Monday, March 20, 2017
10:00 AM -
11:00 AM
Annenberg 105
Special Seminar in Computing and Mathematical Sciences
Better understanding of non-convex methods in machine learning.
Tengyu Ma,
PhD Candidate,
Computer Science,
Princeton University,
Speaker's Bio:
Tengyu Ma is a PhD candidate at the Computer Science Department of Princeton University advised by Sanjeev Arora. His research interests include topics in machine learning and algorithms, such as non-convex optimization, representation learning, deep learning, and convex relaxation for machine learning problems. His research contributes to the theoretical developments of these topics with practical implication. He is a recipient of NIPS'16 best student paper award, Princeton Honorific Fellowship, Siebel Scholarship, IBM PhD Fellowship, and the Simons Award for Graduate Students in Theoretical Computer Science.
Tengyu Ma is a PhD candidate at the Computer Science Department of Princeton University advised by Sanjeev Arora. His research interests include topics in machine learning and algorithms, such as non-convex optimization, representation learning, deep learning, and convex relaxation for machine learning problems. His research contributes to the theoretical developments of these topics with practical implication. He is a recipient of NIPS'16 best student paper award, Princeton Honorific Fellowship, Siebel Scholarship, IBM PhD Fellowship, and the Simons Award for Graduate Students in Theoretical Computer Science.
Recent breakthroughs in machine learning, especially deep learning, often involve learning complex and high-dimensional models on massive datasets with non-convex optimization. Though empirically successful, the formal study of such non-convex methods is much less developed. My research aims to develop new algorithmic approaches and analysis tools in these settings.
The talk will showcase a few results. First, we show that matrix completion — a famous problem in machine learning — can be solved by stochastic gradient descent on the straightforward non-convex objective function in polynomial time. (Formally, we show that all local minima of the objective are also global minima.) Then, we will analyze the landscape of the objective functions for linearized recurrent neural nets and residual nets, and demonstrate that over-parameterization and re-parameterization of the models can make the optimization easier.
Event Sponsors:
For more information, please contact Sheila Shull by phone at 626.395.4560 or by email at [email protected].