Theory for Machine Learning

In this group, we are interested in theoretical aspects of algorithms that relate specifically to machine learning and related domains such as vision. Some of the research problems consist of designing of better optimization models, solvers and provable guarantee of convergence. Learning part of machine learning often corresponds to minimizing a suitable loss function defined for a labelled data. Hence, optimization methods lies at the center of machine learning research in theory and algorithms.

We are interested in all aspects of optimization methods and solvers such as constraint optimization solvers, including optimization on manifolds, stochastic optimization, minmax problems, unconstrained optimization problems. We are currently interested in wide-ranging applications in machine learning and vision such as 3D reconstruction, image and video completion, fast and robust training for deep learning, robust algorithms for generative models, robust training for extreme classification, robust solvers for optimization problems in reinforcement learning, etc.

Active Members:

• Kinal Mehta
• Tanmay Kumar Sinha

Selected Papers:

• “A Deflation Based Fast and Robust Preconditioner for Bundle Adjustment”, Das, Shrutimoy and Katyan, Siddhant and Kumar, Pawan, Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) pp. 1782-1789, 2021

• “Two-Grid Preconditioned Solver for Bundle Adjustment”, S. Katyan and S. Das and P. Kumar, 2020 IEEE Winter Conference on Applications of Computer Vision (WACV) pp. 3588-3595, 2020.