientations to establish upper bounds; the Kernel method, famously applied by Galvin to settle the List Edge Colouring Conjecture for bipartite multigraphs; and Thomassen’s rigid boundary induction, which proved the 5-choosability of planar graphs among others.

Based on a survey co-authored with Somasundaram and Nandana Vasudevan.

Speaker: Yamini Shankar (CS22D002)

Date: Sep 11, 2026

Time: 5 PM – 6 PM

Venue: MR2 - CSB 21 - First Floor

Abstract:

In this thesis, we investigate the practical deployment challenges of WiFi sensing systems, which use ambient Channel State Information (CSI) from existing wireless infrastructure to sense occupancy, activity, and location without dedicated hardware. While prior work has shown CSI sensing works well in controlled settings, we focus on three issues that determine whether it survives contact with the real world: communication scalability, energy efficiency, and privacy.

With WiFi sensing networks scaling to tens or hundreds of nodes, each continuously streaming high-rate CSI, the shared wireless channel becomes badly congested. The key challenge we address here is, “How much of this CSI data is actually necessary, and can we transmit only that?” We develop AutoCompress, a system that selects only informative sensors and compresses CSI spectrograms using importance-guided compression (SSPoC), cutting bandwidth use by up to 500× while minimal impact on activity-recognition accuracy, and improving network throughput by 35.5% with 98–99% lower latency than uncompressed streaming.

Localization is one of CSI sensing’s most useful applications, however, on battery-powered and embedded devices, repeated WiFi scanning for accurate localization drains the battery fast. The key challenge we address here is, “How many scans and which access points are actually needed for good localization, and can this be decided automatically?” We build LowFI, a system that adapts scan count and access point selection to local channel conditions; across multiple device platforms, LowFI cuts non-line-of-sight localization error from ≈8 m to ≈2 m using just 5–7 scans, at roughly a tenth of the energy cost of CSI-fusion alternatives.

Because CSI is exposed at the physical layer, it remains a side channel for surveillance even when the communication itself is encrypted. The key challenge we address here is, “Can we make CSI useless to an eavesdropper without making it useless for communication?” We develop two defenses, NonSense and SSCoP, that perturb only the most sensing-relevant, SNR-constrained portions of the CSI, degrading an adversary’s ability to localize, detect occupancy, recognize activity, and infer gestures, while preserving link quality for legitimate communication.

Building on the sensing, efficiency, and privacy techniques developed in the thesis, we extend these ideas to a new application domain with WiProctor, a multimodal exam-proctoring system that fuses WiFi CSI with video and device-interaction logs to catch covert collaboration happening outside a webcam’s field of view. Across 600 sessions from 20 participants, WiProctor reaches an 88% F1-score and catches 94% of covert-collaboration cases, twice the detection rate of human annotators relying on camera footage alone, while still producing time-aligned evidence a human proctor can review before taking action.

Date: Sep 04, 2026

Time: 5 PM – 5:30 PM

Venue: SSB 133

Abstract:

Snack place in SSB ground floor (behind SSB-133)

Speaker: P Girinath (CS22S021)

Date: Sep 03, 2026

Time: 3 PM – 4 PM

Venue: SSB 334 (A M Turing Hall)

Abstract:

Branch predictor accuracy is a critical determinant of performance in modern out-of-order CPU cores. A subset of branches that exhibit strong data dependence and are inherently difficult to predict—commonly referred to as hard-to-predict (or H2P) branches pose significant performance challenges. Several mechanisms have been proposed to address these H2P branches, including Auto-Predication of Critical Branches (ACB), Control Flow Decoupling (CFD), and Runahead execution. Although each of the above techniques excels for particular classes of branches and carries distinct energy costs, no single solution effectively addresses the full spectrum of H2P behavior.

Our work presents a unified characterization methodology that analyzes retired micro-operation stream to extract dynamic data-flow and control-flow properties of individual branches. We show that the effectiveness of H2P branch mitigation techniques is governed by a fundamental tradeoff between speculative benefit (e.g., latency hiding and misprediction reduction) and resource cost (e.g., execution bandwidth and buffer pressure). Based on this analysis, H2P-BRAIN derives branch-level deployment policies that are embedded as executable annotations, enabling processors to selectively apply H2P mitigation with negligible runtime overhead. Across the SPEC CPU2017 Integer workloads, this approach eliminates performance inversions and improves overall Instructions Per Cycle (IPC) for all three techniques, increasing gains from 2.01% to 2.74% for ACB, 6.33% to 6.57% for CFD, and 8.7% to 10.25% for Runahead. The profitability criteria are derived using SPEC CPU2017 Integer workloads and subsequently applied without retraining to diverse client, server, HPC, Geekbench and floating-point workloads, demonstrating that the identified branch-level characteristics generalize beyond the workloads used for selecting the profitability criteria.

Speaker: Prof. Vincent Tan (NUS)

Date: Aug 28, 2026

Time: 3:30 PM – 4:30 PM

Venue: SSB 334 (A M Turing Hall)

Abstract:

We study the Bayesian fixed-budget best-arm identification problem in which a learner can abstain from making a terminal recommendation. Subject to an abstention budget α, we analyze the probability of undetected error—the risk of recommending a suboptimal arm without abstaining. Our central finding is that abstention induces a phase transition: without abstention, the error probability decays polynomially in the sampling budget T; in contrast, introducing any small positive abstention budget shifts this to an exponential decay. For Gaussian priors and rewards, in the regime T → ∞ followed by α ↓ 0, we establish exact matching information-theoretic lower bounds and algorithmic upper bounds on the optimal error exponent, which takes the form exp(−α²T/(8κ²ν)). The hardness parameter κν represents the prior density of the top-two gap at zero, highlighting that nearly tied instances drive the fundamental error. We introduce an adaptive algorithm, PGWS, that successfully achieves this optimal exponent by expending its abstention budget on statistically ambiguous instances. We further demonstrate that this polynomial-to-exponential improvement is exclusively a Bayesian phenomenon—in the frequentist setting, abstention only affects lower-order exponent terms. We also extend our results beyond the Gaussian model.

Bio:
Vincent Y. F. Tan is a Professor in the Departments of Mathematics and Electrical and Computer Engineering at the National University of Singapore (NUS). His research lies at the intersection of information theory, machine learning, and sequential decision making, with current interests in reinforcement learning, large language models, and statistical learning. He received his Ph.D. from MIT and B.A. and M.Eng. degrees from the University of Cambridge, and has received several awards for excellence in both research and teaching.

Speaker: Dr. Rajani Kant (Padma Shri Awardee)

Date: Aug 25, 2026

Time: 6 PM – 7 PM

Venue: SSB 334 (A M Turing Hall)

Abstract:

Dr. Rajani Kant brings with him 36 years of work in the social development sector, having worked extensively with artisans, farmers, and women across the country. Through his pioneering work with Geographical Indications, he has helped protect not just products, but the livelihoods and identity of entire artisan communities.

This talk is organized by the Artisan Empowerment Lab (AEL), whose ongoing work on GI-tagged craft authentication and traceability draws directly from the kind of ground-level insight Dr. Rajani Kant will be sharing.

It promises to be a wonderful and thought-provoking evening, and I'd love to see many of you there — faculty, staff, students, residents and even those outside the IITM community. The issues discussed are going to be of general interest.

Speaker: Dr Sandhya Saisubramanian (Oregon State University, USA)

Date: Aug 21, 2026

Time: 11 AM – 12 PM

Venue: SSB 334 (A M Turing Hall)

Abstract:

Autonomous agents are expected to operate reliably in open-world environments, involving coordination with other agents and reasoning under imperfect models. In this talk, I will present planning and learning approaches that enable reliable decision making in single and multi-agent settings. The first part of the talk will focus on how individual agents can learn safe, aligned behaviors from human feedback. The second part of the talk will focus on multi-agent coordination strategies under communication constraints and underspecified objectives, to enable safe joint behavior. Together, these directions make progress towards enabling autonomous agents to act safely and reliably under uncertainty. I will conclude the talk with a brief overview of open challenges.

Speaker Bio:
Sandhya Saisubramanian is an Assistant Professor at Oregon State University. She earned her PhD from the University of Massachusetts Amherst. Her research group focuses on developing techniques for reliable autonomy in single and multi-agent systems that operate in fully and partially observable environments. She served as the Program co-chair for ICAPS 2026 and is a recipient of Outstanding Paper Award at RLC 2025, Distinguished Paper Award at IJCAI 2020, and Outstanding Program Committee Member Award at ICAPS 2022.

Speaker: Nitish Yadav (CS23D011)

Date: Aug 19, 2026

Time: 11 AM – 12 PM

Venue: SSB 334 (A M Turing Hall)

Abstract:

Distributed systems use replicated objects to improve fault tolerance for which their key correctness criterion is Strong Eventual Consistency (SEC). SEC ensures that replicas which have seen the same set of updates are in the same state. Operation-based Conflict-Free Replicated Data Types (CRDTs) provide a principled means to achieve SEC by broadcasting update functions, called effectors, and applying them at replicas in an order consistent with causal history. However, buggy CRDT implementations may generate concurrent effectors that do not commute, causing SEC violations. Model checking can detect such violations by exhaustively exploring reachable configurations, but suffers from State-Space Explosion (SSE). In distributed systems setting this is even severe, owing to large number of replicas, updates, delivery choices, and concurrent-effector application orders.

Existing Partial Order Reduction (POR) techniques reduce SSE using Mazurkiewicz Trace Equivalence (MTE), driven by persistent sets and sleep sets. However, we discover that in the CRDT setting, they struggle with certain challenges: (a) MTE is too fine-grained for op-based CRDTs because two non-Mazurkiewicz-equivalent traces may still reach the same configuration, or equivalent configurations modulo timestamps and replica names, eventually leading to repeated SEC checks. (b) Persistent-set choices can further delay SEC bug discovery by postponing transitions generating non-commuting effectors, resulting in discovery of SEC violation at later steps.

To cater to the above challenges, we propose a model checker based on configuration equivalence, unlike MTE, that also incorporates an Eager SEC detection mechanism that pro-actively checks newly generated effectors against available concurrent effectors for SEC violation. The checker is k-bounded sound, i.e., if an SEC bug can be detected within k steps in the full state space, it detects the bug in at most k steps, and often earlier through eager detection. It is also sound and optimal with respect to configuration equivalence, i.e., it explores exactly one representative configuration per equivalence class. Empirical results reveal that the proposed model checker significantly outperforms existing state-of-the-art model checkers: it explores far fewer configurations, scales to higher bounds, takes less time, and can detect buggy implementations at lower bounds.

Speaker: Krishna Nandivada V [CSE Bytes Talk]

Date: Aug 17, 2026

Time: 3:30 PM – 4:30 PM

Venue: CSB-25 (Aryabhatta Hall)

Abstract:

We will try to address some of the popular myths about programs written in Java: "Java programs are comparatively slower”, “Java programs have slow startup times”, “Garbage collection ensures that there are no memory leaks”, "Java programs are useless for Artificial Intelligence and Machine Learning”. We will follow two contradictory approaches in this talk: for each of the myths, we will explain the basis of the myths and also try to throw some light on the reality and some open problems in this space.

I will talk for around 30 minutes max. The main part of the session will be interaction with the audience on some of these topics.

Date: Aug 07, 2026

Time: 5 PM – 5:30 PM

Venue: SSB 133

Abstract:

What: CSE Words
When: August 7, 2026, 5pm -5.30pm.
Venue: Snack place in SSB ground floor (behind SSB 133).

Speaker: Praseetha M (CS19D008)

Date: Aug 07, 2026

Time: 3 PM – 4 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Dynamic Random-Access Memory (DRAM) is the most widely used main memory in modern computing systems. Each bit of data is stored in tiny cells arranged in a grid of rows and columns in the DRAM. As technology scales down, cells become denser and more tightly packed, leading to electromagnetic interference with the normal operation of cells. This interference impacts the reliability, integrity, and confidentiality of the data stored in DRAM, giving rise to a new class of attacks called RowHammer. RowHammer attack demonstrates that repeatedly accessing (activating) a DRAM row can induce bit flips in the neighboring rows, enabling attackers to tamper with or access data belonging to other users. It is a prime example of how a hardware-level vulnerability can lead to software-level exploits, making it one of the most serious security threats identified by researchers. Although numerous mitigation strategies have been proposed in the literature, their widespread adoption in commercial DRAM remains a challenge due to associated costs, performance and energy overheads.

In order to mitigate RowHammer attack, state-of-the-art methods use a counter-based approach for identifying the frequently accessed rows using an array of counters (tracker). When the frequency of a row access crosses the RowHammer Threshold (The maximum number of activations a DRAM row can undergo without causing any bit flip in the nearby rows), it is considered as a hammering activity and the mitigation comes into action. But the main practical difficulty in any counter-based mitigation strategy is the increasing overhead of the tracker with the decreasing trend in RowHammer threshold over the years. We propose to mitigate RowHammer attack without using any tracker by exploiting the physical isolation between DRAM subarrays. The evaluations show that our technique performs better and consumes less power at DRAM compared to state-of-the-art mitigation techniques.

Speaker: Gautam Chandrasekaran, UT Austin, USA

Date: Aug 04, 2026

Time: 3:30 PM – 4:30 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Sparse linear regression is one of the most basic questions in machine learning and statistics. Here, we are given as input a design matrix X∈ℝ^{N×d} and measurements or labels y∈ℝ^N where y=Xw+ξ, and ξ is the noise in the measurements. Importantly, we have the additional constraint that the unknown signal vector w is sparse: it has k non-zero entries where k is much smaller than the ambient dimension. Our goal is to output a prediction vector u that has a small prediction error: 1/N⋅‖Xw−Xu‖^2.
Information-theoretically, we know what is best possible in terms of measurements: under most natural noise distributions, we can get prediction error at most ϵ with roughly N=O(klogd/ϵ) samples. Computationally, this currently needs d^{Ω(k)} run-time. Alternatively, with N=O(d), we can get polynomial-time. Thus, there is an exponential gap (in the dependence on d) between the two and we do not know if it is possible to get do(k) run-time and o(d) samples.
We give the first generic positive result for worst-case design matrices X: For any X, we show that if the support of w is chosen at random, we can get prediction error ϵ with N=poly(k,logd,1/ϵ) samples and run-time poly(d,N). This run-time holds for any design matrix X with condition number up to 2^poly(d).
Previously, such results were known for worst-case w, but only for random design matrices from well-behaved families, matrices that have a very low condition number (poly(logd); e.g., as studied in compressed sensing), or those with special structural properties. This is joint work with Raghu Meka and Konstantinos Stavropoulos https://arxiv.org/abs/2511.06211

Speaker Bio:
Gautam Chandrasekaran is currently pursuing his PhD at UT Austin. His research interests are in learning theory and complexity theory. He graduated from IIT Madras with a B.Tech (DD) degree in CSE in 2022.

Speaker: Gourav Takhar (IIT Kanpur)

Date: Aug 03, 2026

Time: 2 PM – 3:15 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Incorrectness Logic provides a formal framework for proving the existence of bugs in programs. While it has been highly successful in explaining modern bug-finding tools, automating its reasoning principles—particularly its backward variant rule—has remained a significant challenge. Consequently, existing tools typically rely on bounded loop unrolling, which is effective for detecting shallow bugs but often fails to discover bugs that require deeper reasoning.

In this talk, I will present a new proof system, I3, that replaces backward variants with a novel proof construct called underapproximation invariants. Unlike traditional invariants, which must hold for all program executions, underapproximation invariants are required to hold only for a subset of executions. I will discuss how these invariants enable partial incorrectness proofs and show that, under suitable conditions, such proofs can be lifted to guarantees of total incorrectness. Building on this theory, we developed I3Prove, an automated prover that synthesizes underapproximation invariants to establish incorrectness proofs. Experimental evaluation on 106 incorrect-program benchmarks shows that I3Prove proves 97 incorrectness claims, substantially outperforming state-of-the-art bug-finding tools and recent approaches based on abstract interpretation.

Bio:
Gourav Takhar is a Postdoctoral Researcher in the Department of CSE at IIT Kanpur, working with Prof. Subhajit Roy. He received his Ph.D. from IIT Kanpur, where his research focused on synthesis for security and bug detection. His research interests include formal methods, program verification, program synthesis, and computer security. He has published in leading venues such as OOPSLA, CAV, TACAS, ASE, TOSEM, HOST, and ICCAD. His work received the Best Paper Award at ICCAD. His recent research explores automated program verification using synthesis and large language models.

Speaker: Rahul Vashisht (CS18D006)

Date: Aug 03, 2026

Time: 10 AM – 11 AM

Venue: MR - II (CSB 21, First Floor)

Abstract:

Attention mechanisms have revolutionized deep learning, leading to breakthroughs in natural language processing, computer vision, and multimodal tasks such as image captioning. Despite their empirical success, the learning dynamics of attention-based models remain poorly understood. Specifically, it remains unclear how learned parameters evolve and how attention and predictive distributions emerge, evolve, and specialize during training. This opacity limits the ability to debug models, ensure reliability, and build trust in high-stakes applications. Existing studies typically analyze attention either at initialization or at convergence, offering limited insight into how attention mechanisms develop over time.

To address this gap, the thesis presents empirical and analytical studies of the learning dynamics of attention-based models. It examines three attention paradigms- soft attention, hard attention, and latent-variable training under a marginal-likelihood objective, under a Selective Dependence Classification setting. All three paradigms learn a `focus model' that selects a relevant segment of the input and a `classification model' that maps the selected segment to the target label. However, they differ in how the selected segments are aggregated, resulting in distinct training dynamics and predictive distributions.

Building on this study, the thesis analyzes the gradient dynamics of a single-layer self-attention model by separating the query-key and output-value circuits. Under an orthogonal synthetic-data setting, the analysis shows that the two parameter matrices evolve along fixed directions determined by the data. Thus, the resulting coupled scalar differential equations describe the evolution of the parameters under gradient flow. The analysis also shows that factorized parameterizations induce state-dependent preconditioning of the corresponding collapsed dynamics. Increasing the query-key learning rate relative to the output-value learning rate results in sharper attention on task-relevant tokens. This learning-rate intervention also maintains comparable predictive performance. The thesis evaluates these interventions on hate-speech classification, subject-verb agreement, and question answering.

The thesis proposes a formal approach to evaluating interpretability by quantifying the alignment between learned attention distributions and ground-truth relevant segments. Through the Focus-Classify Attention Model framework and a set of objective interpretability metrics, the study identifies a class of failure cases in which accurate classification coexists with misleading attention patterns. To mitigate these cases, sparsity-inducing regularization is shown to substantially improve attention alignment on both synthetic and real-world datasets, with comparable performance.

Finally, the thesis empirically examines how imbalances in self-attention parameters affect attention mass and entropy. A $lambda$-balanced initialization of the circuit parameters produces distinct learned attention patterns and training dynamics. These results demonstrate that the relative scaling of parameter circuits at initialization influences how attention develops during training.

Together, these contributions offer new theoretical insights and practical tools for understanding and shaping how attention is learned. The findings support the development of attention mechanisms that are not only high-performing but also more interpretable and transparent in their decision-making processes.

Speaker: Dr. Prashant Kumar (IIT Delhi)

Date: Jul 31, 2026

Time: 12 PM – 1 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Real-world LiDAR point clouds assisted navigation suffers from challenging constraints due to sensor limitations, limited compute and unavailability of any form of labeled information. To add to this, sparsity dilutes the geometric details due to loss of point density, degrading multiple downstream modules including segmentation and navigation. In this talk,
I will discuss the global connectivity hypothesis, which states that a point cloud’s global connectivity backbone remains intact even after significant drop in density. I will discuss techniques used to test the hypothesis and exploit this connectivity, using graphical representations of point clouds and topological tools from Algebraic Topology to capture the intact topological signatures. These signatures were extracted in the form of a global backbone skeleton of the point cloud using 0-dimensional Persistent Homology, followed by a novel topological regularization that ensures that assimilating/augmenting new static points follows the global topology of the existing static structure outline of the LiDAR point cloud. The results demonstrate considerable improvements in navigation and segmentation. Towards the end, I may briefly talk about a recent discovery of non-trivial topological gaps between real world point clouds and current point cloud benchmarks.
Speaker Bio:
Dr. Prashant Kumar is currently a Postdoctoral Research Associate at the Department of Computer Science and Engineering at IIT Delhi, India working with Prof. Sayan Ranu. He completed his doctoral studies from the School of IT at IIT Delhi under the supervision of Prof. Srikanta Bedathur and Prof. Prem Kalra, and was supported by the prestigious Qualcomm Fellowship. Before that, he worked as a Research Associate at the Indian Institute of Science for about two years under Prof. Chiranjib Bhattacharyya. Even before that, he earned his Masters from IIT Guwahati. His current research has focused on cross-pollinating ideas from Algebraic Topology (AT) and utilizing topological tools for problems ranging from autonomous navigation to industrial point clouds. These tools were successfully applied to solve autonomous navigation challenges for LiDAR in extremely constrained real-world settings. His research has also studied topological properties of industrial point clouds. He has collaborated with researchers and students at academic partners of MILA, Canada and the University of Antwerp, Belgium. He is also actively collaborating with industry partners - Flipkart and Dolby Labs. He has published across top conferences and journals including CVPR, ICCV, KDD, AAAI and Pattern Recognition.

Speaker: Pratheek B (NVIDIA)

Date: Jul 31, 2026

Time: 10 AM – 11:30 AM

Venue: MR1 (SSB 233, First Floor)

Abstract:

The volume of data generated worldwide is growing at an unprecedented rate, and GPUs have emerged as the primary compute engine for processing this data. While GPUs offer massive compute power — reaching thousands of TFLOPS — they are constrained by the relatively low memory bandwidth (around a few TB/s) and limited memory capacity (in tens of GBs). As a result, memory is often the primary performance bottleneck in many GPU applications. In this talk, we explore two key aspects of GPU memory management: address translation and memory oversubscription. Efficient address translation is important in GPUs, as it lies in the critical path of memory accesses, and thus impacts overall GPU performance. Memory oversubscription enables GPU programs to work on datasets larger than the on-board GPU memory, but can lead to severe slowdowns.

In the first part of this talk, we focus on the impact of non-uniformity of Multi-Chip-Module (MCM) design on address translation in GPUs. MCM design allows GPUs to scale beyond limits of single monolithic chips by integrating multiple ‘chiplets' with high bandwidth interconnects. Unfortunately, MCM design leads to non-uniform resource access latencies. We analyzed the impact of MCM design on address translation and found large latency increases due to the disaggregated nature of TLBs and page walkers in MCM GPUs which leads to significant performance degradation. We proposed MCM-aware GPU Virtual Memory (MGVM), which leverages the access patterns of GPU applications to limit remote TLB and page table accesses, ultimately improving performance by 52% across diverse applications.

The second part of this talk looks at challenges posed by GPU memory oversubscription. GPU memory oversubscription enables GPU applications to work with datasets larger than the GPU memory capacity, using the CPU memory as swap space. Unfortunately, applications under GPU memory oversubscription often experience significant slowdowns. Our work, ObservUVM, improves GPU's eviction and prefetching policies by enabling observability into GPU’s memory accesses to pages resident on GPU memory — something current GPUs lack. We show that current eviction and prefetching policies, handled by the driver running on the CPU, are limited in their ability to make informed decisions due to the lack of observability. ObservUVM enables observability into GPU’s memory accesses by repurposing existing hardware access counters, enabling better-informed eviction and prefetching policies. ObservUVM improves UVM performance by around 33% across 14 applications.

