This talk presents novel formulations for learning shared feature representations across multiple tasks. The key idea is to pose the problem as that of learning a shared kernel which enables good generalization in all the given tasks. Firstly, a formulation which extends the framework of multiple kernel learning to the case of multiple tasks and learns a shared kernel as sparse combination of a given set of base kernels is presented. The formulation is then modified to learn sparse feature representations which are shared across multiple tasks. The latter formulation can also be understood as an extension of a state-of-the-art multi-task sparse feature learning formulation for the case of multiple base kernels. Efficient non-Euclidean gradient descent based algorithms for solving the formulations are also presented. Simulation results on real-world datasets show that the proposed formulations generalize better than state-of-the-art. The results also illustrate the efficacy of the non-Euclidean gradient descent based algorithms.
BioSaketha Nath is an assistant professor in the dept. of CSE at IIT-Bombay. His main research interests are in machine learning and convex optimization. He is currently working on kernel learning related algorithms for various ML applications.