Amartya Sanyal is a doctoral student in the Dept. of Computer Science, University of Oxford jointly supervised by Dr. Varun Kanade and Dr. Phil Torr. He graduated with a B.Tech in Computer Science And Engineering with a minor in Linguistic Theory in the top 7% of his class from the Indian Institute of Technology, Kanpur in the year 2017 where he was a student of the Department of Computer Science And Engineering where he worked on projects in non-convex optimization, deep learning and extreme learning. He has interned in the Montreal Institute of Learning Algorithms under Prof. Yoshua Bengio and Twitter Cortex NYC among others where he has worked on developing generalized sequence learning models and adversarial models for natural language modelling.
Deep neural networks have revolutionized the field of machine learning and their use has improved on the prior state of the art significantly in diverse domains such as vision, natural language processing, speech, reinforcement learning, etc. Optimization methods used in the training of these networks can be somewhat ad hoc and lack a unified understanding, primarily because of the non- convex nature of these optimization problems. The work will focus on developing new methods for training networks that involves both changes to optimization methods as well as network design. The research will involve both experimental validation on real-world and synthetic data and mathematical analysis of the proposed methods.