Bio: Pratheek is a Senior Software Engineer at NVIDIA, working on improving GPU memory management. He obtained his Ph.D. from the Indian Institute of Science, Bengaluru in 2026. His research revolves around improving GPU memory management, focusing on address translation and data placement for large-memory workloads. His work spans the domains of GPU micro-architecture, compiler techniques, and operating systems. Previously, he had worked on reverse-engineering Nvidia GPUs

Date: Jul 29, 2026

Time: 2:30 PM – 4:30 PM

Venue: CSB-25 (Aryabhatta Hall)

Abstract:

The Department Orientation Program for the first-year B.Tech students (2026 Batch) will be held on 29 July 2026 (Wednesday) at 2:30 p.m. in CSB-25 (Aryabhatta Hall).

Speaker: Nikhil S. Mande (University of Liverpool)

Date: Jul 27, 2026

Time: 2:30 PM – 4 PM

Venue: Online

Abstract:

Can quantum computers solve problems substantially more efficiently than classical computers? At its broadest, this question is captured by the relationship between the complexity classes BPP and BQP, and resolving this remains far beyond the reach of current techniques. A fruitful approach is to study restricted models of computation in which classical and quantum algorithms can be compared unconditionally.

In this talk, I will introduce two such models (query complexity and communication complexity) and survey some of my work on algorithms, lower bounds, and structural relationships between these models. I will then focus on my recent work on quantum search with generalized wildcards. Here, the goal is to learn an unknown bitstring using queries that test whether selected parts of the string agree with a proposed pattern. We obtain tight quantum query bounds in a general setting parametrized by the collection of subsets that may be queried. A key feature of our approach is that it turns the negative-weight adversary method, normally viewed as a lower-bound technique, into a tool for designing quantum algorithms. I will conclude by mentioning some interesting directions for future research.

Bio: Nikhil S. Mande is a Lecturer (Assistant Professor) in Computer Science at the University of Liverpool. He received his PhD in Computer Science from the Tata Institute of Fundamental Research, Mumbai, in 2018. He subsequently held postdoctoral positions at Georgetown University and at CWI, Amsterdam, before joining Liverpool in January 2023. His PhD focused on classical communication complexity, but his research interests have since broadened to the intersection of quantum computing and computational complexity theory, particularly quantum and classical query and communication complexity.

Speaker: Subin Pulari from HSE University, Moscow

Date: Jul 24, 2026

Time: 2:30 PM – 3:30 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Randomness is central to many areas of mathematics and computer science. Algorithmic Information Theory addresses the foundational question: what does it mean for an individual data sequence to be random? We study this question in settings where randomness is measured under computational resource bounds.
We begin with a broad overview of resource-bounded randomness and dimension. These notions provide a way to quantify the information content of individual sequences under computational constraints.

We then study the connection between one-way functions and polynomial-time dimension (joint work with Nandakumar, Akhil S and Sarma, RANDOM 2026). Polynomial-time dimension measures the asymptotic information rate of an infinite sequence under polynomial-time constraints. It has two natural formulations: a compressibility formulation using polynomial-time-bounded Kolmogorov complexity and a predictability formulation using polynomial-time betting algorithms. In the unbounded, PSPACE, and finite-state settings, these formulations are known to coincide. Whether these two formulations coincide at the polynomial-time level remained open for two decades.

We give a negative answer under the assumption that one-way functions exist. One-way functions imply the existence of polynomial-time samplable distributions whose typical sequences exhibit a strict gap between polynomial-time dimension and polynomial-time Kolmogorov complexity rate, and this gap can be made arbitrarily close to one. Conversely, the existence of such a samplable distribution already implies the existence of infinitely-often one-way functions. Thus, this failure of robustness in polynomial-time dimension is not merely a consequence of cryptographic hardness, but is itself tightly connected to it, revealing a new link between asymptotic information measures, cryptography, and resource-bounded Kolmogorov complexity, with broader connections to metacomplexity.

We next study the power and limitations of bounded-memory computation for randomness extraction and prediction. In recent work on finite-state randomness extraction (joint work with Akhil S, LICS 2026), we show that finite-state transducers cannot, in general, extract higher finite-state dimension from a source, and that this limitation persists even in the multi-source setting. Thus, in sharp contrast with the constructive-dimension setting, multiple independent sources do not necessarily enable bounded-memory computation to extract higher randomness. A related line of work on the Agafonov and Schnorr–Stimm theorems for probabilistic automata (joint work with Bienvenu and Gimbert, FSTTCS 2025) shows that, in the finite-state regime, access to internal randomness does not increase predictive power against normality. Together, these results lead to a broader program of understanding when randomness provides additional computational power and when it does not in restricted models of computation.

Bio: Subin Pulari is an Assistant Professor in the Big Data and Information Retrieval School, Faculty of Computer Science at HSE University, Moscow. Before joining HSE, he was a CNRS Postdoctoral Researcher at LaBRI, Université de Bordeaux. He received his Ph.D. and M.Tech. in Computer Science and Engineering from IIT Kanpur. His research lies in theoretical computer science, with interests in algorithmic information theory, computational complexity, resource-bounded randomness, and computable analysis.

Date: Jul 23, 2026

Time: 3 PM – 4 PM

Venue: CSB-25 (Aryabhatta Hall)

Abstract:

The Department Orientation Program for the first-year M.Tech students (2026 Batch) will be held on 23 July 2026 (Thursday) at 3:00 p.m. in CSB-25 (Aryabhatta Hall).

Speaker: Vinod Vaikuntanathan (MIT)

Date: Jul 23, 2026

Time: 10 AM – 11 AM

Venue: SSB-334 (A M Turing Hall)

Abstract:

I will discuss the notion of robust embeddings, highlighting how cryptographic perspectives help formalize and reason about robustness in representation learning. Second, I will introduce trapdoored matrices, linear transformations equipped with hidden structure, and explain what they enable in machine learning settings. Together, these two results illustrate how cryptographic thinking can lead to new concepts, tools, and questions at the interface of Cryptography and Machine Learning.

Bio: Vinod Vaikuntanathan is the Ford Foundation Professor of Engineering in the EECS department at MIT, a principal investigator at MIT CSAIL, and the chief cryptographer at Duality Technologies. He earned his BTech degree from the Indian Institute of Technology Madras in 2003, and his SM and PhD degrees from MIT in 2005 and 2009, respectively. After a postdoctoral stint at IBM Research, a year as a researcher at Microsoft, and two years as a faculty member at the University of Toronto, Vinod joined the faculty of MIT EECS in September 2013.
Vinod's research is on the foundations of cryptography and its applications to theoretical computer science at large. He is known for his work on fully homomorphic encryption (a powerful cryptographic primitive that enables complex computations on encrypted data), as well as lattice-based cryptography (which lays down a new mathematical foundation for cryptography in the post-quantum world). Recently, he has been interested in the interactions of cryptography with quantum computing, as well as with statistics and machine learning.

Vinod's work has been recognized with the Harold E. Edgerton Faculty Award (2018), the Godel Prize (2022), the Simons Investigator Award (2023), the Distinguished Alumnus Award from IIT Madras (2024), a Guggenheim Fellowship (2026), a Best Paper Award from CRYPTO 2024, and test of time awards from IEEE FOCS and CRYPTO conferences. He was also named a MacVicar Faculty Fellow in 2024 for exceptional teaching and mentoring, and an IACR Fellow in 2026.

Speaker: Vaibhav Krishan, IMSc, Chennai

Date: Jul 22, 2026

Time: 11:30 AM – 12:30 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Practical systems routinely solve problems that are theoretically intractable. A prominent example is the Quantified Boolean Formula (QBF) problem, which asks whether a player has a winning strategy in a finite two-player game. Despite being PSPACE-complete, QBF is a powerful modelling framework for applications including planning, model checking, reactive synthesis, and reasoning about multi-agent systems.

In this talk, I will focus on counting winning strategies in such games. I will present a practical system that efficiently solves important instances of this problem, discuss the theoretical foundations underlying its design, and show how recent theoretical advances can guide the development of more effective practical solvers.

The talk is based on joint work with Sravathi Chede, Leroy Chew, and Anil Shukla (to appear at SAT 2026), as well as ongoing work with Sravathi Chede, Swagato Sanyal, and Anil Shukla.

Speaker Bio:
Vaibhav is a postdoctoral fellow at IMSc, Chennai. He completed his PhD from the CSE Dept at IITB, supervised by Prof. Sundar Vishwanathan and Prof. Nutan Limaye

Speaker: Prof. Arun Ross (Michigan State University)

Date: Jul 20, 2026

Time: 11 AM – 12 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Biometrics is the science of recognizing humans based on their biological and behavioral traits such as face, fingerprint, iris, voice, and gait. Recent advances in Large Language Models (LLMs), Vision-Language Models (VLMs), and Multimodal Large Language Models (MLLMs) are creating new opportunities for biometric recognition, analysis, and explanation. While biometric systems have traditionally relied on task-specific models, generic foundation models based on LLMs, VLMs, and MLLMs offer a more general framework for interpreting biometric data through both visual and semantic cues.

This talk will examine the emerging role of foundation models in biometrics, with examples drawn from recent work on face and iris analysis. Topics include zero-shot and few-shot biometric recognition; the adaptation of general-purpose segmentation models for iris localization; interactive face/iris analysis through foundation models; face demorphing using MLLMs and diffusion models; and the use of language-based explanations for biometric comparisons.

While MLLMs can generate human-readable explanations and support interactive analysis, their explanations may not always correspond to verifiable image evidence or biometric cues. Their performance may depend on prompts, image context, and task formulation, and their explanations may appear convincing even when they are not fully supported by biometric evidence. In this regard, we will present a method based on the likelihood ratio to measure and improve the evidential value of textual explanations produced by MLLMs. Finally, we will enumerate the advantages and disadvantages of utilizing foundation models in the context of biometrics.


Bio:
Arun Ross is the Martin J. Vanderploeg Endowed Professor in the Department of Computer Science and Engineering at Michigan State University, and the Site Director of NSF’s Center for Identification Technology Research (CITeR). He received the B.E. (Hons.) degree in Computer Science from BITS Pilani, India, and the M.S. and PhD degrees in Computer Science and Engineering from Michigan State University. He was in the faculty of West Virginia University between 2003 and 2012 where he received the Benedum Distinguished Scholar Award for excellence in creative research and the WVU Foundation Outstanding Teaching Award.

Ross is an internationally recognized expert in biometrics, privacy, computer vision and deep learning. He is a recipient of the JK Aggarwal Prize (2014) and the Young Biometrics Investigator Award (2013) from the International Association of Pattern Recognition for his contributions to the field of Pattern Recognition and Biometrics. He was designated a Kavli Fellow by the US National Academy of Sciences by virtue of his presentation at the 2006 Kavli Frontiers of Science Symposia. Ross is also a recipient of the NSF CAREER Award.

Ross has advocated for the responsible use of biometrics in multiple forums including the NATO Advanced Research Workshop on Identity and Security in Switzerland in 2018. He testified as an expert panelist in an event organized by the United Nations Counter-Terrorism Committee at the UN Headquarters in 2013. In June 2022, he testified at the US House Science, Space, and Technology Committee on the topic of Biometrics and Personal Privacy. He is a co-author of the monograph “Handbook of Multibiometrics” and the textbook “Introduction to Biometrics”.

Speaker: Divyanshu Upadhyay (CS23S031)

Date: Jul 20, 2026

Time: 11 AM – 12 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Modern Spring Boot applications often implement a single user operation through multiple controller executions and database transactions. As a result, concurrently executing user tasks can interact in subtle ways, leading to serializability violations even when each individual transaction is executed atomically. Such concurrency bugs are difficult to reproduce because they depend on specific execution schedules and therefore often escape conventional testing.
In this work, we present a bounded model checking framework for detecting such task-level conflict serializability violations in transactional Spring Boot applications. The framework models user tasks as sequences of controller transactions and systematically explores their possible interleavings using Java PathFinder to identify anomalous executions. A key aspect of this work is accurately modelling the behavior of Spring Boot applications, including Spring MVC control flow, repository operations, ORM behavior, entity relationships, cascading operations, and database constraints, allowing the analysis to capture dependencies that are not visible through traditional transaction-level reasoning. We evaluate the proposed framework on benchmark and real-world Spring Boot applications, demonstrating its effectiveness in uncovering subtle application-level concurrency bugs that are difficult to detect using conventional testing techniques.

Date: Jul 17, 2026

Time: 3 PM – 6 PM

Venue: Sangam Grounds (Behind Godavari Hostel)

Abstract:

Convocation Batch 2026 - Degree Distribution Program (D3P) of the Department of Computer Science and Engineering for the outgoing batch of 2026 students.
Youtube Live Streaming Link - https://www.youtube.com/live/A_a1EI03g34?t=17m10s

Speaker: Dr. Pulkit Verma

Date: Jul 16, 2026

Time: 11 AM – 12 PM

Venue: 606 Seminar Hall, NAC 2

Abstract:

AI systems are increasingly interacting with users who are not experts in AI. This has led to growing calls for better safety assessment and regulation of AI systems by users, the government and even the industry itself. The assessment of safety and functionality in taskable AI systems designed for continuous learning and task execution in user-specific environments poses a significant challenge. This question, ranging from household robotics to digital assistants capable of making potentially hazardous changes to their operational environments, is crucial to realizing an AI system's full potential. In this talk, I (i) explore why addressing this challenge requires more than simply extending existing verification and validation paradigms, (ii) identify specific requirements and promising research directions for the formal assessment of AI systems, and (iii) present a few solutions that are already developed for such scenarios under restricted settings.

Speaker Bio: Pulkit Verma is an Assistant Professor in the Department of Computer Science and Engineering at IIT Madras, where he leads the Human-Aligned Intelligent Learning Systems (HAILS) Group. He is also a co-PI at the Centre for Responsible AI (CeRAI) at IIT Madras and an affiliate of the Center for Human Compatible AI at UC Berkeley. Before this, he was a Postdoctoral Associate in the Interactive Robotics Group at MIT CSAIL. He completed his Ph.D. at Arizona State University and his M.Tech. at IIT Guwahati.


His research focuses on building safe, reliable, and adaptable AI agents that can work alongside humans effectively. His lab works at the intersection of AI safety, lifelong learning with humans in the loop, and human-robot collaboration, with an emphasis on understanding the minimal requirements an AI system needs so that users can assess and understand the limits of its safe operability. His broader interests span AI planning, interpretability, human-AI collaboration, and robotics.


Pulkit has been recognized with several awards, including the ICAPS 2025 Outstanding Dissertation Award, the 2025 AAAI/ACM SIGAI Innovative AI Education Award, the AAMAS 2022 Best Demo Award, and the 2023 Graduate College Completion Fellowship at ASU. You can learn more about his work at https://pulkitverma.net.

Speaker: Santhosh Raghul (CS22D406)

Date: Jul 14, 2026

Time: 2 PM – 3 PM

Venue: CS 36 (Second Floor)

Abstract:

Graph minors occupy a central position in structural graph theory. In the same spirit, there are containment notions such as ‘conformal minors’ that show up naturally in the study of perfect matchings; we refer the interested reader to Chapter 12 of Lucchesi and Murty’s monograph [Perfect Matchings: A Theory of Matching Covered Graphs, 2024]. We propose to investigate graph classes that forbid specific conformal minors such as θ, K4, the triangular prism C6, C4 (read: cubic-C4)
and K3,3. The first three of these are intrinsically tied to the ear decomposition theory, and the last three have connections to other open problems of significance. (Here, θ and C4 are the only connected bipartite cubic graphs of order two and four, respectively.) The Ear Decomposition Theorem of Lov ́asz and Plummer [Matching Theory, 1986] implies that every matching covered graph, except K2 and cycles, contains at least one of θ and K4 as a conformal minor. Lov ́asz [Combinatorica, 1983] proved the refinement that every nonbipartite matching covered graph contains at least one of K4 and the triangular prism C6 as a conformal minor. These immediately lead to three problems: characterize θ-free (matching covered) graphs,
and likewise, K4-free and C6-free ones. Kothari and Murty [J. Graph Theory, 2016] settled the planar cases of the latter two; the nonplanar cases are open.
We shall discuss two graph classes arising from the perfect matching polytope — ‘Birkhoff– von Neumann’ and ‘PM-compact’, and relate them to C6-free and C4-free classes, respectively. We shall also discuss the famous Pfaffian problem, which has its roots in counting the number of perfect matchings (that is, the dimer problem in theoretical physics), and relate it to the problem of characterizing K3,3-free graphs.
Finally, we provide a characterization of θ-free graphs that immediately implies a poly-time algorithm for the corresponding decision problem. Our result and its proof relies heavily on a seminal result due to Edmonds, Lov ́asz and Pulleyblank [Combinatorica, 1982] pertaining to the tight cut decomposition theory of matching covered graphs. For the class of θ-free graphs, we obtain several corollaries: (i) an upper bound of 2n − 2 on the size, (ii) forbidden conformal minor characterization of Pfaffian graphs, and likewise, of 3-edge-colorable cubic graphs, and (iii) the surprising fact that the cubic ones are precisely those whose each conformal cycle is of length 0 (mod 4). The Petersen graph plays a key role in our results.

Speaker: Anubhav Paul (IITD, FP Launchpad Summer Intern)

Date: Jul 14, 2026

Time: 2 PM – 3 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Formal verification, the practice of proving that a program is correct for every possible input, not merely those that have been tested, is widely regarded as specialist work. This talk argues that this reputation is largely a teaching problem.

Using familiar pen-and-paper arithmetic algorithms as case studies, I show that the central ideas of formal verification, specifications, loop invariants, induction, and termination, are already latent in these algorithms and emerge naturally once we ask why they work. The accompanying open-access book, Learn2Lean, makes this concrete: each chapter implements one algorithm in Lean 4, proves it correct, and pairs the proof with step-by-step visualizations and an in-browser exercise checker that compiles real Lean code against student submissions. The goal is a path into formal verification where the only barrier is understanding the ideas themselves.

No background in Lean or proof assistants required.

Speaker: Nikhil Balaji (IITD)

Date: Jul 13, 2026

Time: 3 PM – 4 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

The sum of square-roots problem is an important computational problem in numerical analysis, with applications to computational geometry: Given positive integers a_1, a_2, ..., a_n and b_1, b_2, dots, b_m, check if sqrt{a_1} + sqrt{a_2} + dots + sqrt{a_n} > sqrt{b_1} + sqrt{b_2} + dots + sqrt{b_m}. The problem has a trivial linear time algorithm on the real RAM (just compute the sums and compare them!) but surprisingly this is not a polynomial time algorithm in the traditional Turing machine model. This problem is an important subroutine in computational geometry, (for example, Euclidean shortest path problem, Euclidean travelling salesperson problem), complicating the transfer of algorithms from the real RAM to the standard integer RAM.

I will survey what is known about this problem and the difficulty in obtaining an efficient algorithm. I will present an interesting observation that gives an "efficient algorithm" for certain special cases of the problem and explain what it would take to make "efficient" actually efficient for the general case.

Based on joint work with Samir Datta.

Speaker bio: Nikhil is an assistant professor in the Department of Computer Science and Engineering at IIT Delhi. His research lies in theoretical computer science, with a particular focus on complexity theory, automata theory and algebraic computation. Much of his recent work and interests have focused on decision problems in the theory of real and complex numbers and their applications to verification. More broadly, he is interested in understanding the boundary between efficient computation and computational hardness through algebraic problems.

Speaker: Udit Narayan (CS23S038)

Date: Jul 07, 2026

Time: 11 AM – 12 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Adaptive traffic signal control (TSC) using reinforcement learning (RL) is a well-studied problem in the literature, and these algorithms operate primarily in the risk-neutral regime. However, the incorporation of risk criteria aligned with the TSC problem and the design of risk-sensitive RL algorithms have not received sufficient attention in the literature. We fill this gap for the distortion risk measure (DRM), a rich class of risk measures that includes the popular conditional value-at-risk (CVaR). We formulate the risk-sensitive TSC problem with a DRM objective. The policy gradient is a popular approach for applying RL to high-dimensional problems. We adopt this approach to solve the DRM-sensitive TSC problem. We perform detailed simulation experiments using SUMO in three settings: networks with a single intersection, two intersections, and a network based on real-world data. We implement our DRM policy gradient algorithm under two distortion functions corresponding to dual power and quadratic risks, and compare their performance to the risk-neutral vanilla policy gradient algorithms as well as a variant with a risk measure based on cumulative prospect theory (CPT). From the experiments, we observe that our algorithms yield lower mean delay and reduced variance compared to the risk-neutral baseline for two objective functions: waiting-time minimization and emissions minimization. Amongst our algorithms, we observed that the DRM policy gradient algorithms achieved 23% and 16% reductions in the mean and standard deviation of waiting time at low congestion levels, and 32% and 73% reductions at high
congestion levels. Furthermore, the CPT-based algorithm reduces the mean emissions by 44% and the standard deviation by 35% compared to the traditional risk-neutral method, but at the cost of a significant increase in waiting time. In contrast, the DRM algorithm variants result in lower emissions compared to the risk-neutral baseline, and a marginal increase in waiting times compared to the waiting-time-based algorithms.

Speaker: Prof. Karthikeyan Sundaresan (Georgia Tech)

Date: Jul 02, 2026

Time: 11 AM – 12 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

With the growing popularity of LEO satellites and the resulting connectivity to ground (dish) terminals, direct-to-cell broadband networking forms the next logical evolution in this progression. While 3GPP is evolving the 5G protocol and architectural needs -- originally designed for terrestrial networks -- to support non-terrestrial networking, several challenges remain. In particular, “network mobility” and “scale”, where relative to the ground stations and
user devices on the ground, the high mobility of the large-scale (LEO satellite) network fabric itself, necessitates a rethink of several of the design principles behind orchestrating these NTNs.
In this talk, we will focus on the problem of 5G session orchestration in NTNs as an example, that involves connecting user devices to appropriate ground gateways through the dynamically evolving satellite constellation, while catering to multiple KPIs (throughput, latency, coverage, utilization, stability) simultaneously. In essence, the more the satellite network mobility can be hidden from the relatively-static end devices, the better the ability to balance these often-conflicting KPIs. Yet, realizing this in practice will involve -- all the way from the architectural designs to accommodate satellite mobility, to the design of efficient, distributed orchestration mechanisms, as well as the platform and framework needed to emulate and validate such mechanisms at constellation-scale. To this end, the talk will outline some of the recent progress made in the design of an open-source NTN digital twin that enables realistic network emulation of dynamic NTNs and their solutions, as well as efficient multi-agent reinforcement learning (MARL) based orchestration solutions that are trained-validated using the digital twin, while being deployed locally on individual satellites for real-time decision making at scale. The talk will conclude with an outlook on the resulting SatGym deployment framework that will be open-sourced for accelerating RL-based research in satellite networks.

Bio: Karthikeyan (Karthik) Sundaresan is a Professor in the School of ECE, Georgia Tech, and a co-director of its Center for Wireless Intelligence. His research interests are broadly in wireless networking and mobile computing, and span both algorithm design as well as system prototyping. He is the recipient of ACM Sigmobile’s Rockstar award (2016) for early career contributions to mobile computing and wireless networking, as well as several best paper awards at prestigious ACM and IEEE conferences. He holds over sixty patents, and received business contribution awards for bringing research technology to commercialization in industry at NEC. He also led the spin-out efforts of two innovative, lab-grown research technologies for infrastructure-free tracking of first responders in GPS-denied environments and sustainable, massive scale product tracking in supply chains. He has participated in various organizational and editorial roles for IEEE and ACM conferences and journals, and served as the PC co-chair for ACM MobiCom’16. He is a Fellow of the National Academy of Inventors and IEEE, and an ACM distinguished scientist.

Speaker: Sheera Shamsu (CS20D001)

Date: Jun 30, 2026

Time: 4 PM – 5 PM

Venue: MR1 (SSB 233, First Floor)

Abstract:

The OCaml programming language finds application across diverse domains, including systems programming, web development, scientific computing, formal verification, and symbolic mathematics. OCaml is a memory-safe programming language that uses a garbage collector (GC) to free unreachable memory automatically. It features a low-latency, high-performance GC tuned for functional programming, with two generations: a minor heap collected using a copying collector and a major heap collected using an incremental mark-and-sweep collector. Alongside the intricacies of an efficient GC design, the OCaml compiler uses efficient object representations for some object classes, such as interior pointers for supporting mutually recursive functions, which further complicates the GC design. The GC is a critical component of the OCaml runtime system, and its correctness is essential for the safety of every OCaml program that runs on top of it.
In this thesis, we propose a strategy for crafting a correct, proof-oriented GC from scratch, designed to evolve with additional language features. Our approach neatly separates abstract GC correctness from OCaml-specific GC correctness, offering the ability to integrate further GC optimisations while preserving the core abstract correctness guarantees. The correctness specifications are based on abstract graph reachability, allowing us to describe GC correctness without referring to implementation details, and making them reusable across different collector designs.

As an initial step to demonstrate the viability of our approach, we developed a verified stop-the-world mark-and-sweep GC for OCaml 4.14.1. We adopt a layered verification architecture that descends from abstract graph reachability, through an OCaml-specific functional GC layer, down to memory-safe executable C code. The abstract layer establishes the correctness of a depth-first search algorithm over graphs using inductively defined reachability. The functional layer proves that the GC correctly identifies and preserves all reachable objects while reclaiming unreachable ones, operating on OCaml-style object layouts including closures and infix objects. The low-level layer verifies memory safety of the implementation, ensuring freedom from buffer overflows, use-after-free, and unsafe pointer dereferences. We further demonstrate the general applicability of the reachability-based specification framework by extending it to two additional collector designs used in the OCaml runtime: an incremental mark-and-sweep collector, which ensures correctness despite mutator-induced changes to the heap during the GC cycle, and a copying collector, whose correctness is expressed through a graph isomorphism that arises from the object relocation mapping. Although these extensions are not implemented and verified, they illustrate that the same foundational specification can be applied to different collector designs without necessitating huge changes in abstract correctness specifications.

The approach is fully mechanised in F★ and its low-level subset Low★. We use the KaRaMeL compiler to extract the verified Low★ implementation to portable C, and integrate the extracted code with the OCaml 4.14.1 runtime. The verification comprises approximately 39,000 lines of F★/Low★ code and proofs. Our GC is evaluated against the standard OCaml GC and the Boehm–Demers–Weiser conservative GC on a range of benchmarks from the Computer Language Benchmarks Game as well as larger real-world OCaml programs. The experimental results show that the verified GC is competitive with the standard OCaml GC on many workloads and consistently outperforms the conservative collector, demonstrating that strong formal correctness guarantees need not come at an unacceptable performance cost.

Speaker: Vimala S (CS19D750)

Date: Jun 30, 2026

Time: 3 PM – 4 PM

Venue: MR1 (SSB 233, First Floor)

Abstract:

Replication is fundamental to building scalable and fault-tolerant distributed systems, but concurrent updates across replicas introduce conflicts that must be resolved without compromising correctness. Conflict-free Replicated Data Types (CRDTs) address this challenge by ensuring convergence through carefully designed merge operations that satisfy algebraic properties such as commutativity, associativity, and idempotence. This convergence is commonly formalized as strong eventual consistency, a standard correctness criterion requiring that replicas that have received the same set of updates eventually converge to the same state. CRDTs depend on the monotonic growth of state within a join-semilattice, where replicas reconcile by computing the least upper bound of their states. While this method guarantees convergence by design, it can lead to inefficient implementations due to the inherent growth of metadata.

Mergeable Replicated Data Types (MRDTs) generalize this approach by removing the requirement for monotonic growth or lattice-based joins. Instead, MRDTs are replicated data structures in which replicas apply updates locally and reconcile divergent states using a three-way merge function based on the lowest common ancestor (LCA). This flexibility enables efficient implementations while still ensuring convergence when merge functions are designed correctly. However, reasoning about their correctness is challenging due to arbitrary interleavings of updates and merges and the absence of strong algebraic guarantees. This thesis develops automated frameworks for specifying and verifying the correctness of MRDT implementations with respect to both convergence and functional behavior. The proposed approaches are also applicable to state-based CRDTs.

The first contribution of this thesis is a pragmatic framework for building and verifying efficient MRDTs, called Peepul. The framework formalizes the execution model of MRDTs by casting them in the mold of distributed version control systems and equipping them with a three-way merge function for reconciling conflicting versions. Beyond convergence, the framework enables the verification of rich functional specifications. To achieve this, the thesis introduces a replication-aware simulation relation that connects high-level specifications to efficient purely functional implementations.

While the correctness of replicated data types is commonly specified using strong eventual consistency, more expressive specifications that relate the converged state to the updates received at each replica are often desirable. Replication-aware linearizability is one such specification, requiring that each replica always be in a state that can be obtained by linearizing the updates received at that replica. The second contribution of this thesis is a fully automated technique for verifying replication-aware linearizability for MRDTs, called Neem. This technique identifies novel algebraic properties of operations and merge functions that are sufficient to establish linearizability, going beyond the standard properties of commutativity, associativity, and idempotence. It also introduces an inductive verification method called bottom-up linearization to automatically check these properties. The technique applies to both MRDTs and state-based CRDTs.

Both techniques are implemented in the F* programming language as part of a verification framework that enables developers to write implementations and mechanically prove their correctness. A wide range of data types has been verified using this framework, including sets, maps, counters, flags, lists, queues, and a JSON data structure, demonstrating that the approach scales to realistic and complex systems. The verified implementations are extracted as OCaml code and deployed in Irmin, a distributed database built on Git-like principles.

Speaker: Nitin Kumar Dhiman (CS09D024)

Date: Jun 30, 2026

Time: 10:30 AM – 11:30 AM

Venue: MR1 (SSB 233, First Floor)

Abstract:

Autonomous mobile robots are increasingly essential for high-stakes missions—such as search and rescue, surveillance, and tactical reconnaissance—where prior maps are often unavailable, outdated, or unstructured. In these scenarios, a robot must operate with zero initial knowledge of the environment, constructing a comprehensive understanding of the space incrementally. This requires a seamless interplay between robust navigation, systematic perception, and strategic interaction. This thesis addresses the fundamental challenges of autonomous exploration in unknown environments through four distinct research themes.

The first theme establishes a Multi-floor Navigation Framework (MFNF) to address the topological complexity of multi-storey buildings. Unlike single-floor navigation, multi-floor environments require robots to reason about vertical connectivity (stairs, elevators), manage disconnected maps, and localize within a state space that includes a discrete floor index. The proposed framework integrates the floor index directly into the robot's state representation and details the implementation of costmaps, path planners, and Simultaneous Localization and Mapping (SLAM) optimized for multi-floor environments with vertical transitions.


The second theme introduces a Simultaneous Exploration and Coverage (SEC) strategy designed for the systematic inspection of structures within user-defined regions or dynamically detected boundaries. This is particularly relevant for tactical clearing in security applications, where specific zones must be fully "cleared" before the mission progresses. The SEC algorithm employs a viewpoint selection process that prioritizes exhaustive coverage of the current region, while maintaining the capacity to venture into broader environmental frontiers.


The third theme, Dual-Side Coverage (DSC), addresses a fundamental limitation in autonomous exploration where thin surfaces are prematurely marked as "covered" in traditional volumetric representations. Because a single voxel typically represents a thin surface, standard algorithms often fail to distinguish between the inspected and non-inspected sides, leading to incomplete structural vetting. To resolve this, the DSC strategy generates hypothesized viewpoints targeting the non-visible facets of detected surfaces. These hypotheses are then subjected to a validation process—assessing both target visibility from the candidate pose and robot reachability—before being integrated into the planning cycle to ensure the structure is comprehensively covered from all perspectives.


The fourth theme, Exploration Among Movable Objects (EAMO), shifts the paradigm from passive observation to active interaction. This approach enables robots to physically modify their environment to resolve occlusions and unlock new paths. The proposed algorithm utilizes representation abstractions that treat movable objects as "gateways" between disconnected regions. Viewpoints are categorized into two types: Type-N (Navigation), associated with position-based perception in free space, and Type-O (Object), associated with repositioning obstacles to gain critical spatial awareness in obstructed areas.


Collectively, these contributions demonstrate that while robots begin with no pre-existing information, they can build online spatial representations using only onboard sensors. Together, these strategies provide a foundation for resilient autonomous agents capable of navigating and interacting in indoor structures.

Speaker: Venkata Naga Sai Rohit Kanteti (CS23S019)

Date: Jun 30, 2026

Time: 9 AM – 10 AM

Venue: SSB-334 (A M Turing Hall)

Abstract:

The Ahead-Of-Time (AOT) compilation scheme of OpenJ9 analyzes and optimizes classes and methods at runtime and populates the Shared Classes Cache (SCC) with compiled binaries. Future runs can reuse the native code stored in the SCC. This approach reduces startup and ramp-up times and improves performance by reducing the need for on-the-fly compilation. To ensure that it is safe to use previously stored AOT-compiled code, OpenJ9 uses a conservative approach: if a method 𝑚 has changed since the initial AOT compilation, the AOT-compiled code of all possibly dependent methods (methods for which 𝑚was “peeked” during their compilation) is considered invalid, and those methods are recompiled from scratch. We find that, because of this conservative scheme, after the redefinition of a method whose AOT-compiled code is stored in the SCC, OpenJ9 aborts the AOT loading of many methods and instead JIT compiles them. This can be redundant in many cases, as a simple change may not affect many methods.

To address this challenge, in this work, we propose a framework called SmartAOT that can be used to improve the reuse of AOT-compiled methods. The key insight behind SmarAOT is that the AOT-compiled code of a method m can be reused if the modifications to other parts of the application do not change the effect of the interprocedural optimizations invoked on m, and consequently do not change the AOT-compiled code of m. SmartAOT provides a way to encode such conditions for each of the optimizations, after obtaining the updated analysis results in an incremental fashion. We instantiate SmartAOT for two important interprocedural optimizations of OpenJ9: lock elision and function inlining. We implemented our proposed techniques in the Eclipse OpenJ9 JVM. We present an evaluation using five long-running enterprise applications and show that SmartAOT leads to significantly reduced AOT load failures (geomean: 65.7%), faster ramp-up times (geomean: 13.8%), and reduced compilation times (geomean: 13.7%).

Speaker: Harshil Mittal (Post Doc IIT Madras)

Date: Jun 29, 2026

Time: 11:30 AM – 12:30 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Pure dynamic programming algorithms are DPs whose recurrences use only min and + operations (e.g., Bellman-Ford algorithm for Shortest s-t path problem); these can be viewed as circuits consisting of min and + gates. Typically, algebraic circuits consist of + and × gates, where + and × are addition & multiplication operations of some field, respectively; (min, +)-circuits are similar except that one works over a min-plus semiring (instead of a field), i.e., where addition and multiplication of the semiring are defined as minimum operation and usual addition, respectively.

Over fields, algebraic models such as circuits, formulas, and algebraic branching programs (ABPs) are well studied. In this talk, we will discuss these models over min-plus semirings (motivated by an understanding of the power/limitations of pure DPs) and some of the known results. We define VNP over min-plus semirings. Unlike over rings, we complement the values in the certificate for free, as complementation is impossible over min-plus semirings.

We prove a dichotomy theorem that states that if we only complement logarithmically many values, this class is the same as VP over min-plus semirings. If we complement super-logarithmically many values, then VNP is different from VP.

We consider constant-width ABPs over the min-plus semiring (which are also called incremental dynamic programs that are restricted to use only a constant number of registers) and show that even simple problems like computing the min-weight 2-edge-matching is impossible with width 2 (or 2 registers). However, with width 3 (or 3 registers), such programs are universal. More generally, we show that constant-depth formulas are efficiently simulated by constant-width ABPs.

(Based on a joint work with Balagopal Komarath and Jayalal Sarma, to appear at ICALP 2026)

Speaker: Sudeep Chowdhary (CS23S002)

Date: Jun 25, 2026

Time: 11:30 AM – 12:30 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Compiler optimizations such as stack allocation and scalarization rely on escape analysis to eliminate heap allocations for objects whose lifetimes are confined to their allocating methods. These techniques follow an all-or-nothing approach: if even a single field of an object may be accessed outside the allocating method, the entire object is marked as escaping and must be allocated on the heap – even if all the remaining fields are accessed only within that method.

We present COPE (Class Outlining for Partial scalarization of Escaping objects), a static whole-program technique that enables partial scalarization of escaping objects. COPE outlines the method-local fields of each class (fields accessed only within the methods that allocate the receiver objects) into newly synthesized classes, whose instances are guaranteed to never escape their allocating methods. COPE then tries to scalarize these instances into local variables, thereby shrinking the original escaping objects without introducing any pointer-chasing overhead. Two inter-procedural pre-passes – constructor inlining and field-read hoisting – rewrite the program to significantly increase the number of method-local fields. We implemented COPE in the Soot framework and evaluated it on thirteen benchmarks from the DaCapo and SPECjvm2008 suites, running on the production JVM Eclipse OpenJ9. COPE partially scalarized objects that the JVM alone could not scalarize or stack-allocate, reducing the heap usage by up to 14.41%, the garbage-collection cycles by up to 10.47%, and the execution time by up to 21.1%.

Speaker: Pooja Kumari (CS20D006)

Date: Jun 23, 2026

Time: 11 AM – 12 PM

Venue: Online GMeet

Abstract:

A camera captures the world in an instant, but not always as the world truly appears. Have you ever taken a photo of a tall building, only to see it leaning unnaturally? Or tried to capture a document with your phone, only to end up with skewed and unreadable text? These distortions are not merely visual imperfections, they directly affect the reliability of computer vision systems used for document analysis, OCR, image registration, scene understanding, robotics, and real-world visual automation. Distortions like these, whether caused by perspective, lens aberrations, or uneven illumination pose major challenges in computer vision. This research delves into the problem of image rectification and distortion correction across four key domains: texture images, document images, scene text images, and general outdoor-indoor images, with the goal of making visual data more reliable for downstream applications. While existing techniques address specific distortions, they often fall short in handling real-world scenarios where multiple distortions coexist.

In the thesis, we present a series of works that tackle different facets of this problem with deep learning frameworks. It begins with document image rectification, where prior methods have mostly focused on projective distortions, ignoring real-world issues like shading, contrast, and illumination variations. Proposed DocTLNet, a transfer learning-based network, bridges this gap by leveraging both generic rectification knowledge and document-specific features. But what happens when a document lacks clear boundaries? The thesis introduces DocAttenRect, an attention-based network that processes incomplete or localized text regions, overcoming a major limitation of state-of-the-art rectification methods. This architecture intelligently focuses on text layout rather than relying on visible paper edges. Document rectification is further elevated through a Conditional Stable Diffusion Transformer, merging Thin Plate Splines (TPS) with global self-attention to preserve pristine structural integrity while progressively refining residual warping.

For general indoor and outdoor distortion correction and rectification, the research pioneers the application of generative diffusion models to multi-distortion correction. It introduces a Conditional Stable Diffusion framework paired with a Guided Rectification Network (GRN) to dynamically correct overlapping fisheye, barrel, and perspective errors, an approach that marks the first instance of diffusion-based multi-distortion rectification. Furthermore, it also introduces a transformer-driven diffusion based rectification model that overcomes the limitations of existing architectures by leveraging self-attention mechanisms.
Finally, the thesis demonstrates the profound real-world impact of these geometric corrections through proposed STR-VLNet. It explores how accurate image rectification can enhance the effectiveness and reliability of scene text recognition. Text captured in natural environments, such as street signs, billboards, and storefronts often suffers from perspective distortions, warping, and occlusions, making it challenging for OCR models to extract meaningful information. By correcting these distortions before text recognition, our approach ensures that the text appears in a more standardized, readable format, thereby improving both recognition accuracy and contextual understanding. By fusing CLIP-based vision-language embeddings with rectification, this framework standardizes severely warped scene text on billboards and storefronts, drastically improving downstream OCR accuracy and semantic understanding.

Speaker: Anbarasu (CS12D007)

Date: Jun 19, 2026

Time: 9 AM – 10 AM

Venue: Online GMeet

Abstract:

This thesis focuses on conversational recommender systems built on the Case-Based Reasoning (CBR) paradigm, which solves problems by drawing on prior experiences. These systems simulate the interaction between a customer and a seller in a shop, guiding users through product options in an online environment. The main goal is to help people effectively explore available options while reducing their cognitive load during the decision-making process.

The system uses Preference-based Feedback (PBF), whereby users iteratively select their preferred product from a presented set of recommendations. This feedback is used to update the user model and guide subsequent recommendations. Since PBF is the only form of feedback available, this thesis proposes methods to exploit it more effectively by leveraging product relationships such as similarity, dominance, tradeoffs, and evidence.

The first contribution introduces the Predicted Preference Region-based Method (PPRM), which pre-computes product dominance regions and uses them to improve preference inference during recommendation dialogues. The thesis then develops novel crisp and fuzzy representations of tradeoffs, together with similarity measures that explicitly model the user's willingness to exchange one product feature for another. Building on this, tradeoff evidence is accumulated across interaction cycles to further improve recommendation efficiency.

The thesis also proposes an evidence-based recommendation framework in which feedback on one product contributes to the utility estimation of other products, moving beyond traditional approaches that treat products independently. Finally, a graph-theoretic preference model is introduced, representing products and user preferences as a weighted graph whose stationary distribution provides a final ranking that captures higher-order preference relationships.

The work further discusses the relevance of structured, interpretable preference modelling in the era of Large Language Models (LLMs), highlighting the importance of domain-grounded reasoning and explicit tradeoff modelling in data-sparse recommendation domains.

Experimental evaluations demonstrate significant improvements in recommendation efficiency, reducing the number of interaction cycles required for users to identify suitable products while lowering cognitive effort. Overall, the thesis contributes towards more effective, interpretable, and cognitively inspired conversational recommender systems.

Speaker: Dr. Debabrota Basu (University of Lille, France)

Date: Jun 16, 2026

Time: 4 PM – 5 PM

Venue: 606 Seminar Hall, NAC 2

Abstract:

Modern AI-driven Decision Makers (ADMs) are not passive optimisers-- they interactively influence the systems that they operate within. For example, an automated loan approval system does not only optimise a bank’s long-term profits but also change the wealth dynamics in the society and often enforces the existing socioeconomic biases. First, we formalise the wealth dynamics induced by a loan approving ADM interacting with a multi-demographic population as a “performative” Markov Decision Process (MDP), and develop Eutopia, a lending-process simulator enabled with a novel performative data generator. Thus, we extend the scope of the existing long-term fairness literature that considers only exogenous dynamics of data. Next, to achieve the long-term optimal outcomes in performative MDPs, we propose PePG (Basu et. al, 2026)-- the first policy gradient algorithm for Performative Reinforcement Learning (PRL). We show that PePG provably converges to performative optimality, and achieves better sample-efficiency than the existing performatively stable RL algorithms. Finally, we leverage Eutopia to test PePG against the classical RL algorithms. Experimental results demonstrate that performative dynamics and social outcome aware learning enhances long-term efficiency, equity, and inclusivity.


Bio: Debabrota Basu holds the Inria Starting Faculty Position in the Scool team (previously Sequel) of Inria Centre at Université de Lille and CNRS, France. He completed his PhD from National University of Singapore (NUS) and École Normale Supérieure (ENS), Paris. His research interest is to develop algorithms and analysis leading to statistically and computationally efficient responsible AI systems. Specifically, he studies how to develop robust, private, fair, and explainable algorithms for online learning, bandits, and reinforcement learning problems. In 2022, ANR endowed the young researcher (JCJC) award for his works on theory of responsible AI. In IJCAI'23, he presented the tutorial on "Auditing Bias of Machine Learning Algorithms: Tools and Overview". In 2024, he got elected as a scholar of European Learning and Intelligent Systems Society (ELLIS). For details, please visit: https://debabrota-basu.github.io/

Speaker: Harish Chavan (CS23S001)

Date: Jun 16, 2026

Time: 11 AM – 12 PM

Venue: MR1 (SSB 233, First Floor)

Abstract:

Blockchain-based payment systems depend on continuous connectivity and incur confirmation latency, making them impractical in intermittently connected environments. TEE-based offline payment schemes enable offline value transfer, but multi-hop forwarding remains inefficient to date, as current techniques for settlement depend on sequential reconciliation of intermediate participants. Addressing this structural limitation is the central motivation of STAMP, a TEE-based offline payment protocol that binds settlement to provenance rather than to intermediate accounts.

This talk builds STAMP up from its two core components. The first is the cheque: a TEE-attested unit of offline value sealed with a genesis triple that preserves an unforgeable binding to the original on-chain owner across every forwarding hop. The second is the user node: each participant's device is partitioned into a normal-world host that handles connectivity and a trusted enclave that owns the keys and the sealed ledger, so no private key or balance ever leaves the TEE. With these in place, we walk through the working of the protocol, initialization, deposit, offline payment and forwarding, and reconciliation, showing how the final recipient of an N-hop chain settles the entire chain through a single contract call, without requiring any intermediate participant to reconnect first.

We then present an evaluation on OP-TEE running on Raspberry Pi device against the closest prior TEE-based offline payment scheme of Jie et al. (IEEE Transactions on Computers, 2024): STAMP outperforms the baseline from N = 4 hops onwards, reaching a 21% end-to-end improvement at N = 32, where the baseline requires N sequential on-chain reconciliation against STAMP's single reconciliation.

Speaker: Prof. Gopal Pandurangan (Houston, USA)

Date: Jun 15, 2026

Time: 3 PM – 4 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Abstract: This talk posits a new quantum framework for designing and
analyzing communication-efficient distributed quantum algorithms.
Using this framework, we present distributed quantum algorithms for
several fundamental problems in distributed computing. Our quantum
algorithms are significantly more message-efficient than their
classical counterparts and surpass classical lower bounds.

Joint work with Fabien Dufoulon (Lancaster University) and Frédéric
Magniez (Université Paris Cité, CNRS, IRIF).

Bio: Gopal Pandurangan is a Moores Professor of Computer Science at
the University of Houston, USA. His research interests include the
theory and algorithms for distributed computing, networks, quantum
computing, machine learning, and big data, and he has over 150
refereed publications. He is a Fellow of the Institute of Electrical
and Electronics Engineers (IEEE), a winner of the 2026 Dijkstra Prize,
a recipient of the University of Houston Research Excellence Award,
and a winner of the ACM PODC 2025 Best Paper Award.

Speaker: Sagnik Mukhopadhyay

Date: Jun 10, 2026

Time: 3 PM – 4 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

We show a few communication protocols for solving fundamental graph problems such as minimum cut, vertex connectivity and bipartite maximum matching. We also mention how these communication protocols help us come up with efficient algorithms for these problems in various other models of computation.

Date: May 25, 2026

Time: 1 PM – 2 PM

Venue: CS 13

Abstract:

Semester End Thanksgiving Lunch and farewell to Jai Ganesh

Speaker: Vinod Shambhu Gupta (CS23S007)

Date: May 21, 2026

Time: 2 PM – 3 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Motivated by application in reducing training data set size in supervised learning, Hart (IEEE Trans. Inf. Theory'68) introduced the problem called Minimum Consistent Subset. We study a more general version of the problem in graph metric, introduced by Banerjee et al. (LATIN'18); coherent with the existing literature, we will directly state this version of the problem and call it Minimum Consistent Subset.

In the Minimum Consistent Subset problem, we are given a connected graph G, an integer k, and a coloring function chi : V(G) rightarrow [c], and the goal is to determine whether there exists a non-empty set vertex subset S of size at most k such that every vertex v in G has at least one closest vertex in S with the same color as v. The problem is already NP-hard on trees (Banik et al., FSTTCS'24), and has received quite significant attention in different algorithmic frameworks designed to cope with NP-hardness.

We study the problem with parameterizations by simple properties of the input graph such as maximum degree (Delta(G)), diameter (diam(G)), treewidth (tw(G))---that captures tree-likeness of the graph; and obtain the following results:

---An FPT algorithm for the (combined) parameter diam(G) + tw(G) +c.
---The problem is W[1]-hard for the (combined) parameter k+Delta(G).

Date: May 20, 2026

Time: 5 PM – 6 PM

Venue: CS Building

Abstract:

IIT Madras - Research Exploration Program (Remote, 4 Weeks, 10 positions)
Work on structured research problems designed with faculty from IITM
🔬 Explore research
🧠 Learn how to interpret research papers + think like a researcher
📊 Build a mini project / literature review
🎤 Present your work for evaluation

Domains : Algorithms | Data Science | ML & AI | Cryptography | Systems Design

Application deadline: 20, May, 2026
Program start : 1, June, 2026

If you’ve been curious about research but didn’t know where to start - this is your entry point.

Speaker: Stephen L (CS22S026)

Date: May 18, 2026

Time: 11 AM – 12 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

The need for articulation-aware representations that explicitly capture the linguistic structure underlying sign formation hampers building effective sign language recognition(SLR) and generation systems especially in signer independent scenarios. This work investigates various encoder strategies to improve Isolated sign language recognition(ISLR) performance across datasets with varied signing conditions. The primary objective is to move beyond conventional appearance-based and pose-based representations, which are sensitive to background variations and signer variations. These factors degrade SLR performance, highlighting the need to explicitly model the articulation dynamics that govern sign formation.We propose a novel articulation-driven encoder that transforms raw pose sequences into a structured feature space defined by manual sign phonological parameters, without relying on RGB frames or raw joint coordinates. This study analyzes the proposed method and compares it with the baseline model across various datasets. We also analyze the effectiveness of our proposed encoder in continuous sign language recognition approaches. Secondly, this work builds a seamless sign generation system by leveraging articulation-aware representations, where sign stitching is performed based on manual sign parameters. We compare the proposed generation system with publicly available sign generation models. Our findings suggest that the proposed articulation-aware representations consistently outperform the baseline models across datasets, demonstrating the effectiveness of explicit articulatory modeling for sign language recognition and generation

Speaker: Umesh Shankar (t-Meet talk)

Date: May 13, 2026

Time: 3:30 PM – 4:30 PM

Venue: CS 36 (CSB)

Abstract:

The classical Ferry Cover problem, inspired by Alcuin’s 8th-century river-crossing puzzles, asks for the minimum boat capacity required to transport the vertices of a graph across a river such that no edge is ever left alone on either bank. Formally, this requires that the banks always induce stable (independent) sets. In this talk, we present a generalization of this problem where the banks must satisfy an arbitrary graph property. We first show that for any hereditary property—such as acyclicity
or planarity—the structural characterizations of “small-boat” and “large- boat” graphs extend directly from existing literature. Our primary focus then shifts to the connected-bank variant, a non-hereditary property where both banks must remain connected subgraphs throughout the entire transfer process.
Our main result is a complete structural characterization of boat-1 graphs which yields a linear-time recognition algorithm. Finally, we initiate the study of boat-n graphs for n ≥ 2, establishing tight bounds for stars and exploring algorithmic results for trees.

Speaker Profile:
Umesh Shankar (https://sites.google.com/view/umeshshankar/home) is a Research Associate in Prof. Sunil Chandran’s Graph theory and Combinatorics lab in the Department of Computer Science and Automation at IISc Bangalore. His research interests lie in Enumerative and Extremal Combinatorics. He completed his PhD from the Department of Mathematics at IIT Bombay where his thesis study was on algebraic properties of polynomials arising in enumerative combinatorics under the supervision of Prof. Krishnan Sivasubramanian. Outside of his thesis study at IITB, he was a part of the Extremal Combinatorics research group headed by Prof. Niranjan Balachandran.

Looking forward to seeing many of you there.
As usual: NO tea/coffee/snacks. Only math(s). ;-)

Speaker: Nency Bansal (CS21D002)

Date: May 11, 2026

Time: 4 PM – 5 PM

Venue: CS 36 (CSB)

Abstract:

Identifying causal relationships and distinguishing them from associations is a central scientific endeavor with many applications; knowing causal links between genes and diseases, for instance, can focus drug discovery on curing diseases beyond just symptom management. Despite several studies on automatically extracting relations between entities from large biomedical literature corpora like PubMed, (i) only a few studies extract causal relations from abstracts, and (ii) even fewer summarize corpus-level evidence for causal links.

For addressing the first research gap, we carefully curated a new CRE Dataset (CRED) of 5618 (causal and non-causal) gene-disease pairs, spanning 204 diseases and 500 genes, across 267 published abstracts with inter-annotator agreement as 89%. To assess CRED’s utility in classifying causal vs. non-causal pairs, we compared multiple classifiers and found SVM (Support Vector Machine) trained on embeddings from a deep learning transformer model called BioBERT to perform the best (Average F1 score 0.70).



For addressing the second research gap, we developed a method to infer Corpus-Wide Causal Score (CWCS) of a gene-disease (G-D) pair by integrating two pieces of evidence: (i) network-based causal signals in a prior gene regulatory network, quantified as a CWCS-Net score using an existing multilayer network centrality algorithm; and (ii) corpus-wide literature evidence, quantified as a CWCS-TD (TD for Truth Discovery) score using a newly-developed TD algorithm. Our CWCS-TD (scoring) algorithm jointly and iteratively estimates causal scores for multiple G-D pairs while modeling the reliability of PubMed abstracts co-mentioning them; and represents an advance in the field of TD algorithms due to its incorporation of bibliometric features of publications to address the challenge of sparsity of abstracts that assert a G-D causal relation. Using OMIM as an external expert-curated reference to evaluate classifications of G-D pairs as causal or not, our CWCS method achieved a causal class F1 score of 0.600 across ten diseases, outperforming both LLMs, GPT-4o and MMed-Llama 3 (this performance trend also persists when using area under the precision-recall curve as the evaluation metric). Both LLMs exhibit high recall accompanied by comparatively low precision, resulting in lower causal class F1 scores (0.505 for GPT-4o and 0.522 for MMed-Llama 3) due to large number of false positive predictions.

Finally, to make these methods practically useful for biomedical researchers, geneticists, and clinicians, we developed a web-based interactive tool for assessing whether a gene is likely to causally influence a disease rather than merely being co-mentioned with it in the literature. The tool supports two complementary modes of analysis, one for each abstract and another for corpus-wide. We plan to apply them to the whole PubMed corpus, comprising approximately 4.5 million abstracts, mentioning at least a gene and a disease.

Overall, these three works contribute a unified framework for biomedical causal relation extraction, moving from abstract-level causal classification to corpus-wide evidence aggregation and interactive exploration. By combining curated annotations, domain-aware causal scoring, and network-guided prioritization, this work represents a meaningful advancement in biomedical CRE and BioNLP for systematic gene–disease causal discovery.

Date: May 11, 2026

Time: 2 PM – 5 PM

Venue: CSB

Abstract:

PhD Interviews for May 2026 Admissions

Date: May 11, 2026

Time: 8 AM – 12 PM

Venue: CSB and DCF

Abstract:

PhD Written Test for May 2026 Admissions

Speaker: Prof. Pandurangan

Date: May 07, 2026

Time: 3 PM – 4 PM

Venue: CS 24

Abstract:

Abstract thinking and problem solving are among distinguishing abilities of the human race. Mathematical thinking (MT) is at the heart of solving real-world problems. Algorithmic thinking (AT) is a specialised style of thinking. It is nothing but mathematical thinking needed to arrive at a computational process ( CT!!) that is efficient and suitable for computer based problem solving..
Amazingly great ideas have emerged during the golden period for Algorithmic thinking. Most great algorithms, when viewed as a finished product, have several puzzling and surprising elements and we feel at the end a bit mysterious about the whole solution. Prim, while writing his Bell labs report has discussed a few insightful points and when we know them, Prim's algorithm becomes a PURE DELIGHT!!!!

Date: May 06, 2026

Time: 2 PM – 5 PM

Venue: CSB

Abstract:

MS / Direct PhD Interviews for May 2026 Admissions

Date: May 06, 2026

Time: 8 AM – 12 PM

Venue: CSB and DCF

Abstract:

MS / Direct PhD Written Test for May 2026 Admissions

Speaker: C Rohin Menon (CS23D402)

Date: May 05, 2026

Time: 9 AM – 10 AM

Venue: MR1 (SSB 233, First Floor)

Abstract:

Side-channel attacks pose a critical threat to the security of cryptographic hardware by exploiting unintended physical leakages, such as power consumption, to recover secret keys. Among the various sources of leakage, glitches, which are spurious transitions caused by mismatched signal arrival times, play a dominant yet often overlooked role. Traditionally treated as design artifacts to be minimized, glitches are widely considered detrimental to both efficiency and security.

In this talk, we challenge this conventional perspective and present GLiTCH, a novel framework that leverages controlled glitch manipulation to enhance side-channel resistance in digital circuits. The key insight is that not all glitches are harmful: while some transitions are correlated with secret data and increase leakage, others act as noise and can obscure sensitive information. By strategically shaping signal arrival times through gate sizing, GLiTCH reduces correlated glitches while amplifying uncorrelated ones, thereby improving resistance to differential power analysis attacks.

The proposed framework combines functional and timing-aware analysis to identify critical leakage points and formulates glitch manipulation as a geometric programming problem. Experimental results on standard cryptographic primitives, including AES, CLEFIA, and SM4, demonstrate significant improvements in security, achieving up to 52.82% increase in guessing entropy with moderate area overheads.

Complementing this approach, we briefly discuss Athena, an analytical framework for fast, pre-silicon estimation of side-channel leakage, which enables efficient identification of timing-induced vulnerabilities during the design phase.

Together, these contributions highlight a new design paradigm: rather than eliminating glitches, carefully engineered timing behavior can be harnessed to improve hardware security.

Speaker: Yuvraj Talukdar (CS23D009)

Date: Apr 30, 2026

Time: 3:30 PM – 4:30 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Graphics Processing Units (GPUs) have become indispensable for high-performance computing, powering everything from deep learning to scientific simulations. However, programming GPUs correctly and portably remains a profound challenge. A fundamental source of this difficulty is that CUDA programs are routinely written with implicit assumptions about the number of threads per block (NTPB) — a hardware-specific parameter that varies across GPU generations and vendors, and can even vary at runtime due to per-kernel resource consumption.

This proposal presents the Unique Worker (UW) Model for CUDA — a principled compiler-driven solution that decouples program logic from hardware-specific thread count parameters. Programs written under the UW model can be automatically transformed to execute correctly on any GPU, regardless of the NTPB that hardware can feasibly provide at runtime. This guarantee, termed semantic portability, is complementary to performance portability: correctness is guaranteed everywhere, while performance can be separately optimized.

The proposed translation scheme is implemented as an AST-to-AST transformation in LLVM 20 and has been evaluated on eleven benchmarks spanning two NVIDIA hardware platforms (Tesla V100 and RTX 3050 Mobile). Key results include: correct execution across all tested NTPB configurations, successful execution of four programs that previously crashed due to resource constraints, and geometric mean optimization gains of approximately 96% over unoptimized translations.

Date: Apr 27, 2026

Time: 2 PM – 3 PM

Venue: MR1

Abstract:

As you know Profs. P Sreenivasa Kumar and C Chandrasekhar have retired from IIT Madras. This semester will be the last teaching semester of their glorious careers. The department is organizing a farewell meeting for them on April 27th at 2pm. Let us share our happy memories/anecdotes good wishes with the retirees and celebrate this occasion.

Speaker: Krishna Sivalingam and Ayon Chakraborty, Moderator: Krishna Nandivada

Date: Apr 24, 2026

Time: 3:30 PM – 4:30 PM

Venue: CS 15

Abstract:

Great news! As per the DCC approved in March 2026, we’re kicking off our CSE FIRE side chat event on April 24th, from 3:30 to 4:15 p.m. This time, we’ll have the amazing Profs. Krishna Sivalingam and Ayon Chakraborthy sharing their insights with us. We’re planning some exciting discussions with these firefighters on stage. We’ll have about 20-25 minutes of moderated Q&A, followed by 5-10 minutes for audience questions. After that, we’ll have some free time for casual chats while enjoying some tasty snacks.

Speaker: Nilesh Anantha Subramanian (CS20D300)

Date: Apr 24, 2026

Time: 3 PM – 4 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

An overarching goal of computational systems biology is to model interactions among bioprocesses such as gene regulation, metabolism, and immune regulation. These models help understand how natural or engineered perturbations affect humans and microbial systems, respectively. Despite the well-established models for healthy and perturbed states, significant gaps persist. For instance, current human studies have not fully characterized the interactions between co-occurring diseases. And to address this gap, we have developed a systematic methodology to quantify the interaction effect of two diseases on molecular markers and clinical traits under a co-morbid condition (Subramanian et al., BMC Bioinformatics 2025). Other significant gaps in modeling microbial systems pertain to modeling two-way feedback interactions between gene regulatory networks (GRNs) and genome-scale metabolic models (GSMMs), and modeling non-linear gene-reaction flux relationships - addressing these gaps, especially the first one, will be the focus of this talk.

Most integrative models of gene regulation and metabolism focus on one-way flow of information from genes to metabolic fluxes, while metabolite-to-gene feedback is often neglected. Towards this, we developed CausalFlux, a framework to capture the bidirectional gene–metabolite feedback to predict steady-state reaction fluxes under wild-type and perturbed conditions. Our proposed algorithm, CausalFlux, combines Bayesian gene regulatory networks with constraint-based metabolic models through iterative causal interventions under a wild-type or gene knockout (KO) condition. Relative to TRIMER, a one-way feedback model, CausalFlux improved flux prediction performance in 92% of testbed simulation settings and achieved a higher balanced accuracy for predicting essential genes in real-world settings in E. coli datasets (0.79 vs. 0.71). Overall, these results demonstrate that CausalFlux can better predict reaction fluxes and growth outcomes by incorporating metabolite feedback into gene regulation. Importantly, further ablation studies also showed that removing key metabolite-to-gene feedback edges reduces prediction performance, thereby highlighting the crucial role of bidirectional feedback in biological systems.

Finally, I will conclude the talk by outlining future directions for exploring the interactions between genes and reaction fluxes in determining yeast growth under gene KO conditions. Through the computational workflows and methods developed in these works, we aim to understand the behaviour of various bioprocesses under perturbations and thereby suggest mechanistic hypotheses and potential treatment strategies.

Speaker: Riya Verma (CS21D012)

Date: Apr 24, 2026

Time: 12 PM – 1 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Continual learning systems must adapt to evolving data distributions while preserving previously acquired knowledge, particularly in realistic settings where data arrives sequentially, classes are imbalanced, and storage of past samples is restricted. This seminar presents a unified overview of four recent research contributions that address these challenges through advances in feature representation, adaptive architectures, and online learning mechanisms. The first work focuses on few-shot class incremental learning, proposing a multi-scale semantic enrichment strategy that extracts hierarchical feature representations from multiple abstraction levels. Combined with a dual angular margin contrastive prototype learning framework and layer-wise feature augmentation, the method improves inter-class separability and enhances generalization to novel classes with limited samples. Addressing the challenge of long-tailed class imbalance, the second study introduces a covariance- controlled feature space augmentation and rectification framework. By leveraging statistical covariance properties to enrich minority class representations and dynamically adjusting classification margins, the
method reduces representation drift and mitigates catastrophic forgetting during incremental learning. The third work introduces ReFACTNet, a recurrent fast-adaptation framework designed for non- exemplar online continual learning. By integrating sparse fast adapter modules, compressed recurrent memory, and causal attention, the approach enables efficient adaptation to new data while maintaining stability and supporting temporal context modeling and backward knowledge transfer. Finally, the fourth work proposes an uncertainty-guided semantic feature adaptation architecture for non-exemplar online class-incremental learning under long-tailed distributions. Incorporating Bayesian temporal class relevance estimation, semantic token scoring, and geometry-aware token alignment, the approach dynamically adapts learning based on class frequency and uncertainty, improving robustness to rare and evolving classes in streaming environments. Together, these contributions provide a unified roadmap for advancing continual learning across exemplar-free, few-shot, long-tailed, online, and uncertainty-aware settings. The proposed methods consistently achieve state-of-the-art or competitive results on benchmark datasets, pushing the boundary of achievable goals in realistic continual learning environments.

Speaker: John Ebenezer Augustine [CSE Bytes Talk]

Date: Apr 22, 2026

Time: 3:30 PM – 4:30 PM

Venue: CSB-25 (Aryabhatta Hall)

Abstract:

Trust is typically viewed as a phenomenon between two individuals: E.g., Aanya trusts Raj because he has always delivered good products to her. This extends to individuals who may not be human. E.g., two AI/software agents may trust each other either because they are programmed to do so or because of established consistency in their interactions. This uncanny parallel between human-to-human interaction and agent-to-agent interaction extends to the macro level. We are bound to observe parallels between trust in human societies and trust in large-scale multi-agent systems. After motivating why such collective trust is needed, I will discuss some important milestones on the study of collective trust from computer science (and a bit of other) research. Importantly, I will discuss why ensuring collective trust is challenging in the modern AI era.

The second half of the talk will be on some concrete ideas towards ensuring healthy collective trust in societies of humans and agents. In particular, I will present the need for (and the challenges in) building a large-scale non-monopolistic digital public infrastructure that can anchor trust between individuals (of both the human and agentic variety), thereby paving the way towards healthy collective trust.

Speaker: Prof. Tushar Jois (CUNY)

Date: Apr 22, 2026

Time: 11 AM – 12:30 PM

Venue: MR1 (SSB 233, First Floor)

Abstract:

This talk focuses on communication challenges during protests disrupted by Internet shutdowns. Prof. Tushar Jois will explain why smartphone mesh messaging has not been widely adopted, highlight the limitations of earlier work, and present Amigo (CCS ’25), showing why cryptography alone is insufficient and why practical low‑level networking research is essential.

Speaker Profile:
Tushar Jois is an Assistant Professor of Electrical Engineering at City College of New York and Computer Science at the CUNY Graduate Center. He earned his PhD from Johns Hopkins University and was a visiting scholar at UIUC and Dartmouth. His research focuses on computer security, cryptography, and privacy. At Spacelab, his group builds deployable cryptographic systems with provable security. His work has been supported by NSF, the Open Technology Fund, Sui Foundation, Google, and CUNY.

Speaker: Naresh Goud Boddu (JPMorgan Chase in Singapore)

Date: Apr 20, 2026

Time: 3 PM – 4 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Quantum information is reshaping our understanding of both computation and security. Quantum algorithms challenge widely used cryptographic assumptions, while quantum systems introduce new adversarial capabilities based on entanglement and measurement. This shift raises foundational questions in cryptography: how do classical notions of security and information evolve in the presence of quantum adversaries?

In this talk, I will present a research program focused on developing cryptographic primitives that remain secure against quantum side information and quantum tampering. I will describe constructions of quantum-secure non-malleable extractors and non-malleable codes, and show how quantum information enables tamper-detecting codes in the split-state model, a task provably impossible in the classical setting. These results demonstrate that entanglement can function not only as an attack resource, but also as a defensive tool that enables fundamentally new forms of protection.

I will also discuss results in quantum communication complexity that clarify fundamental limitations of quantum communication relative to classical communication.

Together, these results contribute to a unified theory of security and information in the presence of quantum adversaries, bridging quantum information theory and modern cryptography.

Speaker Bio:
Naresh Goud Boddu is a researcher in quantum information, working at the intersection of post-quantum cryptography, quantum cryptography, and communication complexity. He most recently served as Vice President and Applied Research Lead at JPMorgan Chase in Singapore, where he led work on post-quantum cryptographic applications for industry. Previously, he was a Postdoctoral Fellow at NTT Research in California in the Cryptography and Information Security Laboratories, mentored by Prof. Vipul Goyal and Prof. Dakshita Khurana.

Naresh received his Ph.D. in Computer Science from the Centre for Quantum Technologies at the National University of Singapore under the supervision of Prof. Rahul Jain. His doctoral research focused on quantum tamper-resilient cryptography and quantum communication complexity. His work develops cryptographic primitives secure against quantum adversaries, including quantum-secure non-malleable extractors, split-state non-malleable codes, and split-state tamper-detecting codes that uniquely leverage quantum phenomena. He has also contributed to understanding structural limitations of quantum communication relative to classical communication.

His work has appeared in leading venues including FOCS, TCC, Asiacrypt, IEEE Transactions on Information Theory, and Quantum, and has been presented at major quantum conferences such as QIP, TQC, QCrypt, and AQIS. He is a recipient of the Best Early-Career Paper Award at Asiacrypt 2025. Naresh has served on the program committee of QCrypt 2024 and will serve on the program committee of Asiacrypt 2026.

Speaker: Sri Saravanan R (CS23S011)

Date: Apr 20, 2026

Time: 3 PM – 4 PM

Venue: CSB-25 (Aryabhatta Hall)

Abstract:

Deep neural networks with increasing depth consistently outperform their shallower counterparts across a wide range of architectures and datasets. However, a principled theoretical explanation for this phenomenon remains at a superficial level. While explanations such as representational power, input space folding, optimization conditioning, the neural tangent kernel and hierarchical feature learning have been proposed, they all have glaring deficiencies. We develop a new narrative that postulates the following. From the viewpoint of any given layer, the other layers act as a lens that applies importance weights to the input, facilitating the emergence of features without global scope that would otherwise be overlooked. Gradient methods simply update the parameters of a layer to best fit the data that is importance-weighted by a lens made of other layers. Individual neuron parameters serve a dual role of `separator for data lensed by other layers' and `component of lens for other layers'. This enables a positive feedback loop that forms the core mechanism of feature learning under gradient descent in deep networks. We support our narrative with theoretical analysis and empirical results on the recently proposed Deep Linearly Gated Network (DLGN), an architecture combining elements of deep linear and ReLU networks. We also show how our analysis leads to a natural notion of hierarchical feature discovery in the feature learning regime.

Speaker: Ryan Gibb (Cambridge)

Date: Apr 16, 2026

Time: 3 PM – 4 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Package managers are legion. Every language and operating system has its own solution, each with subtly different semantics for dependency resolution. This fragmentation prevents multilingual projects from expressing precise dependencies across language ecosystems, meaning external system and hardware dependencies are implicit and unversioned, and obscuring security vulnerabilities that lie in the full dependency graph. We present the Package Calculus, a formalism for dependency resolution that unifies the core semantics of diverse package managers. Through a series of formal reductions, we show how real-world package manager features can be reduced to our core calculus. We define the semantics of translating between distinct package managers and show that we can perform dependency resolution across ecosystems.

Speaker Bio:

Ryan Gibb is a PhD student in the Systems Research Group and Energy and Environment Group at the Computer Lab, University of Cambridge. In his free time, he writes open-source software and uses self-hosted networked services.

Speaker: Vishal S (CS22S010)

Date: Apr 13, 2026

Time: 2:30 PM – 3:30 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

The rapid growth of smart devices across various domains has led to an increase in the collection of personal information from these devices. Such data may be exploited for targeted advertising, personalizing content or for even more intrusive purposes such as identity theft and cyberbullying. Although smart devices offer the users options to manage their privacy settings, each application typically requests multiple permissions and a smart device usually has many applications. Managing these permissions places a heavy cognitive burden on users, making it error-prone. There is also a need for the privacy management to be customizable based on the context in which the user is using the device.

In our work, we provide a comprehensive ontology, PerManOnt, that conceptualizes a context based permission management system aiming to reduce the cognitive burden on users and improve expressivity of the permission decisions of the users while also modelling knowledge about the permissions and applications. PerManOnt is an extension of an existing ontology for permission management and aims to make the system more evolvable. We analyze a real-world dataset and find that 90% of the users' privacy behaviour can be expressed with an average of approximately four preferences per user specified as per the proposed methodology.

Speaker: Rishiyur Nikhil (CTO, Bluspec)

Date: Apr 10, 2026

Time: 2 PM – 3:30 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Property Based Testing of software is well established--- Haskell QuickCheck was the pioneer; Wikipedia lists ports to about 40 other programming languages. In this talk we describe our experience using QuickCheck with _hardware_, to verify "Fife", a 6-stage pipelined RISC-V CPU. We describe the setup for TestRIG (from the University of Cambridge): its use of Haskell QuickCheck for RISC-V test generation and for shrinking; use of the official RISC-V Formal Model written in the Sail ISA Description Language; necessary configuration of TestRIG and the Sail RISC-V Formal Model.

We describe necessary instrumentation for "Direct Instruction Injection" and effect-reporting in any implementation being tested (much of it reusable across implementations). We describe how to deal with a complication: a pipelined CPU's speculative instruction execution, due to branch and PC prediction, and traps. We describe proposed extensions to TestRIG to handle non-deterministic events (such as interrupts) and micro-architectural properties (not shared across implementations). We also briefly describe BlueCheck, an implementation of QuickCheck written in the hardware design language Bluespec BSV, so that BlueCheck can itself be run in hardware (such as an FPGA) along with the CPU design being tested, for much greater testing speed.

Speaker Profile:
Rishiyur Nikhil is CTO and co-founder of Bluspec, Inc., where he promotes high-level languages and methodologies for the design and verification of complex digital hardware. He is a key contributor to the BSV and BH open-source High-Level Hardware Design Languages and their applications, which use ideas from Haskell and atomic rule-based specs. He has created several open-source RISC-V CPU and System designs, and chaired the RISC-V Foundation's technical group that selected the RISC-V ISA formal spec in Sail.

His degrees are from U. Pennsylvania (CS PhD) and IIT Kanpur (EE BTech). His interests and R&D career (at MIT, Digital Equipment Corp. and some startups) span full systems, from formal methods and high-level programming languages to CPUs and system micro-architectures.

Speaker: Brintha V P (CS18D017)

Date: Apr 08, 2026

Time: 2 PM – 3 PM

Venue: Paanini Meeting Room (MR-2, CSB 21, First Floor)

Abstract:

Microbes such as bacteria and viruses inhabit various organisms, including humans, and their states (species composition and function) can play important roles in the health and disease of the host organism. Advances in DNA sequencing technologies and associated computational methods for analyzing sequencing data have revealed the diversity, abundance, and functionality of different microbes in a sample of interest, facilitating personalized healthcare and disease surveillance. But computational challenges remain in analyzing microbial sequencing data, in terms of handling sequencing errors, identifying and quantifying novel bacterial variants, and developing scalable workflows. In this talk, we discuss the design and development of computational methods that address such challenges in the context of three applications related to Tuberculosis (TB), a respiratory disease caused by a bacterial pathogen.

First, we address the problem of delineating different bacterial strains, including novel ones that are mixed in a TB sample of interest. Existing methods for identifying and quantifying mixed infections from microbial sequencing reads either rely on reference databases and miss novel strains or use fully reference-free approaches that cannot resolve strain identities. We present Demixer, a probabilistic generative model that extends the topic model in text mining called latent Dirichlet allocation to jointly model known mutations from reference databases and thereby discover novel strains. We have carefully tuned our method using parallelization and other heuristics to perform efficient multi-sample inference. In both synthetic and real-world datasets, our proposed method precisely determined the identities and proportions of the mixed strains. As Demixer is a shallow bag-of-words based probabilistic model, we also compared Demixer with deep models and found that Demixer is able to better learn the mutational profiles of both seen and unseen strains.

Second, we developed a computational workflow to study the community of microbes during single-disease or double-disease (such as during co-infection of TB and COVID-19) conditions. Specifically, we generated a novel TB+COVID microbial sequencing dataset and systematically analyzed it using a workflow of statistical methods and linear regression models to identify potential biomarkers. Our findings highlight that co-infection with COVID-19 can exacerbate the severity of lung infection in TB patients by altering the microbial composition and its related pathways. Our workflow identified certain key pathogens in patient samples, raising the prospect of personalizing the antibiotic regimen of such patients to improve treatment outcomes.

Finally, we benchmark state-of-the-art approaches for predicting second-line drug resistance in TB. Advances in sequencing have enabled traditional machine learning (ML) and deep learning (DL) approaches to predict drug resistance from genomic data. While first-line drug resistance is predicted reliably, the performance for second-line drugs remains limited. To better understand these gaps, we performed a comprehensive benchmarking of 17 traditional ML and DL models from six studies across 14 second-line drugs using a rigorous evaluation framework. Our results show that shallow ML models outperform complex DL models in predicting resistance to second-line drugs. On an independent dataset, however, shallow ML and DL models perform comparably, with no models achieving substantial improvement over catalogue-based approach. Overall, this benchmarking study suggests using a combination of methods for drug resistance prediction and specifies recommendations for future prediction methods to make robust multi-drug resistance predictions.

Overall, we have developed and applied different computational methods for efficient analysis of microbial sequencing data pertaining to TB, and working with collaborators at the National Institute of Research in Tuberculosis to test the applicability of these methods to help with the Government's mission of TB eradication.

Speaker: Nik Swamy (Microsoft Research)

Date: Apr 07, 2026

Time: 11 AM – 12 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Can AI assist in building programs with formal proofs of correctness? My recent experience using Copilot CLI to develop programs with proofs in F* and Pulse has been eye-opening. In just the past few weeks, working with agents, I have programmed nearly 200 KLOC of verified code and proofs in a variety of contexts, ranging from concurrent data structures, formalizations of classic algorithms textbooks, porting a verified garbage collector from FPLaunchPad, and even some proofs of production C code at Microsoft.

I will speak about my experience and offer some tools and techniques for others to try out similar AI-assisted proof-oriented programming tasks using a fresh consolidated release of F* and related tools, coupled with https://github.com/FStarLang/proof-copilot, agentic tools for proof-oriented programming.

While AI-assistance offers many new opportunities for program proof, I will also offer some thoughts about the many challenges it poses.

Speaker Bio:
Nik Swamy is a Senior Principal Researcher in the RiSE group at MSR Redmond. His work covers various topics including type systems, program logics, functional programming, program verification and interactive theorem proving. He often thinks about how to use these techniques to build provably secure programs, including web applications, web browsers, crypto protocol implementations, and low-level systems code.

Many of his current projects connect somehow to F*, a programming language and verification tool for higher-order, effectful programs. You can learn more about F*, including trying an interactive tutorial, at https://fstar-lang.org(opens in new tab). F* is the main programming language used in Project Everest(opens in new tab), an effort to build and deploy provably secure communication software that he co-leads.

Speaker: Abhishek Ranjan (CS23S036)

Date: Apr 06, 2026

Time: 4 PM – 5 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Code-mixing is pervasive in multilingual Indian communication, yet contemporary machine translation (MT) systems default to formal monolingual output. This work explores, models, and evaluates English to code-mixed translation across eleven Indian languages to better align with how users actually write and speak.

A user preference study revealed a strong preference for code-mixed translations over formal equivalents in everyday communication, firmly justifying this as a primary modeling target. Because traditional surface-overlap metrics are unreliable for this task, we constructed a new benchmark and validated an LLM-as-a-Judge paradigm to ensure robust, human-aligned evaluation.

Benchmarking showed that while massive open-source models perform best, they are impractical for everyday deployment. To close this gap, we developed a teacher-driven synthetic data pipeline to create large-scale parallel corpora featuring native-script, mixed-script, and Romanised text, alongside aligned speech-to-text pairs. Fine-tuning compact models on this data successfully bridged the performance gap, demonstrating the feasibility of deploying highly capable, user-aligned code-mixed translation systems on standard consumer hardware.

Speaker: Dr. Aditya Gangrade

Date: Apr 06, 2026

Time: 9:30 AM – 11 AM

Venue: SSB 233 (MR-1)

Abstract:

In-Context Learning (ICL) is a remarkable phenomenon whereby transformer-based LLMs can use data contained within their prompts to adapt their responses, without changing their weights. This suggests that such models encode learning mechanisms. The recent literature has used statistical learning problems as a test-bed to investigate ICL, and established that ICL can be realised for a wide range of function classes. However, the mechanisms these models use to learn are poorly characterised.

I will describe work on extracting and analysing learning algorithms embedded in the weights of transformers trained to perform ICL in two settings: linear-activation transformers for linear regression, and softmax-activation transformers for linear classification. Through the former, I will illustrate a high-level 'simplify and validate' strategy that allows extraction, and through the latter, I will describe a symmetry-driven strategy for evoking structure in these weights. In these settings, we recover concrete iterative procedures that use existing ideas (Newton-Schulz; mean-shift methods) in new ways that are distinct from gradient descent. Further, we show that transformers trained on variations of these problems implement modified versions of the same dynamics. This suggests that such models recover certain `stable' algorithmic motifs, and adapt them in response to problem structure.

Based on work done jointly with Patrick Lutz, Themistoklis Haris, Arjun Chandra, Hadi Daneshmand, and Venkatesh Saligrama.

Speaker Bio:
Aditya Gangrade is a research scientist at the ECE department at Boston University. He obtained his Ph. D. in Systems Engineering from Boston University, and previously held postdoctoral positions at Carnegie Mellon University and the University of Michigan. His research interests span theoretical and methodological aspects of machine learning, with recent focus on safety in sequential decision making, and on in-context learning phenomena.

Speaker: Bulakh Abhinay Nitin (CS12S024)

Date: Apr 02, 2026

Time: 4 PM – 5 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

This seminar will present a framework for supporting cloud integration directly within the operating system kernel, enabling cloud utilization without modifying applications. For the specific use case of cloud backups and storage, we demonstrate file synchronization policies on a modified Linux kernel. This leverages the Message Filter framework which is a C++ wrapper for Linux kernel

Speaker: Dr. Pritam Sarkar

Date: Mar 30, 2026

Time: 3:30 PM – 5 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Building artificial intelligence with meaningful real-world impact requires models that can understand and interact in both the virtual and physical world. This talk outlines four key capabilities necessary to achieve this: foundational world knowledge, alignment with human values and expectations, reasoning ability, and the capacity to self-improve. The talk begins with an overview of our past contributions toward these capabilities, with a particular focus on building foundational world knowledge in multimodal models and improving their alignment with human values and expectations. The next part of the talk delves into a fundamental challenge in current alignment approaches: the alignment tax. While existing methods aim to make models safer and more reliable, they often degrade general capabilities—reducing response diversity and making models overly cautious or less useful. To address this, we introduce Refined Regularized Preference Optimization (RRPO), a fine-grained alignment method. By penalizing only specific error tokens rather than entire responses, RRPO mitigates harmful behaviors while simultaneously improving performance across diverse vision tasks. Finally, the talk concludes with an overview of ongoing work on enabling stronger visual reasoning and outlines a research vision for developing machines that can adapt and improve over time, fostering long-term usefulness and human–AI trust.

Date: Mar 28, 2026

Time: 9 AM – 7:30 PM

Venue: CSE Department, IIT Madras

Abstract:

CS ExeBit stands as one of India's most anticipated technology festivals, bringing together brilliant minds from across the nation. It's a celebration of innovation, knowledge, and the boundless possibilities of computer science.

With diverse events, engaging competitions, and inspiring talks from industry leaders, ExeBit creates a platform for students to showcase their skills, learn cutting-edge technologies, and connect with fellow enthusiasts and mentors.

Date: Mar 27, 2026

Time: 4 PM – 6 PM

Venue: TTJ Auditorium, IC&SR Building

Abstract:

Inauguration of the Arvind Raghunathan Center for Theoretical Computer Science by Mr. Arvind Raghunathan, Founder, Roc360 and Founter/CEO, House IQ in the presence of Prof. V. Kamakoti, Director, IIT Madras

Speaker: Dr. Sunandita Patra

Date: Mar 24, 2026

Time: 9:30 AM – 11 AM

Venue: SSB 233 (MR-1)

Abstract:

The talk will focus on enabling autonomous actors, such as digital agents or robots, to take deliberative actions towards achieving their long horizon goals, in the face of uncertainty and dynamic events. Today, dynamic events or failures often require human intervention, system restarts, retraining, or redesign, for example, when robots get stuck in dead ends or digital agents collapse under unanticipated events. Existing methods either rely on rule-based reasoning, where it is extremely difficult for human experts to compile and maintain a complete set of rules, or black-box machine learning models that need to be trained extensively for individual tasks. Neither approach is well suited when actors are operating in a dynamically changing environment. This work aims to overcome these limitations by creating integrated planning and learning algorithms that are practical to be executed in real-world environments (spanning finance, robotics and cybersecurity) by incorporating within a single framework: (a) deliberative acting, (b) online planning, and (c) data-driven learning from the actor's experiences. By unifying structured reasoning with data-driven learning, the goal of this research is to push towards the next generation of autonomous systems, general-purpose agentic AI that can plan and act deliberately across a set of diverse tasks and domains.

Speaker: Dr. Ashwin Verma

Date: Mar 23, 2026

Time: 11 AM – 12:30 PM

Venue: SSB 233

Abstract:

Modern distributed systems, from peer-to-peer networks to collaborative learning platforms, operate under fundamental information constraints: communication is limited, network topology evolves, and data sources may be unreliable. In such systems, the information structure is not merely an implementation detail; it fundamentally shapes algorithm design and convergence behavior. This talk presents algorithmic and theoretical contributions organized around two core challenges.
First, I analyze communication-efficient distributed convex optimization over graphs. Standard gossip and consensus-based methods rely on fixed or random communication schedules. I instead study adaptive protocols in which nodes prioritize communication with maximally disagreeing neighbors, inducing state-dependent network evolution. I establish almost-sure convergence over dynamically evolving graphs without relying on spectral gap assumptions, and characterize how adaptive communication affects convergence behavior.
Second, I study learning from unreliable sources without ground truth. In a crowdsourced binary classification setting, workers provide noisy labels with unknown error rates and no labeled validation data. I establish that three workers are necessary and sufficient for parameter recovery, while two workers admit infinitely many indistinguishable models. I then develop an online stochastic-approximation algorithm that provably converges without supervised signals.
Across these problems, a common theme emerges: convergence guarantees depend jointly on objective geometry and the evolution of information constraints. Treating information structure as an algorithmic design variable enables principled approaches to distributed learning under uncertainty.
In the final part of the talk, I briefly highlight additional directions related to resource-aware optimization and learning with endogenous information.

Speaker: Returaj Burnwal (CS21D406)

Date: Mar 20, 2026

Time: 12 PM – 1 PM

Venue: SSB 233

Abstract:

We propose to study the problem of offline safe imitation learning (IL). Here, the agent aims to learn both "what to do" and "what not to do'' by imitating preferred behaviors from heterogeneous expert demonstrations while simultaneously learning to avoid unsafe or non-preferred behaviors from negative demonstrations. Most existing IL approaches rely exclusively on expert demonstrations to guide the agent toward preferred behavior, such a strategy focuses solely on optimal actions and does not explicitly convey unsafe or non-preferred behaviors. As a result, the agent may fail to develop safety awareness, since it has never been trained on what should be avoided. Moreover, non-preferred demonstrations that violate safety constraints are often collected naturally. For example, vehicle accidents and toxic content reported by chatbot users can be used as non-preferred demonstrations. By incorporating these non-preferred demonstrations in the IL framework, the agent can learn not just what to do but also what not to do to remain safe.

Speaker: Rohit Vaish (CSE, IIT Delhi)

Date: Mar 17, 2026

Time: 3 PM – 4 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Ensuring fairness is a fundamental aspect of human nature.
In times when an increasing amount of decision-making is being handed
over to automated systems, it is more important than ever to provide
solutions with provable, mathematically rigorous guarantees. The area
of fair division provides a formal theoretical framework to reason about
such questions in the context of resource allocation.

This talk will focus on the fair allocation of indivisible resources,
which arises in real-world settings such as inheritance division and
university course allocation. Traditional approaches to this problem
involve either randomized allocations that are fair in expectation or
deterministic allocations that are approximately fair. I will discuss an
algorithmic framework that unifies randomization and approximation.
Specifically, I will present an algorithm for finding a randomized
allocation of indivisible goods that is ex-ante envy-free and
ex-post envy-free up to one good. I will also touch upon some
open problems and potential avenues for future research.

Based on joint work with Haris Aziz (UNSW Sydney), Rupert Freeman (University of Virginia), and Nisarg Shah (University of Toronto).

Speaker: Swaprava Nath (CSE, IIT Bombay)

Date: Mar 17, 2026

Time: 11 AM – 12 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Firms (businesses, service providers, entertainment organizations, political parties, etc.) advertise on social networks to draw people's attention and improve their awareness of the brands of the firms. In all such cases, the competitive nature of their engagements gives rise to a game where the firms need to decide how to distribute their budget over the consumers on a network to maximize their brand's awareness. The firms (players) need to optimize the budget allocation to the vertices (consumers) of the network so that the spread improves via direct (e.g., advertisements or free promotional offers) and indirect marketing (e.g., word-of-mouth). We propose a two-timescale model of decisions, where communication between vertices occurs on a faster timescale and the strategy update of firms occurs on a slower timescale. We show that under fairly standard conditions, the best response dynamics of the firms converge to a pure strategy Nash equilibrium. However, such equilibria can be away from a socially optimal one. We provide a characterization of the contest success functions and provide examples for the designers of such contests (e.g., regulators, social network providers, etc.) such that the Nash equilibrium becomes unique and social welfare maximizing. Our experiments show that for realistic scenarios, such contest success functions perform fairly well.

Speaker: Dr. Nikita Yadav (Microsoft)

Date: Mar 13, 2026

Time: 11 AM – 12 PM

Venue: MR1 (SSB 233, First Floor)

Abstract:

Control-Flow Attestation (CFA) is a technique for verifying the integrity of program execution by attesting to its control-flow path. It is particularly useful for auditing execution, allowing one to prove that a program ran as expected or to perform a post-mortem analysis when execution deviates from expected behaviour. Since its introduction in 2016, CFA has primarily been applied in the embedded systems domain. CFA techniques work by instrumenting the prover program to collect runtime control-flow path measurements, which are committed to a Trusted Execution Environment (TEE), such as the secure world of an ARM TrustZone-enabled processor. These measurements are signed by the TEE and later verified by a trusted verifier, who checks whether the program followed the expected path for a given input. However, two key challenges hinder the practical deployment of CFA: (1) scaling to whole-program execution, and (2) supporting complex environments beyond bare-metal systems, i.e., non-bare-metal settings where the prover program runs atop an operating system (OS). In this talk, we will discuss approaches to address these limitations by answering two key questions: (1) What causes the high runtime overhead in existing CFA systems, and how can it be reduced? (2) How can CFA be made secure in non-bare-metal settings where the OS-level adversaries may compromise the system?

Speakr Bio:
Nikita Yadav is a Postdoctoral Researcher in the Future of Scalable Software Engineering group at Microsoft. Prior to this, she completed her PhD at the Indian Institute of Science (IISc), Bangalore, where she worked with Prof. Vinod Ganapathy on problems related to computer systems, with a focus on security, trusted hardware, and proof-of-execution protocols. She previously earned her bachelor’s degree in CSE from the National Institute of Technology (NIT), Kurukshetra and her master’s degree from IISc. She was awarded PMRF in May 2021, and her research projects have spanned diverse domains, including blockchain, decentralized networks, and embedded systems. Her work has been published in top-tier security conferences, including ACM CCS, NDSS and USENIX Security. In 2024, she received the Google PhD Fellowship and was honoured with the 'Murthy Govindaraju Women in Computer Science Research Endowment Award' at IISc.

Speaker: Mr. Vipin Bharathan

Date: Mar 12, 2026

Time: 5 PM – 6 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

The session will cover topics such as:
Post-Quantum Cryptography (PQC) and Crypto
Stablecoins
DeFi and Traditional Finance
CBDC adoption across China, India, and the Euro Zone
Blockchain outlook and trends for 2026

Speaker Bio:
Vipin Bharathan (1982 B.Tech and 1984 M.Tech, IIT Madras) is a thought leader in Web3 and emerging technologies and the founder of dlt.nyc, a consulting practice focused on distributed ledger technology, capital markets, tokenization, and regulation. He is a Forbes contributor on blockchain and digital assets and actively participates in the Hyperledger community as Chair of the Capital Markets Special Interest Group and the Identity Working Group. He has contributed to the Token Taxonomy Framework, defining key token attributes, and authored a paper proposing a novel architecture for Central Bank Digital Currencies (CBDCs). With extensive experience as a developer, development manager, product owner, and technical strategist in financial systems, he has worked with numerous programming languages from Algol and COBOL to Python, Rust, and Go. He also helped build the open-source CBDC prototype “eThaler” in Hyperledger Labs, later adapted into a carbon token and demonstrated to global institutions such as the IMF and the World Bank.

Speaker: Yamini Shankar (CS22D002)

Date: Mar 11, 2026

Time: 11 AM – 12 PM

Venue: SSB 134 (Ground Floor)

Abstract:

Radio Frequency (RF)-based sensing provides a privacy-preserving and non-intrusive alternative to vision-based systems for applications such as human activity recognition, indoor monitoring, and environmental sensing. Recent advances in WiFi sensing enable IoT devices to extract fine-grained physical layer measurements: Channel State Information (CSI), and apply machine learning techniques for perception tasks. However, scaling such sensing systems introduces a critical challenge: the continuous transmission of high-dimensional CSI data from multiple distributed sensors can overwhelm network resources, leading to congestion, increased latency, and degraded performance for regular network traffic.
To address this challenge, we propose AutoCompress, a model-aware framework designed to reduce the network overhead of distributed RF sensing systems. The key idea is to intelligently select only the most informative time-frequency components of CSI spectrograms prior to transmission. AutoCompress leverages a sparsity based technique to compute importance scores that identify the most relevant elements for inferencing. These scores guide a bandwidth-aware sensor-subcarrier selection strategy enabling significant compression of sensing data while maintaining inference accuracy.
AutoCompress is evaluated using the UT-HAR dataset and a real-world Nexmon-based WiFi sensing testbed. The results show that AutoCompress achieves over 4000× reduction in CSI data transmission, improves network throughput by 35.5%, and reduces sensing latency by more than 90%, while preserving sensing accuracy.

Speaker: Soumen Pachal (CS22D009)

Date: Mar 11, 2026

Time: 10 AM – 11 AM

Venue: Paanini Meeting Room (MR 2, CSB 21 First Floor)

Abstract:

In this talk, we discuss a family of generalized Hessian estimators of the objective using random direction stochastic approximation (RDSA) by utilizing only noisy function measurements. The form of each estimator and the order of the bias depend on the number of function measurements. In particular, we demonstrate that estimators with more function measurements exhibit lower-order estimation bias. We show the asymptotic unbiasedness of the estimators. We also perform asymptotic and non-asymptotic convergence analyses for stochastic Newton methods that incorporate our generalized Hessian estimators. Finally, we perform numerical experiments to validate our theoretical findings.

Speaker: Prof. Shweta Agrawal

Date: Mar 09, 2026

Time: 3:30 PM – 5:30 PM

Venue: SSB 134

Abstract:

Cryptography is a beautiful branch of theoretical computer science that seeks to provide guarantees to the art of secret keeping. The questions it poses are fundamental -- does the universe permit asymmetry of computation? Its practical utility requires no argument -- ad-hoc security solutions repeatedly fall prey to attack and crime is increasingly digital. Its scientific charm lies in its deeply paradoxical nature, in its uncanny ability to enable what should really not be possible.


In this talk, we'll take a peek at some of the surprising constructs of cryptography and get a taste of how to walk the tightrope between mathematical structure and computational hardness. We'll see how different mathematical structures have different personalities and will cooperate for some tasks while stubbornly refusing to budge on others. We'll try to approach the exciting, and not always happy, relationship of cryptography with quantum computing.

No background in cryptography will be assumed.

Speaker Bio: Shweta Agrawal is a professor and the Venky Harinarayan and Anand Rajaraman Faculty Fellow at the Department of Computer Science and Engineering at IIT Madras. She works in the area of cryptography, with a focus on post-quantum cryptography. Her work has received over 5000 citations and has appeared at flagship venues in cryptography. She has won multiple awards and honours such as the Rashtriya Vigyan Puraskar (Shanti Swarup Bhatnagar award), Swarnajayanti award, the ACM India award for Outstanding Contributions to Computing by a Woman, a best paper award at flagship conference Eurocrypt, Google India faculty award, BY award for excellence in research and teaching and best reviewer awards at CCS, Eurocrypt, Asiacrypt. She has been an invited speaker at prestigious global conferences like Asiacrypt and “Women in Mathematics”, served as program co-chair for the flagship conference Asiacrypt and serves as associate editor for the "SIAM Journal of Computing", "Journal of Cryptology" and "Information and Computation".
============================

Speaker: Vipul Harsh

Date: Mar 06, 2026

Time: 10 AM – 11:30 AM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Modern Internet-scale services must identify and mitigate customer-impacting incidents quickly. Despite the development of many Root Cause Analysis (RCA) algorithms—including recent LLM-assisted solutions—existing approaches struggle with the "long tail" of novel failure modes and the sheer scale of telemetry. In this talk, I argue that the path forward requires a paradigm shift from developing point-solution algorithms to a systems-first approach. I will introduce MoCE: a Mixture-of-Experts (MoE) framework that provides high-level abstractions for failure diagnosis. This framework enables developers to express complex troubleshooting logic succinctly using MoCE’s domain specific operators while providing the underlying systems support for scalable telemetry processing. Finally, I will briefly describe how these abstractions empower reliable, autonomous agents to perform interactive diagnosis and discuss ongoing and promising future work based on these ideas.

Speaker: Amitabh Trehan (Durham University)

Date: Mar 03, 2026

Time: 2 PM – 3 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

In the field of distributed algorithm design, it is often standard to abstract the network as an undirected graph with the nodes as vertices and connections as edges. About the simplest process one can imagine on a network/graph is flooding: A node is in possession of a message M which has to be eventually sent to every node on the graph (this is called achieving broadcast)- the node sends M immediately to all its neighbours and they send to all their neighbours they did not just receive the message from and so on. Clearly, this achieves broadcast. But, how to achieve termination? i.e. the copies of the message should not circulate indefinitely.

At the advent of distributed computing, more than 50 years ago, a simple solution was devised - keep a copy of M, and if M is received again, simply discard this M. However, this requires memory/state and a stack of earlier received messages. Surprisingly, we discovered [PODC2019,STACS2020,DC2023] that state is unnecessary to achieve terminating broadcast - the same process without any state or memory beyond the immediate receipt (hence, called Amnesiac Flooding (AF)), due to some still slightly mysterious properties of simple undirected graphs, terminates in asymptotically optimal time on every graph. Intriguingly, we have recently discovered [DISC2025] that AF is Unique! i.e. under certain reasonable conditions, AF is the one and only algorithm that achieves terminating broadcast. Are there other examples of Unique algorithms in literature, and is counting the number of algorithms for solving a problem a concept we can reasonably postulate?

AF on Wikipedia: https://en.wikipedia.org/wiki/Amnesiac_flooding

Speaker Bio:
Amitabh Trehan is an associate professor at the department of Computer Science, Durham University, U.K., where he heads the NESTiD (Network Engineering, Science, and Theory in Durham) research group. In the academic year 2025-26, he’s on a sabbatical reconnecting with Indian academia! He is currently a visiting professor in the CSE Department at IIT Madras.

Speaker: All CSE Research Scholars

Date: Feb 28, 2026

Time: 9:30 AM – 7 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Talks (9:30 AM - 12:15 PM) in AM Turning Hall (SSB 334) and Posters (2 PM - 7 PM) in CSB 13

Speaker: G Mahesh (CS22S009)

Date: Feb 26, 2026

Time: 3 PM – 4 PM

Venue: MR1 (SSB 233, First Floor)

Abstract:

Embedded firmware is a critical component of Internet-of-Things (IoT) devices deployed in security-critical domains. Automated vulnerability discovery for such firmware remains challenging due to its tight coupling with the underlying hardware and the absence of source code or detailed specifications. While emulation-based firmware fuzzing has shown promise, existing approaches rely on coarse or heuristic peripheral models that fail to handle complex validation logic, limiting exploration of deep execution paths.
We present FlashFuzz, a fuzzing framework that enables effective exploration of embedded firmware by accurately modeling peripheral dependent control flow. FlashFuzz identifies peripheral input entry points and tracks the propagation of peripheral data through firmware transformations using lightweight taint analysis. By correlating peripheral reads with conditional branches, FlashFuzz selectively extracts and solves constraints imposed by firmware checks to synthesize concrete inputs that steer execution past validation barriers. To maintain scalability, FlashFuzz limits expensive symbolic reasoning and leverages static analysis to infer input sizes and semantic properties.
We implement FlashFuzz on an emulation-based fuzzing platform and evaluate it on 22 real-world firmware binaries, demonstrating improved code coverage and enhanced vulnerability discovery compared to state-of-the-art firmware fuzzers.

Speaker: Chhaya (postdoc at the Indian Statistical Institute, Kolkata)

Date: Feb 26, 2026

Time: 2 PM – 3 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

In the graph stream model of computation, an algorithm processes the edges of an n-vertex input graph in one or more sequential passes while using a memory that is sublinear in the input size. The streaming model poses significant challenges for algorithmically constructing long paths. Many known algorithms that are tasked with extending an existing path as a sub-routine require an entire pass over the input to add a single additional edge. This raises a fundamental question: Are multiple passes inherently necessary to construct paths of non-trivial lengths, or can a single pass suffice? To address this question, we systematically study the Longest Path problem in the one-pass streaming model. In this problem, given a desired approximation factor α, the objective is to compute a path of length at least lp(G)/α, where lp(G) is the length of a longest path in the input graph G.

We give algorithms as well as space lower bound results for both undirected and directed graphs. Besides the insertion-only model, where the input stream solely consists of the edges of the input graph, we also study the insertion-deletion model, where previously inserted edges may be deleted again. Our results include:
1. We show that for undirected graphs, in both the insertion-only and the insertion- deletion streaming models, there are semi-streaming algorithms, i.e., algorithms that use space O(n poly log n), that compute a path of length at least d/3 with high probability, where d is the average degree of the input graph. These algorithms can also yield an α-approximation to Longest Path using space \tilde O(n^2/α).
2. Next, we show that such a result cannot be achieved for directed graphs, even in the insertion-only model. We show that computing a (n^{1−o(1)})-approximation to Longest Path in directed graphs in the insertion-only model requires space Ω(n^2). This result is in line with recent results that demonstrate that processing directed graphs is often significantly harder than undirected graphs in the streaming model.
3. We further complement our results with two additional lower bounds. First, we show that semi-streaming space is insufficient for small constant factor approximations to Longest Path for undirected graphs in the insertion-only model. Last, in undirected graphs in the insertion-deletion model, we show that computing an α-approximation requires space Ω(n^2/α^3).

Speaker Bio:
Chhaya has a PhD in mathematics from the London School of Economics & Political Science. Prior to doing a PhD she has worked in tech Industry for 8 years in various capacities. She is currently doing a postdoc at the Indian Statistical Institute, Kolkata.

Speaker: Nagashri Krishnakumar (IIT Madras)

Date: Feb 24, 2026

Time: 4 PM – 5 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Solving Mazes - more formally, the graph reachability is a fundamental algorithmic problem in space complexity: given a graph and two special vertices $s$ and $t$, the task is to determine if there is a path from $s$ to $t$ in the graph. While the directed reachability is complete for $NL$, the undirected reachability is $L$-complete. The word problem is a fundamental computational problem over a groupoid $(mathcal{A, ullet})$: given a word over the alphabet $mathcal{A}$, the task is to determine if the word evaluates to an element in $mathcal{A}$ w.r.t. the operation $ ullet$. The word problem is undecidable even for finitely presented groups.

A bridge between the Reachability and Word problem: the Labelled Reachability problem for undirected graphs where the edge labels are monoid elements is known to capture space-bounded complexity classes $L$ and $NL$. Given a labelled graph $G(V, E)$ (labelled with $phi: E o M$), $s,t in V$ and an accepting subset $F subseteq M$, the problem asks whether there is a walk $P$ from $s$ to $t$ in $G$ where $phi(P) in F$. When the accepting element is also a part of the input, the problem has been studied by Ramaswamy {em et al} (2019) for the case when the monoid is aperiodic and when it is a group.

Motivated by the success of space-bounded algorithms for the graph reachability in the undirected case, we study the labelled reachability when the accepting set is also fixed. We establish upper and lower bounds for undirected labelled reachability over various monoids and accepting subsets. We give a logspace algorithm for reachability over all monoids when the accepting set is the monoid identity and show $NL$-hardness when the accepting element is a restricted idempotent. We also consider arbitrary accepting subsets over commutative monoids and restricted union of groups monoid to show logspace algorithms, and dichotomies for the building blocks of aperiodic monoids.

Speaker: Dr. Partha Ranganathan

Date: Feb 23, 2026

Time: 10 AM – 11 AM

Venue: DSAI, Room No. 606

Abstract:

Abstract: We are in the middle of an intelligence revolution, with impactful new AI innovations across various sectors from consumer products to enterprise cloud offerings, and even fundamental scientific research. These amazing advances have been made possible by equally impressive progress in computing infrastructure. But as we look at continued exponential growth, the growth, volatility, and diversity of the rapid pace of innovation demands dramatic improvements in compute efficiency, reliability, and sustainability. This talk will discuss disruptive innovations and new opportunities: from “mud to cloud,” from datacenter infrastructure to broader cloud computing services; and from “chip to ship,” from the design of hardware to their deployment and usage in production. Codesign and collaboration - across the hardware-software stack, across disciplines, and across communities - will be key to this exciting new future.

Bio: Partha Ranganathan is currently a VP, Technical Fellow at Google where he is driving next-generation computing systems for the AI-era. He has pioneered numerous impactful innovations touching billions of users, including widely-adopted ideas across all mobile and cloud systems. He has published 100+ papers, is the co-author of the popular Datacenter as a Computer textbook, and is a co-inventor on 125+ patents. He has been recognized with numerous awards; notably: top 35 young innovators in the world by MIT Tech Review, top-15 enterprise technology rock star by Business Insider, IEEE & ACM Fellow, and Rice University & IIT Madras distinguished alumni awards. Fun fact: Partha is one of the few computer scientists to also have an Emmy award!

To mark the occasion, we will be hosting lunch with the guest at BEGH, and we would be happy to have interested faculty members join us

Speaker: Robert Samuel (CS23S042)

Date: Feb 19, 2026

Time: 4 PM – 5 PM

Venue: MR1 (SSB 233)

Abstract:

This work presents DHIR (Distributed Heterogeneous IR), a task-based intermediate representation for compiling programs to modern HPC clusters with multi-core CPUs and accelerators such as GPUs. The core idea is to represent cluster execution explicitly in the compiler IR: programs are decomposed into tasks, tasks are mapped to target devices/nodes, and data movement and synchronization requirements are made visible so the compiler can generate correct distributed code. DHIR is implemented as an MLIR dDHIRialect so it can interoperate with existing MLIR infrastructure. It provides three core abstractions: (1) dhir.task, which encapsulates a unit of work together with explicit read/write memory sets and a device target; (2) dhir.partition, which marks dependence-free loop iteration spaces that can be safely distributed across devices and then lowered into multiple tasks; and (3) dhir.schedule, which acts as the top-level orchestration scope where the compiler builds a task dependency graph (RAW/WAR/WAW), assigns tasks to devices using target information, and inserts necessary communication and synchronization when dependencies cross device boundaries. Using this approach, a single source program can be retargeted across different cluster configurations (CPU-only, GPU-only, or heterogeneous CPU+GPU) by recompiling with a different device/topology configuration, instead of rewriting large amounts of MPI/OpenMP/CUDA code by hand. In our evaluation on representative PolyBench kernels, DHIR-generated code achieves performance close to hand-optimized implementations while substantially reducing developer effort for porting and retargeting.

Speaker: Ajay Singh, ICS FORTH, Greece

Date: Feb 11, 2026

Time: 2:30 PM – 3:30 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Safe memory reclamation (SMR) is crucial to ensure memory safety in optimistic and lock-free concurrent data structures in non-garbage collected programming languages. Yet, designing an ideal SMR algorithm remains challenging. The main difficulties include achieving high speed and scalability, keeping the interface easy to use for programmers, ensuring applicability to a broad class of data structures, bounding the memory footprint, and avoiding asymmetric overhead across data structure operations. Addressing these issues is critical for enabling the deployment of concurrent data structures in production systems.


In this talk, I will present three new methods for safe memory reclamation that are each designed to address a distinct shortcoming of state-of-the-art safe memory reclamation algorithms. Together, these methods demonstrate how codesigning safe memory reclamation algorithms with neighboring layers of the programming environment, including operating system and hardware cache-coherence, can overcome longstanding challenges efficiently.

Speaker Bio:
Ajay Singh obtained his PhD in Computer Science from University of Waterloo in 2024. His Life's goal is to build data structures and algorithms for high performing next generation concurrent systems. He is a recipient of the Distinguished Paper Award at PPoPP 2025, the Best Artefact Award at PPoPP 2021, the David R Cheriton Graduate Scholarship (awarded to top graduate students of the department) 2022. Currently, he is a postdoctoral researcher at Institute of Computer Science (ICS) FORTH in Greece, working with Panagiota Fatourou. Before his PhD he worked as Software Engineer at Samsung Research Institute Noida where he contributed to development of successful feature phones such as Samsung Guru, Samsung Metro, Samsung Guru FM Plus.

Speaker: Dhara Thakkar, Nagoya University, Japan

Date: Feb 06, 2026

Time: 11:15 AM – 12:45 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

In this talk, I will discuss recent progress on the role of classical proofs in quantum complexity theory, focusing on group-theoretic problems in the black-box model. In particular, I will present QCMA protocols for two long-standing open questions in this area. First, we show that the Group Non-Membership problem--known to lie in QMA but not in MA--is in fact in QCMA, settling a 2006 conjecture by Aaronson and Kuperberg. This result follows from a more general result that the Group Order Verification problem lies in QCMA, answering an open question posed by Watrous in 2000. Our techniques also give improved quantum upper bounds on the complexity of many other group-theoretical problems, such as group isomorphism in black-box group settings.

This talk is based on the joint work with François Le Gall and Harumichi Nishimura.

Speaker Profile:
Dhara Thakkar is currently a Designated Assistant Professor at the Graduate School of Mathematics, Nagoya University. She completed her PhD in Theoretical Computer Science at IIT Gandhinagar. Her PhD thesis was awarded the ACM India Doctoral Dissertation Award 2025. Her research interests are in theoretical computer science, in particular, Algebraic Computation, Computational Complexity Theory, Group Theory, Quantum Complexity Theory, Group-based Cryptography, Algorithms and Data Structures, and Parallel Algorithms.

Speaker: Gayathri S (CS23S043)

Date: Feb 06, 2026

Time: 10 AM – 11 AM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Speculative execution is a cornerstone of performance optimization in modern out-of-order processors, yet it has introduced a broad class of microarchitectural side channels. This talk investigates performance-induced security vulnerabilities that arise from aggressive memory disambiguation mechanisms in speculative processors. While most prior defenses focus on speculative data leakage, recent work has shown that the memory ordering logic itself can leak fine-grained structural information through deterministic timing variations, enabling unprivileged adversaries to infer address mappings and amplify other microarchitectural attacks. A central challenge is that these vulnerabilities stem directly from performance-critical speculation mechanisms, creating an inherent challenge in balancing security and efficiency. Traditional mitigation approaches either disable speculation entirely or introduce prohibitive performance overheads, making them impractical for deployment. This thesis presents a lightweight hardware defense that mitigates these attacks by obfuscating deterministic patterns in speculative memory-dependency resolution while preserving the performance benefits of aggressive out-of-order execution. We validate the proposed approach through cycle-accurate microarchitectural simulation and hardware synthesis, demonstrating effective suppression of timing-based leakage with negligible performance degradation and minimal hardware cost. This work contributes both a practical defense mechanism and insights into design principles for reconciling performance-oriented speculation with security requirements in future processor architectures.

Speaker: Anil Kumar (CS18D001) Thesis defense

Date: Jan 29, 2026

Time: 11 AM – 12 PM

Venue: Online

Abstract:

The Santa Claus Problem, also known as the Max-Min Fair Allocation Problem, is a fundamental NP-hard allocation problem in which a set of gifts must be distributed among a set of kids. Each gift provides a kid-dependent happiness value, and the total happiness of a kid is the sum of the happiness values of the gifts assigned to them. The objective is to allocate gifts so as to maximize the minimum total happiness received by any kid. A widely studied variant is the Restricted Santa Claus Problem, where each gift gives a uniform non-zero happiness value to a subset of kids and zero to the rest.

In this work, we study the problem from the perspective of approximation algorithms using nonlinear mathematical programming techniques. We design polynomial time approximation algorithms for the Restricted Santa Claus Problem by combining the kid–gift clustering technique of Bansal and Sviridenko, which we refer to as BS-Clustering, with geometric programming. This approach yields a (12 + ϵ) approximation algorithm, improving upon the previous 36-approximation bound. We further introduce the (τ, α, ϵ)-Santa Claus Problem, in which the input is a quadruple (I, τ, α, ϵ), where I is a Santa Claus instance, τ, α ≥ 1, and ϵ ∈ (0, 1). It is guaranteed that the instance admits an allocation in which each kid has happiness at least τ / α. Using geometric programming, we design a polynomial time algorithm that computes an allocation guaranteeing happiness at least τ / (α+ϵ) for every kid. As a consequence, we obtain a polynomial time (3.534 + ϵ) approximation algorithm for the Restricted Santa Claus Problem, improving upon the previously best known (4 + ϵ) bound.


We also investigate the parameterized complexity of the Restricted Santa Claus Problem with respect to structural parameters of the incidence graph, including treewidth (ω), clique-width (λ), diameter (δ), and the target happiness value (τ ). We show that the problem is W[1]-hard under the combined parameters (ω, λ, δ), even when the incidence graph is planar, and under (τ, λ, δ). On the positive side, we establish fixed-parameter tractability of the general Santa Claus Problem
under the combined parameters (ω, τ ) and (m, τ ), where m is the number of kids.

Speaker: Nikhil Shinde (CS23S018)

Date: Jan 29, 2026

Time: 10 AM – 11 AM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Programmable data-plane (PDP) switches enable flexible packet processing
but remain constrained in scheduling capabilities, particularly for
priority-based policies that require strict ordering of packets
according to user-defined ranks. Push-In First-Out (PIFO) provides an
ideal abstraction for programmable packet scheduling; however, its
requirement for line-rate packet sorting makes it impractical at high
network speeds (100G and beyond). PIFO is a priority queue that enables
ordering enqueued packets based on their priority. Alternate accurate
schedulers like SIFTER, involve complex packet movement between the
scheduling queues. Approximate schedulers such as SP-PIFO and AIFO, RIFO
reduce complexity but introduce priority inversions, degrading latency,
fairness, and flow completion times (FCTs) as they rely on First-In
First-Out (FIFO) queues or use Active Queue Management (AQM) as a
substitute to scheduling. To address the scalability limitations of
accurate schedulers while avoiding the correctness issues of approximate
schedulers, this work proposes two architectures -- called P3PO and
PPIFO -- that are scalable, inversion-free and parallelize the sorting
workload across multiple priority queues.

P3PO and PPIFO decompose a large global priority queue into multiple
smaller, bounded-capacity priority queues that operate in a parallel
manner. P3PO uses a cascaded set of priority queues with priority-based
bounds to ensure that packets are inserted into the earliest queue
capable of maintaining ordering; overflow is handled by cascading
evictions into lower-priority queues. Under-utilization of a queue
triggers promotion of packets from lower priority queues to higher
priority queues to maintain availability for timely dequeue. PPIFO
distributes packets across multiple independent priority queues based on
occupancy, using a de-multiplexer for balanced enqueue and a multiplexer
for globally selecting the highest-priority packet at dequeue. Both
designs ensure bounded sorting complexity, preserve strict priority
ordering, and eliminate priority inversions.

The architectures are implemented in the NetBench packet-level
simulator, and evaluated using empirical datacenter workloads
(Web-search and Data Mining) under in-cast and all-to-all traffic
patterns. Experiments are conducted on star and leaf-spine topologies
using Start-Time Fair Queueing (STFQ) and pFabric-style ranking. Results
show that P3PO achieves fairness and flow completion time (FCT)
performance comparable to ideal PIFO and performs better than
approximate schedulers. We show that P3PO and PPIFO have less packet
comparisons as compared to ideal PIFO, thus reducing the sorting
latency. For flows under 2 MB, P3PO improves FCTs by 14–62%, and for
flows exceeding 2 MB by 47–60% as compared to approximate
schedulers. Packet drops are reduced by 66–95% due to improved queue
utilization and elimination of inversion-driven tail latencies against
approximate schedulers.

Speaker: Ritwik Badola (CS22M016)

Date: Jan 28, 2026

Time: 11 AM – 12 PM

Venue: MR1 (SSB 233, First Floor)

Abstract:

This talk investigates the security posture of small office/home office (SOHO) internet-connected Wi-Fi routers and IP-camera devices through an end-to-end supply chain lens by analysing device firmware. Using a multi-vendor corpus and automated scanning, we quantify vulnerability and patch lag, map operating system kernel baselines to publicly known kernel issues and fixes, and use limited lifecycles and market signals to contextualise whether affected devices remain present in the ecosystem. Across vendors, including brand-new devices procured off the shelf in 2025, we find that many devices run significantly outdated operating system kernels with large backlogs of publicly disclosed vulnerabilities (CVEs) present in them, and often bundle additional third-party components (TPCs) that are similarly old and vulnerability-prone.


A central finding is that these devices vendors are effectively locked to specific kernel versions by system-on-chip (SoC) SDKs; Wi-Fi router and IP-camera makers often remain on older kernels to retain chipset support and SoC vendor-provided proprietary software components. This produces a vulnerability debt that is inherited along the supply chain from SoC vendors to device manufacturers to device vendors and ultimately borne by end users. We further show that kernel upgrades are possible and plausible on supported platforms by contrasting with OpenWRT-based maintenance models, highlighting how open-source driver enablement can reduce upgrade cost and friction. We present a platform and SDK-aware analysis framework, report empirical findings across vendors, and discuss lifecycle and policy gaps, along with limitations and areas requiring further validation, and what these results imply for improving updates and reducing long-lived risk at in these devices at the network edge.

Speaker: Sivaramakrishnan K C [CSE Bytes Talk]

Date: Jan 27, 2026

Time: 3:30 PM – 4:30 PM

Venue: CSB-25 (Aryabhatta Hall)

Abstract:

OCaml is a 25+ year-old, industrial-strength functional programming language widely used in industry and academia. OCaml is particularly favoured for its ease of writing correct and efficient programs. Notable open-source projects using OCaml include the Rocq theorem prover, MirageOS Unikernel libraries, and static analysis tools and programming languages such as Hack, Flow, Infer, CompCert, etc. Some notable industrial users include Jane Street, Meta, Ahrefs and Docker. Despite its age, the OCaml community remains strong and continues to grow, bringing cutting-edge programming language research and language tooling to its users without compromising critical legacy code. In this talk, I reflect on how the OCaml programming language evolves, focusing on its community and also shed some light on how you may contribute to the OCaml development.

Speaker: Ashesh Ashesh (Milan Italy)

Date: Jan 22, 2026

Time: 2:30 PM – 3:30 PM

Venue: SSB 233 (ALC MR1)

Abstract:

Fluorescence microscopy is limited by optics, fluorophore chemistry, and photon exposure,
forcing trade-offs in speed, resolution, and depth. In this talk, I will discuss my PhD research
that addresses these challenges. Specifically, my PhD research enables imaging of multiple cellular structures within a single fluorescent channel, allowing faster imaging with less photon exposure. Technically speaking, given a superimposed image (e.g., containing nucleus and tubulin), the objective is to predict the constituent images separately.

This talk focuses on my first work, μSplit [1]. Early in my PhD, we found that regular deep architectures performed best with large image patches, but GPU memory limits hindered scalability. We thus developed μSplit, a novel meta-architecture enabling memory-efficient use of large image context. Built on Hierarchical-VAE (HVAE) and U-Net variants, it modifies HVAE’s ELBO loss for non-autoencoding tasks, modifies KL loss for high-frequency details extraction, and reformulates the encoder output for stable training. We also created a synthetic dataset to evaluate our network’s capability to extract large image context. Lastly, we explored tiling artifacts, analyzed two mitigation strategies, and demonstrated the superiority of one, both empirically and via out-of-distribution arguments.

References: [1] Ashesh Ashesh, Alexander Krull, Moises Di Sante, Francesco Pasqualini, and Florian Jug. μSplit: Image decomposition for fluorescence microscopy. ICCV, 2023.


Speaker Bio:
Ashesh is a postdoctoral fellow at Human Technopole, Milan, Italy. He recently completed his PhD in Computer Science at TU Dresden, Germany, conducted in Florian Jug’s lab at Human Technopole’s Computational Biology Center.
His doctoral research focused on image decomposition, specifically unmixing superimposed fluorescence microscopy images into constituent channels. With first-author publications in top CV/ML venues such as ECCV, ICCV, and NeurIPS, and a recent one accepted to Nature Methods, his work offers a robust solution to this challenge. His thesis earned a nomination for TU Dresden’s PhD prize nominations (pending decision), an €8,700 EMBO grant for a research visit to ENS de Lyon on self-supervised finetuning and uncertainty quantification, and the Best Oral Presentation Award at the 2024 HT PhD & Postdoc Symposium.
Previously, Ashesh earned a dual B.Tech+M.Tech in Computer Science (2015) from IIT Delhi,
India. He brings over three years of industry experience as a Data Scientist and served as Research Assistant at National Taiwan University under Prof. Hsuan-Tien Lin, initiating multiple computer vision projects, culminating in publications.

Speaker: Chaitanya Agarwal (New York University)

Date: Jan 19, 2026

Time: 2 PM – 3 PM

Venue: SSB 134

Abstract:

Authenticated data structures (ADSs) allow untrusted third parties to carry out operations which produce proofs that can be used to verify an operation’s output. Such data structures are challenging to develop and implement correctly. In this talk, I will talk about programming frameworks that exploit some structural properties of ADSs to make the task of correctly implementing ADSs easier: (1) lambda-auth - a domain-specific extension of the OCaml programming language, that generates correct and secure ADS binaries automatically; (2) Authentikit - a library version of lambda-auth currently implemented in OCaml. I will also talk about recent work by us (https://dl.acm.org/doi/abs/10.1145/3719027.3744801) that gives a formal proof of security and correctness of Authentikit.

Speaker Bio: Chaitanya Agarwal (https://culechetoo.github.io) is a 3rd year computer science PhD student at the New York University, advised by Joseph Tassarotti. He is broadly interested in programming languages and formal verification with a particular focus on verification of security applications. In the past, he has also worked with Thomas Wies on abstract-interpretation analysis for recursive, higher-order programs, and with Shibashis Guha, on developing statistical-model-checking techniques for Markov Decision Processes (MDPs). Chaitanya obtained his B.Tech. from IIIT Delhi.

Speaker: Naresh Goud Boddu (VP, JP Morgan Singapore)

Date: Jan 14, 2026

Time: 10 AM – 11 AM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Tamper-detection codes (TDCs) are fundamental objects at the intersection of cryptography and coding theory. A TDC encodes messages in such a manner that tampering with the codeword causes the decoder to either output the original message or reject it. In this work, we study quantum analogs of one of the most well-studied adversarial tampering models: the so-called t-split-state tampering model, where the codeword is divided into t shares and each share is tampered with locally. It is impossible to achieve tamper detection in the split-state model using classical codewords. Nevertheless, we demonstrate that the situation changes significantly if the message can be encoded into a multipartite quantum state entangled across the t shares. Concretely, we define a family of quantum TDCs on any t ≥ 3 shares, which can detect arbitrary split-state tampering so long as the adversaries are unentangled, or even limited to a finite amount of pre-shared entanglement. Previously, this was only known in the limit of asymptotically large t. As our flagship application, we show how to augment threshold secret sharing schemes with similar tamper-detecting guarantees. We complement our results by establishing connections between quantum TDCs and quantum encryption schemes.

Speaker Bio:
Naresh Goud Boddu is Vice President and Applied Research Lead at J.P. Morgan Singapore. He completed his Ph.D. (Thesis) in September 2022 under Prof. Rahul Jain at the Centre for Quantum Technologies, National University of Singapore. Following his Ph.D., he was a Postdoctoral Researcher at NTT Research, where he worked with Prof. Vipul Goyal and Prof. Dakshita Khurana until November 2024. At J.P. Morgan, he has worked on lattice-based zero-knowledge proof applications for financial institutions. His research interests lie at the intersection of post-quantum cryptography, quantum cryptography, and quantum information. More information is available at https://sites.google.com/view/ngboddu/home

Speaker: Pratap Singh (Carnegie Mellon University)

Date: Jan 13, 2026

Time: 11 AM – 12 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Cryptographic security protocols, such as TLS or WireGuard, form the foundation of a secure Internet, but vulnerabilities are discovered in them with alarming frequency. Formal verification promises a foundational solution to this problem, with significant prior work on mechanizing cryptographic proofs of high-level protocol designs, and more recent work on verifying particular implementations of single protocols against higher-level specifications.

In this talk, I will discuss our work on Owl, a verifier and secure compiler for cryptographic protocols. Owl's verifier uses a novel combination of information-flow and refinement types to prove security in the computational model, enabling greater modularity and automation than prior computational tools. Owl's compiler translates well-typed protocols into performant, interoperable, side-channel-resistant Rust libraries that are automatically verified to be correct. With Owl, developers can define and prove security for their protocols in an intuitive, high-level language, and obtain for free a drop-in executable implementation with formal guarantees of correctness and security.

We evaluate Owl using a range of case studies. We verify the core logic of 14 protocols, including SSH and Kerberos. We additionally develop two large-scale, interoperable case studies for WireGuard and Hybrid Public-Key Encryption (HPKE), yielding verified implementations that interoperate with, and match the performance of, existing industrial baselines on end-to-end benchmarks.

Speaker Bio:
Pratap Singh is a fourth-year PhD student in Computer Science at Carnegie Mellon University, advised by Bryan Parno. His research leverages formal methods techniques to produce provably secure systems, exemplified by the Owl tool for verified cryptographic protocol implementations. He received a bachelor's degree from Harvard University in 2021, where he worked with Stephen Chong on compiler correctness proofs and Adam Chlipala on the Bedrock2 verified compiler. He is supported by the NSF Graduate Research Fellowship, and his research has received the Distinguished Artifact Award at USENIX Security.

Speaker: Ms. Srividhya S (CS23S024)

Date: Jan 13, 2026

Time: 11 AM – 12 PM

Venue: MR1 (SSB 233, First Floor)

Abstract:

Dynamic pricing in high-dimensional markets poses fundamental challenges of scalability, uncertainty, and interpretability. Existing low-rank bandit formulations learn efficiently but rely on latent features that obscure how individual product attributes influence price. We address this by introducing an interpretable Additive Feature Decomposition-based Low-Dimensional Demand (AFDLD) model, where product prices are expressed as the sum of attribute-level contributions and substitution effects are explicitly modeled. Building on this structure, we propose ADEPT (Additive DEcomposition for Pricing with cross elasticity and Time-adaptive learning) - a projection-free, gradient-free online learning algorithm that operates directly in attribute space and achieves a sublinear regret of O (√\uD835\uDC51\uD835\uDC47^{3/4}). Through controlled synthetic studies and real-world datasets, we show that ADEPT (i) learns near-optimal prices under dynamic market conditions, (ii) adapts rapidly to shocks and drifts, and (iii) yields transparent, attribute-level price explanations. The results demonstrate that interpretability and efficiency in autonomous pricing agents can be achieved jointly through structured, attribute-driven representations.

Speaker: Prof. Venkatesan Guruswami (University of California, Berkeley)

Date: Jan 07, 2026

Time: 11 AM – 12 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

We show that any binary linear code of block length  n that is locally correctable with 3 queries against a constant fraction of adversarial errors must have dimension at most $O(\log^2 n \log \log n)$. This nearly matches the known construction of 3-query locally correctable codes (LCCs) based on quadratic Reed–Muller codes, which have dimension $\Theta(\log^2 n)$. Our result significantly improves the best known upper bounds in the binary case: the $O(\log^8 n)$ bound of Kothari and Manohar (STOC 2024), later improved to $O(\log^4 n)$ by Yankovitz (FOCS 2024).
Prior work bounded the dimension of 3-query LCCs by reducing them to 2-query locally decodable codes and invoking strong known bounds for the latter. Our approach is more direct and proceeds by bounding the covering radius of the dual code, borrowing inspiration from (Iceland and Samorodnitsky, 2018). Specifically, we show that in any 3-query LCC with encoding map $x \mapsto (v_1 \cdot x, \dots, v_n \cdot x)$, every binary vector of length k can be expressed as a $\tilde{O}(\log n)$-sparse linear combination of the $v_i$'s. This then yields the desired dimension bound via a simple counting argument.

The crux of our proof is a procedure that compresses any sufficiently large linear combination of the $v_i$’s into a sparser one, by exploiting the abundance of short linear dependencies amongst them. This step hinges on a recent result of Alon, Bucić, Sauermann, Zakharov, and Zamir (2023) on the existence of rainbow cycles in properly edge-colored graphs. We apply this theorem in a black-box manner to graphs that capture the combinatorial structure induced by the local correction constraints.

Joint work with Omar Alrabiah.

Speaker Bio:
Prof. Venkatesan Guruswami is the Chancellor's Professor at the Dept of EECS, Professor at the Dept of Mathematics, and the Director of the Simons Institute for the Theory of Computing at UC Berkeley. Dr Guruswami's research interests span several topics in theoretical computer science, including the theory of error-correcting codes, the approximability of fundamental optimisation problems, explicit combinatorial constructions, pseudorandomness, probabilistically checkable proofs, computational complexity theory, and algebraic algorithms.

Dr. Guruswami currently serves as the Editor-in-Chief for the Journal of ACM and has in the past served as the Editor-in-Chief for the ACM Transactions on Computation Theory. Dr. Guruswami is a recipient of the Presburger Award (2012), Packard Fellowship (2005), Sloan Fellowship (2005), NSF CAREER award (2004), the ACM Doctoral Dissertation Award (2002), and the IEEE Information Theory Society Paper Award (2000).

Speaker: Prof. Jeffrey Ullman (Stanford University)

Date: Jan 05, 2026

Time: 2:30 PM – 3:30 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

We shall introduce four algorithms that run very fast on large amounts of data, although typically the answers they give are approximate rather than precise. (1) Locality-sensitive hashing (2) Approximate counting (3) Sampling (4) Counting triangles in graphs.

Speaker: Sharat Ibrahimpur (Faculty Candidate)

Date: Jan 05, 2026

Time: 11:30 AM – 12:30 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Uncertainty is ubiquitous in real-world problems, highlighting a growing need for new mathematical and algorithmic tools for optimization under uncertainty. In this talk, I will discuss two prominent hurdles that arise in stochastic combinatorial optimization, along with recent advances that address these challenges.


The first setting considers a generic combinatorial optimization problem in which the costs are random variables with known distributions, and feasible solutions induce multi-dimensional cost vectors. The algorithmic goal is to compute an oblivious solution that minimizes the expected norm of the associated cost vector for a given monotone, symmetric norm. One can show that optimizing the norm of the expected cost vector may yield solutions whose approximation quality degrades with the dimensionality of the problem. I will present a powerful theorem, dubbed approximate stochastic majorization, that enables constant-factor approximations for norm-based objectives across a wide class of combinatorial optimization problems.


I will then turn to stochastic decision environments where forgoing adaptivity for simplicity or analytical convenience can lead to poor approximation guarantees. I will illustrate this through a novel variant of two-stage stochastic load balancing that enables a quantitative trade-off between the practical benefits of non-adaptive algorithms and their performance limitations. I will present a striking power-of-two-choices result for stochastic load balancing, showing that constant-factor approximations are achievable with limited adaptivity relative to the omniscient and adaptive optimums.

Speaker Bio:
Sharat Ibrahimpur (https://i-sharat.github.io/) is a postdoctoral researcher in the Computer Science department at ETH Zurich. Previously, he held postdoctoral positions at the University of Bonn and the London School of Economics. Sharat earned his M.Math. and Ph.D. degrees in Combinatorics and Optimization from the University of Waterloo. He completed his undergraduate degree in Applied Mathematics at the Indian Institute of Technology Roorkee. Sharat's research focuses on designing approximation algorithms for combinatorial optimization problems, with an emphasis on optimization under uncertainty.

Speaker: Vaidy Sivaraman (Ohio University)

Date: Dec 26, 2025

Time: 10 AM – 11 AM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Several topics connecting graph structure and graph invariants, particularly chromatic number, will be discussed. Several open problems and special cases of famous conjectures like Erdos-Hajnal conjecture will feature in this talk.

Speaker Bio:
Vaidy Sivaraman received his PhD in 2012 from The Ohio State University with Prof. Neil Robertson as his supervisor. He is currently an associate professor of mathematics at Mississippi State University. His research interests are in structural graph theory, mainly induced subgraphs, graph coloring, and hereditary graph classes. He is a recipient of the Simons Collaboration grant for mathematicians in 2021, and has been the dissertation supervisor of two PhD students. He has posed several problems and conjectures, some of which have been solved/resolved.

Speaker: Sugyani Mahapatra (CS18D015)

Date: Dec 22, 2025

Time: 11 AM – 12 PM

Venue: SSB 426

Abstract:

A graph representation of a complex system such as a biological or social system is often not a single "true" network, but a family of distinct yet related networks that together represent our knowledge of the system's interactions. For instance, multiple gene-gene networks could be learnt using observations of the same cell line from different time points, conditions, or bootstrap resamplings. Network science measures of nodes in such an ensemble of graphs could yield important insights about the underlying system's behaviors and dynamics.

Given this setting of multiple input graphs, each defined over the same set of nodes but having distinct yet similar/overlapping sets of edges, we can ask the following broad questions:
(1) Can we design algorithms to compute a network science measure (like node centrality) on this ensemble of graphs more efficiently than the naive approach of analyzing each input graph separately?
Such algorithms are feasible if the input graphs are similar to each other so that they can be organized hierarchically into a directed tree (arborescence), and the measure can be computed using dynamic/incremental algorithms. We have designed and earlier presented parameterized/approximation algorithms for the related problem of finding Steiner arborescence in a hypercube.
(2) Can we identify concrete applications (e.g., in biology) wherein ensemble analysis of multiple graphs leads to demonstrably robust insights?
Identifying hub genes in a gene-gene coexpression network is one such application, wherein we show that ensemble-based centrality analysis of genes can lead to gene rankings that are robust to uncertainty. In this talk, I will expand on this application by proposing an uncertainty-aware ranking of genes derived from an ensemble of centrality estimates, and comparing this ranking's performance to traditional centrality based rankings. In several simulated and real-world gene networks, across different evaluation measures pertaining to recovery of hub genes and tissue-specific pathways, our proposed gene rankings show noticeable performance gains over baseline rankings that do not incorporate uncertainty. Our ensemble analysis is more compute intensive though, and we are currently exploring if the hierarchical framework of organizing graphs described in (1) above can be used to improve running times in practice.

Overall, we present a concrete methodology that enables analysis of a collection of similar graphs inferred from measurements of a biological system to derive a stable prioritization of genes. As such, this work paves the way for similar integrative analyses of other system-level datasets in the future.

Speaker: Vignesh Manoharan (UT Austin, USA)

Date: Dec 19, 2025

Time: 3 PM – 4 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Routing messages is an important problem in distributed networks, and efficient fault-tolerant routing requires methods to maintain shortest distances between nodes when faults occur. In this talk, I will introduce fault-tolerant versions of shortest path problems in graphs, and discuss algorithms and lower bounds for these problems in the "CONGEST" model of bandwidth-limited distributed computation. Specifically, I will present results for the Replacement Paths problem and the Distance Sensitivity Oracle problem, both of which deal with computing shortest paths in a graph when a single edge fails.

Speaker Bio: Vignesh Manoharan did his B.Tech (DD) in 2019 from IIT Madras. He finished his PhD from the University of Texas at Austin. His research interests include graph algorithms and fault tolerance, especially in parallel and distributed models of computation.

Speaker: Jayant Haritsa (IISc, Bengaluru)

Date: Dec 18, 2025

Time: 3 PM – 4 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

When complex SQL queries suffer slow executions despite query optimization, DBAs typically invoke automated query rewriting tools to recommend ``lean'' equivalents that are conducive to faster execution. The rewritings are usually achieved via transformation rules, but these rules are limited in scope and difficult to update in production systems. We investigate here how the remarkable cognitive capabilities of LLMs can be leveraged for performant query rewriting while incorporating safeguards and optimizations to ensure correctness and efficiency. Our study shows that these goals can be progressively achieved through incorporation of (a) an ensemble suite of basic prompts, (b) database-sensitive prompts via redundancy removal and selectivity-based rewriting rules, and (c) LLM token probability-guided rewrite paths. Further, a suite of logic-based and statistical tools can be used to check for semantic violations in the rewrites prior to DBA consideration. We have implemented the above LLM-infused techniques in the LITHE system, and evaluated complex analytic queries from standard benchmarks on contemporary database platforms. The results show significant performance improvements for slow queries, over both SOTA rewriters and the native optimizer. Overall, LITHE is a promising step toward viable LLM-based advisory tools for enhancing enterprise query performance.

Speaker: Shivika Narang

Date: Dec 17, 2025

Time: 11 AM – 1 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

We initiate the study of fair distribution of delivery tasks among a set of agents wherein delivery jobs are placed along the vertices of a graph. Our goal is to fairly distribute delivery costs (modeled as a submodular function) among a fixed set of agents while satisfying some desirable notions of economic efficiency. We adopt well-established fairness concepts -- such as envy-freeness up to one item (EF1) and minimax share (MMS) -- to our setting and show that fairness is often incompatible with the efficiency notion of social optimality. We then characterize instances that admit fair and socially optimal solutions by exploiting graph structures. We further show that achieving fairness along with Pareto optimality is computationally intractable. We complement this by designing an XP algorithm (parameterized by the number of agents) for finding MMS and Pareto optimal solutions on every tree instance, and show that the same algorithm can be modified to find efficient solutions along with EF1, when such solutions exist. The latter crucially relies on an intriguing result that in our setting EF1 and Pareto optimality jointly imply MMS. We conclude by theoretically and experimentally analyzing the price of fairness.

Joint work with Hadi Hosseini and Tomasz Was.

Speaker Bio: Shivika Narang is a postdoc at UNSW working with Haris Aziz and Toby Walsh, since March 2024. During Fall 2023, she was a postdoc at MSRI/SLMath in the Algorithms, Fairness and Equity program. She completed her PhD at the Indian Institute of Science, Bengaluru, advised by Prof Y Narahari. There, she was a recipient of the Tata Consultancy Services (TCS) Research Scholarship. Her current work is focused on fairness and efficiency in computational social choice, especially matching and allocation problems.

Speaker: Santhosh Raghul

Date: Dec 12, 2025

Time: 3:30 PM – 4:30 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

This presentation comprises two parts; these correspond to the courses CS6999 and CS7999, respectively. In the first part, we shall discuss a couple of celebrated results per-taining to cubic (that is, 3-regular) graphs from the literature, whereas the second part shall focus on our contributions to further the understanding of cubic graphs

Speaker: Dr. Kashyap Rajeevsarathy (IISER Bhopal)

Date: Dec 11, 2025

Time: 3:30 PM – 4:30 PM

Venue: CS 36 (CSB)

Abstract:

There are several infinite graphs associated with surfaces that combinatorially encode the simple closed curves on them. These graphs play a central role in the study of symmetries, topology, and the geometry of surfaces. In this talk, we will introduce one such graph, namely the curve graph, and some of its derivative graphs. We will discuss some basic properties of these graphs and their importance in the field of geometric topology.

Speaker Bio:
Dr. Kashyap Rajeevsarathy is a Professor and the Head of the Department of Mathematics at the Indian Institute of Science Education and Research (IISER) in Bhopal. He specializes in geometric topology, focusing on the mapping class groups of surfaces, while also having a secondary interest in algebraic graph theory. He earned his Bachelor's degree in mathematics from Vivekananda College in Chennai before pursuing his Master's degree at IIT Madras. He completed his Ph.D. in Mathematics at the University of Oklahoma in 2011, under the supervision of Professor Darryl McCullough. After completing his doctorate, he joined IISER Bhopal. He was awarded the Ramanujan National Fellowship from the Department of Science and Technology in 2011.

Speaker: Daniel Paul-Pena (University of California, Santa Cruz)

Date: Dec 09, 2025

Time: 3 PM – 4 PM

Venue: SSB 233 (MR1)

Abstract:

Subgraph and homomorphism counting are fundamental algorithmic problems. Recent work (Bera et al. SODA 2021, JACM 2022) provided a characterization of the patterns that can be counted in linear time for bounded degeneracy inputs. An older result by Nesetril and Ossona de Mendez proved that for input graphs in classes with bounded expansion all the patterns can be counted in linear time. What lies between? We present a hierarchy of dichotomies that closes the gap between both results, completely and precisely characterizing the instances that are linear time computable. The result uses ideas from graph sparsity, generalized notions of tree decompositions, and techniques from subgraph counting in sparse graphs. This is a joint work with C. Seshadhri.

Speaker Bio: Daniel Paul-Pena is a 5th year PhD Candidate at the University of California, Santa Cruz, under the supervision of Prof. C. Seshadhri. His main research interest is in subgraph counting, with special emphasis in finding algorithms that translate well into practice.

Speaker: Mr. S R Swamy Saranam Chongala (PhD Research Scholar, CS14D403

Date: Dec 08, 2025

Time: 11 AM – 1 PM

Venue: MR1

Abstract:

Modern applications depend on architectures that effectively exploit data locality, but growing working-set sizes have outpaced the scalability of SRAM-based caches. As a result, DRAM-based last-level caches are increasingly used, despite challenges such as large tag storage, higher access latency, and greater power consumption. This thesis addresses these limitations by proposing three techniques that leverage temporal, spatial, and algorithmic locality across the memory hierarchy and heterogeneous SoCs. First, we propose tagless DRAM cache (TDC) to mitigate the overhead of the tag array in DRAM caches and reduce average access latency. Our technique exploits the spatial locality exhibited by applications to design efficient large LLCs with minimal metadata overhead, while improving access latency and power consumption. In the second technique, we propose novel cache management strategies that decouple L3-L4 replacement decisions to reduce inter-cache interference. We introduce energy-efficient mechanisms tailored for tagless last-level caches (LLCs): restricted block caching (RBC) and victim tag buffer caching (VBC). These mechanisms incorporate L4 eviction penalties into L3 replacement logic with minimal overhead. Finally, building on the insights from locality exploitation in caches, we propose micro-architectural changes to efficiently harness algorithmic locality in SoCs with heterogeneous compute. Our technique aims to optimize data movement between CPUs and NPUs. We introduce a hybrid processing unit (HPU), a novel architecture that tightly integrates NPUs and CPUs. The HPU combines the strengths of both processors to improve data movement and performance for AI applications.

Speaker: Dr Tabish A. Syed (McGill University)

Date: Dec 08, 2025

Time: 10 AM – 11 AM

Venue: SSB 233

Abstract:

The rhythmic pumping of the mammalian heart arises from the intricately coordinated contraction of cardiomyocytes. However, the detailed three-dimensional organization of these cells has long been difficult to resolve. In this talk, I will introduce an integrated framework that applies computer vision methods to 3D confocal microscopy data to reconstruct cardiomyocyte orientation and geometry at micrometer resolution; an improvement of roughly three orders of magnitude over conventional approaches. I will show how this pipeline maps individual cardiomyocytes into coherent myofibers and further into higher-order laminar sheet structures that define ventricular architecture. By estimating orientation fields directly from volumetric microscopy data, we produce a comprehensive, multiscale reconstruction of myocardial organization at the cellular scale. Our work offers new insights into the structural basis of cardiac mechanics and establishes a foundation for high-resolution, data-driven models of heart function.

Speaker Bio:
Tabish A. Syed, is a Postdoctoral Fellow at the Centre for Neuroscience at McGill University Health Center-Research Institute. He earned his PhD in Computer Science from McGill University, specializing in computer vision with a focus on biological image analysis. His research focuses on developing well grounded explainable computational methods for biological image analysis. In the past his work has been featured on the cover of the EMBO Journal. He is particularly interested in geometry and applied math approaches for image analysis. His current work focuses on leveraging computer vision and machine learning techniques for uncovering the role of astroglia in freely behaving in mice, specifically, he is interested in exploring the roles of astrocytes in learning and behaviour.

Speaker: Nitin Saurabh (IIT Hyderabad)

Date: Dec 04, 2025

Time: 2:30 PM – 3:30 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

A polynomial p(x) is said to approximate a Boolean function f:{0,1}^n -> {0,1} if mod(f(x)-p(x)) <= 1/3 for all x in {0,1}^n. The approximate degree of f, denoted adeg(f), is defined to be the minimal degree of an approximating polynomial for f. It was introduced roughly 30 years back in a seminal work of Nisan and Szegedy (Computational Complexity 1994). Since then it has become a powerful and versatile tool in theoretical computer science finding spectacular applications in circuit complexity, quantum query complexity, communication complexity, computational learning, and differential privacy, to name a few. In a beautiful and significant work, Sherstov (Theory of Computing 2013) showed that adeg(f o g) = O(adeg(f)adeg(g)), where f o g is obtained by substituting variables of f by independent copies of g. The approximate degree composition problem posits that this upper bound is tight. That is, adeg(f o g) = Omega(adeg(f) adeg(g)) for all f and g. I will present a recent work of mine (along with my collaborators) on this problem. Furthermore, I will also present an overview of my research work and future plans.

Speaker Bio: Nitin Saurabh joined IIT Hyderabad CS Dept as a faculty member in 2022. Prior to that, he spent a brief stint at IIIT Hyderabad. Prior to that, he was a postdoctoral fellow at Max Planck Institut für Informatik, Saarbrücken, Germany and Charles University, Prague, Czech. He completed his PhD at the Institute of Mathematical Sciences (IMSc) in 2016 and did a BSc (Honours) in Mathematics and Computer Science from CMI in 2010.

Speaker: Kanchan Kumar Tiwari (IIT Madras)

Date: Nov 28, 2025

Time: 10 AM – 12 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

This work will present DATURA, a Deep Learning(DL) based Adaptive Traffic-aware Unified Resource Auto-scaling framework for virtual network functions (VNFs) based slicing in 5G/6G networks. DATURA integrates a deep reinforcement learning (DRL) based traffic generator with a predictive GRU-MLP scaling architecture, forming a closed-loop, traffic-adaptive, and latency-aware scaling system. The framework addresses two key practical challenges that persist in existing predictive scaling solutions: the lack of realistic, dynamically varying traffic environments for evaluation, and the limited adaptability under non-stationary workloads. At the core of DATURA lies a Gated Recurrent Unit combined with a Multi-layer Perceptron (GRU-MLP) scaling controller, which forecasts per-VNF and per-slice latencies from streaming telemetry. Temporal dependencies are captured through the GRU layer, while two multitask MLP heads estimate local (per-VNF) and global (slice-level) latency trends. Once deployed, DATURA operates fully online, executing inference-driven scaling decisions in real time. The underlying models are periodically retrained with updated telemetry, enabling the framework to adapt to long-term traffic evolution without service disruption. The predicted latencies are then translated by a regression layer into proactive scaling decisions and resource allocations that maintain SLA compliance while avoiding unnecessary over-provisioning. This design replaces manual threshold tuning and online optimisation with a lightweight inference process that introduces negligible decision latency. Finally, a reactive safeguard layer ensures stability and robustness under unforeseen workload deviations. The DRL traffic generator employs Proximal Policy Optimisation (PPO) agents trained to learn transition probability matrices (TPM) that capture realistic transitions between low, medium, high, and very-high-load regimes. This enables the generation of controllable and bursty traffic traces that emulate real-world network fluctuations and stress-test scaling algorithms under reproducible conditions. The generator supports flexible composition of traffic patterns, enabling controlled mixes of load conditions that reflect realistic multi-slice behaviour and expose elasticity limits under contention. The framework is implemented in a SimPy-based 5G Core Network model for supporting UE registration and PDU session establishment, capturing control-plane dynamics. Detailed experiments for diverse traffic intensities show that DATURA consistently outperforms several existing schemes, achieving up to 73% latency reduction, 55% throughput gain, and 47% higher provisioning efficiency. These results demonstrate that DATURA provides a unified, traffic-aware, and data-driven solution for predictive and scalable VNF management in softwarised 5G/6G infrastructures.

Speaker: Dr. Akshima (NYU Shanghai)

Date: Nov 20, 2025

Time: 11 AM – 12 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

In the preprocessing Decisional Diffie-Hellman (DDH) problem, we are given a cyclic group G with a generator g and of prime order N. We want to prepare some advice of size S, such that we can efficiently distinguish (g^x,g^y,g^xy) from (g^x,g^y,g^z) in time T for uniformly and independently chosen x,y,z from Z_N. This is a central cryptographic problem whose computational hardness underpins many widely deployed schemes, such as the Diffie-Hellman Key exchange protocol. Corrigan-Gibbs and Kogan [CK18] and then Coretti et al. [CDG18] studied lower bounds for this problem for generic algorithms. However, both the works left a gap from the best known (generic) attack for the problem. In our STOC24 paper, we close this gap up to poly-log factors and consequently show that DDH is asymptotically as hard as Discrete Logarithm problem for preprocessing algorithms. The talk will include some preliminary definitions, detailed description of the models, results and a high level idea of the techniques from the joint work with Tyler Besselman, Siyao Guo, Zhiye Xie and Yuping Ye, available at https://eprint.iacr.org/2024/1171.pdf.

Speaker: Dr. Siddharth Mehrotra (University of Amsterdam and Delft University of Technology (TU Delft))

Date: Nov 13, 2025

Time: 2:30 PM – 3:30 PM

Venue: Online

Abstract:

Trust is a cornerstone of any society. Nowadays, many AI systems are being developed that have the potential to make a difference in people’s lives, from health apps to robot companions. However, to reach their potential, people need to have appropriate levels of trust in these AI systems, i.e., people should not over- or under-trust AI as it can lead to misuse and disuse. Therefore, AI systems need to understand how humans trust them and what to do to promote appropriate trust. To investigate this question, in this talk I will adopt three lenses namely: a formal, a social and an application lens to ensure a holistic exploration of appropriate human trust, drawing on formal theories, social considerations, and practical insights. At first, we will delve into the technical intricacies by understanding the logic of building appropriate trust. Second, we will study the ways in which humans interact with and rely on AI systems and finally, we will apply these findings to a real-world context of predictive policing in the Netherlands by taking a deep dive. At the end, I will very briefly introduce some of my recent works on insurance as an interaction mechanism for trust calibration, findings on what is lacking in FAccT & AIES community to understand AI trustworthiness and designing a LLM digital companion for young adult informal caregivers.

Speaker: Rini Sharon (Uniphore)

Date: Nov 10, 2025

Time: 11 AM – 12 PM

Venue: SSB 233

Abstract:

This talk explores machine learning methods that bridge human cognitive signals and machine understanding. It covers my work on decoding speech and EEG to recognize imagined speech through multi-phasal modeling, improving decoding speed and generalizability in brain–computer interfaces. The second half extends into building scalable, multilingual speech systems, including but not limited to adaptive ASR, multimodal abuse detection, semantic endpointing in real-world conversational settings.

Speaker: Siddharth Priya (University of Waterloo)

Date: Nov 04, 2025

Time: 3 PM – 4 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Ensuring the correctness of systems software is critical, but traditional testing offers no formal guarantees. Formal verification provides stronger assurance, and bounded model checking has proven to be a particularly powerful approach. Yet, despite its promise, widespread adoption in everyday software development remains limited. This talk explores why — focusing on two core challenges: bridging the gap between developer intent and the design of verification tools, and advancing automated reasoning to scale verification to industrial programs. The speaker will share strategies for improving usability and demonstrate how enhanced reasoning can make formal verification more practical, scalable, and developer-friendly.

Speaker: Mukulika Maity [CSE Bytes Talk]

Date: Oct 29, 2025

Time: 3:30 PM – 4:30 PM

Venue: CSB-25 (Aryabhatta Hall)

Abstract:

In this talk, I will talk about the newest HTTP protocol, HTTP/3. It runs on top of QUIC as the transport protocol. I will briefly explain how QUIC works. We will then talk about how the most popular streaming application, YouTube, works and its interaction with the QUIC protocol. QUIC runs on top of UDP, and a few firewalls tend to block or rate-limit UDP traffic; therefore, the browsers ensure compatibility by enabling connection racing via simultaneously initiating a TCP connection with the QUIC one. Each time the QUIC protocol suffers, connection racing is activated, and whichever protocol wins the race is further used for the application. In this talk, I will present how browsers implement this connection racing mechanism and analyze its impact on YouTube. We perform a large-scale measurement study across different browsers (Chrome/Chromium and Firefox), which helps to analyze why and how the repeated connection racing between protocols affects adaptive streaming QoE over 6265 YouTube sessions covering 5647 hours of streaming. Interestingly, we observe that the imperfection in the implementation of this connection racing in popular browsers impacts YouTube's QoE, preventing QUIC from attaining its full potential. I will present the modifications we made in the browser to improve YouTube QoE. This talk will highlight how subtle design flaws in browser implementations, specifically in handling connection racing, can lead to significant degradation in application QoE

Speaker: Jagdeep Singh (Mississippi State University)

Date: Oct 28, 2025

Time: 4 PM – 5 PM

Venue: Online

Abstract:

Complement reducible graphs, or cographs, form one of the most well-understood hereditary classes of graphs: they admit linear-time recognition, possess multiple elegant characterizations, and have a simple forbidden subgraph description (no induced path of length three). In this talk, I will describe two natural generalizations of cographs; 2-cographs and sesquicographs. I will then present results showing that apexing (adding a single vertex or edge) preserves the finiteness of forbidden induced subgraphs, along with explicit bounds on their size. Finally, I will extend this framework from graphs to binary matroids, introducing comatroids as the natural matroid analogue of cographs. Analogous extension results hold in this setting. This work includes joint work with James Oxley (Electron. J. Combin., SIAM J. Discrete Math.), Thomas Zaslavsky and Vaidy Sivaraman (Discrete Math.) LINK: meet.google.com/agg-kgrr-ycp

Speaker: Dr. Kankana Roy (Karolinska Institute, Sweden)

Date: Oct 27, 2025

Time: 3:15 PM – 4:30 PM

Venue: SSB23

Abstract:

In this presentation, we will explore the problem of driver’s cellphone usage detection in vehicles for driver inattentiveness assessment. Current datasets in the literature, which are widely used in this domain to build and test models to be deployed in the real world, often suffer from class distribution bias, frequently overrepresenting inattentive behavior in the dataset to aid the supervised models. This imbalance may limit the generalizability of such models in practical, real-world driving conditions, where mobile usage is a relatively rare event. To address this challenge, we propose a novel approach using unsupervised dictionary learning by treating the problem as anomaly detection rather than supervised classification. This work illustrates one approach, combining hand detection, feature engineering, dictionary learning, and real-world applicability through a novel idea that reduces reliance on biased labels. Our results demonstrate comparable detection accuracy to fully supervised methods reported in the literature while offering better robustness to dataset imbalance. Overall, our study offers a practical and computationally efficient alternative to conventional supervised pipelines for driver inattentiveness detection, emphasizing generalizability and real-world feasibility.

Speaker: Pappu Kumar (CS20D015-IITM)

Date: Oct 24, 2025

Time: 3:30 PM – 4:30 PM

Venue: SSB-334 (A M Turing Hall)

Abstract:

Leader election (LE) is one of the most fundamental and well-studied problems in distributed computing, with applications ranging from classical to modern systems. The obvious lower bounds on time and message complexity for explicit leader election are $Omega$(1) and $Omega$(n) on complete networks, and $Omega$(n) and $Omega$(n) on ring networks, respectively, where n is the number of nodes in the network. A long line of research has aimed at achieving bounds that are tight with respect to these lower bounds. In search of theoretically interesting results, the algorithms proposed in the past tend to be complex, hard to implement, and often hide large constants under asymptotic notation (e.g., big O), making them impractical in most scenarios. In this paper, we propose significantly simpler and easier-to-implement randomized leader election algorithms for asynchronous complete networks and asynchronous ring networks. We prove correctness for all our proposed algorithms. To study their performances, we simulated our algorithms along with several key algorithms from the literature using the NS-3 simulator, and conducted systematic performance comparison on large networks of about a million nodes. Our algorithms achieve less time and messages in our experiments than established LE algorithms. web conference link https://meet.google.com/yee-azeo-ujf

Speaker: Divyarthi Mohan (Boston University)

Date: Oct 24, 2025

Time: 2 PM – 3 PM

Venue: MR1

Abstract:

Digital markets and platforms have shaped the algorithmic landscape into a complex ecosystem of strategic, self-interested entities. This has motivated the study and development of mechanisms or algorithms that are robust to strategic behaviour, using tools from algorithms, game theory and economics. Standard assumptions in mechanism design are too strong to capture the informational challenges present in many real scenarios, from ad auctions where bidders' values depend on competitors' private market data, to resource allocation where there is uncertainty about future demands. In this talk, I will provide an overview of my recent work that tackles three important challenges—strategic behavior, interdependence, and online decision making—going beyond standard assumptions. In particular, I will focus on my work establishing the first constant-approximation algorithms for prophet and secretary problems with interdependent values.

Speaker: Robert Samuel (IIT Madras)

Date: Oct 23, 2025

Time: 10 AM – 11 AM

Venue: CS26

Abstract:

MLIR (multi-level IR) is making waves due to its applicability in multiple domains, and so much in the ML area that some people call it machine learning IR. It is considered to be a new avatar of LLVM. Robert, whom you know as a TA for our course, has been working on MLIR and has developed a dialect for distributed heterogeneous computing. He is going to teach us the basics of MLIR in this guest lecture in CS6843 Program Analysis.

Speaker: Srinivas Narayana [CSE Bytes Talk]

Date: Oct 15, 2025

Time: 4 PM – 5 PM

Venue: CSB-25 (Aryabhatta Hall)

Abstract:

Extended Berkeley Packet Filter (eBPF) has emerged as a powerful method to extend the capabilities of the Linux operating system kernel. eBPF allows users to write code in high-level languages (like C or Rust) and attach them for event-triggered execution at specific execution hooks in the kernel. Among many others, examples of prominent hooks include the network device driver, process scheduler, and the system call dispatcher. To ensure the safe execution of a user-developed eBPF program within the kernel, Linux uses an in-kernel static code analyzer. The analyzer allows a program to execute within the kernel only if it can formally prove that the program is crash-free, always accesses memory within safe bounds, and avoids leaking sensitive kernel data for any input. eBPF has found production-grade deployment is all the hyperscalars as well as several major computing organizations, with some companies anecdotally running dozens of programs on every single server in their fleet. Typical use cases include fleet-wide observability and profiling, flexible and efficient container networking, high-speed connection load balancing, and system call security.
This talk will introduce eBPF in concept, provide an overview of the eBPF static code analyzer---the cornerstone of safe kernel software extensibility, and show a short demonstration with real eBPF programs. Time permitting, I will introduce recent research from my group on foundational scientific questions surrounding this technology.

Speaker: Sreejith K. Pallathumadam

Date: Sep 30, 2025

Time: 3:30 PM – 4:30 PM

Venue: CSB 34

Abstract:

A total coloring of a simple undirected graph $G$ is an assignment of colors to its vertices and edges such that the colors assigned to the vertices form a proper vertex coloring, the colors assigned to the edges form a proper edge coloring, and the color of every edge is different from that of its ends. It is easy to verify that $G$ needs at least $Delta + 1$ colors for any total coloring of $G$ where $Delta$ is the maximum degree of $G$. Surprisingly, Behzad (emph{PhD Thesis, Michigan State University, 1965}) and Vizing (emph{Russian Mathematical Surveys, 1968}) independently conjectured that $Delta + 2$ colors are sufficient. Note that a triangle needs only $Delta + 1$ (i.e. three) colors for a total coloring while a 4-cycle requires $Delta + 2$ (i.e. four) colors. Interestingly, the conjecture is still open. Given a total coloring of a graph $G$, for each vertex $u$, let $colors(u)$ denote the set comprising the color assigned to $u$ and the colors assigned to the edges incident with $u$. A total coloring of a graph $G$ is emph{Adjacent Vertex Distinguishing (AVD)} if for each $uv in E(G)$, $colors(u) neq colors(v)$. The emph{AVD Total Coloring Conjecture} of Zhang, Chen, Li, Yao, Lu and Wang (emph{Science in China Series A: Mathematics, 2005}) states that every graph $G$ admits an AVD total coloring using at most $Delta + 3$ colors. It is easy to verify that a triangle requires $Delta + 3$ (i.e. five) colors for an AVD total coloring. For $s in mathbb{N}$, a graph $G$ is $s$-degenerate if every subgraph of $G$ has a vertex of degree at most $s$. For example, forests are 1-degenerate (what about the converse?), outerplanar graphs are 2-degenerate, triangle-free planar graphs are 3-degenerate and planar graphs are 5-degenerate. Miao, Shi, Hu and Luo (emph{Discrete Mathematics, 2016}) proved the AVD Total Coloring Conjecture for 2-degenerate graphs. We recently proved the conjecture for 3-degenerate graphs; in this talk, we shall discuss this result.

Speaker: Ashrujit Ghoshal [CSE Bytes Talk]

Date: Sep 17, 2025

Time: 3:30 PM – 4:30 PM

Venue: CSB-25 (Aryabhatta Hall)

Abstract:

In order to build secure and efficient cryptography, we constantly lean on two pillars of theoretical computer science: algorithmic techniques and complexity-theoretic insights. In this talk, I will explore the algorithmic side via two themes of research that I have worked on.

In the first part of the talk, I will focus on Private Information Retrieval (PIR), a cryptographic mechanism that allows users to query databases without revealing their access patterns. Beyond its compelling potential applications such as private browsing and private DNS, PIR constructions often exhibit remarkable mathematical elegance and simplicity. I will define PIR, present one such elegant construction, outline some of my recent results, and end this part with a discussion of exciting open problems in the area.

In the second part of the talk, I will turn to the analysis of cryptography. I will discuss how limitations on resources of attackers are modelled formally, and how these limitations inform the design choices of deployed cryptographic systems. I will conclude with an example-driven exposition of one of the simplest to describe yet surprisingly powerful tools in this area: compression arguments, an algorithmic technique that transforms the task of proving an impossibility result into an algorithmic question!

Together, these themes illustrate how algorithmic thinking enriches cryptography, providing both tools for construction and frameworks for analysis.

No prior background in cryptography is required. Just bring your curiosity (and perhaps a healthy suspicion of eavesdroppers)

Speaker: Nishad Kothari [CSE Bytes Talk]

Date: Sep 10, 2025

Time: 3:30 PM – 4:30 PM

Venue: CSB-25 (Aryabhatta Hall)

Abstract:

The study of perfect matchings (aka 1-factors) goes back all the way to the Four Color Conjecture (1852) and related developments in the late 1800s due to Tait (1880) and Petersen (1891). The former established that the Four Color Conjecture is equivalent to the statement that every 2-connected 3-regular planar graph is 3-edge-colorable (or, in other words, has three disjoint perfect matchings), whereas the latter proved that every 2-connected 3-regular graph has a perfect matching (that is, a spanning 1-regular subgraph).

One of the earliest breakthroughs was Tutte’s 1-factor Theorem (1947) that gave a complete characterization of matchable graphs — that is, graphs that have a perfect matching.

Using Tutte’s Theorem, one may easily deduce a strengthening of the aforementioned result of Petersen: in a 2-connected 3-regular graph, each edge belongs to some perfect matching. This leads to one natural direction of research that is discussed below. For a graph G, an edge e is solitary if it belongs to precisely one perfect matching. For example, each edge of K4 is solitary, whereas no edge of the Petersen graph is solitary.
Recently, Goedgebeur, Mattiolo, Mazzuoccolo, Renders and Wolf (http://arxiv.org/abs/2402.08538, 2024) proved that every 3-connected 3-regular graph has at most six solitary edges, and they gave a complete characterization of all such graphs that have three or more solitary edges.

In a joint work with a recently graduated MS scholar D. V. V. Narayana, Davide Mattiolo, and a former intern Kalyani Gohokar, we (http://arxiv.org/abs/2409.00534, 2025) generalize their results to 3-edge-connected r-graphs; when r=3, this is precisely the class of 3-connected 3-regular graphs. Our main tool is the dependence relation introduced by Carvalho, Lucchesi and Murty (Combinatorica, 1999).
In this talk, I will discuss all of the above in as much detail as possible within the timeframe of one hour. No prior background, except for basic graph theory knowledge, is required.

Speaker: Rupesh Nasre [CSE Bytes Talk]

Date: Aug 13, 2025

Time: 3:30 PM – 4:30 PM

Venue: CSB-25 (Aryabhatta Hall)

Abstract:

Graph algorithms are notorious; more than our students. They pose challenges toward analysis and parallelization. In this discussion, we will uncover limitations of compiler-based approaches and understand why graph algorithms need dynamic techniques. We will also discuss various challenges posed by different computing hardware (e.g., CPU, GPU, distributed systems) and specialized graphs such as hierarchies and GNNs.

Overall, I plan the discussion to pose more questions than answers. See you there.

Prerequisites: curiosity or urge for coffee or both

Speaker: Gopalakrishnan Srinivasan [CSE Bytes Talk]

Date: May 15, 2025

Time: 3:30 PM – 4:30 PM

Venue: CSB-25 (Aryabhatta Hall)

Abstract:

Spiking Neural Networks (SNNs), widely regarded as the third generation of artificial neural networks, offer a plausible solution to approaching the brains’ processing capability for cognitive tasks. With biologically realistic perspective on input processing, SNN performs neural computations using spikes in an event-driven manner. The spike-based asynchronous computing capability can be exploited to achieve improved energy efficiency in hardware implementations. In this talk, I will describe our work on different SNN training methodologies, namely, ANN-SNN conversion and spike-based error backpropagation using surrogate gradient for the discontinuous spiking nonlinearity. I will subsequently present our recent findings on the enhanced fault-tolerance of SNNs to bit-flip attacks on model parameters.