Belkin Lab @ UC San Diego

Research

We are a research group working on understanding modern machine learning, and especially deep learning. We take a scientific approach, using tools across the spectrum from theory to experiment. For example, we want to characterize the behavior of deep learning systems, understand which design choices are necessary or sufficient for these behaviors, and identify universal structures underlying machine learning methods (deep or otherwise).

We are particularly interested in the areas of generalization, interpolation, optimization, overparameterization, and relations between these objects. In the past, our group has published influential results in Double Descent, Neural Tangent Kernels, overparameterization, and frameworks for generalization.

If Deep Learning is the solution, what is the problem?

Upcoming Talks:

• Nov 22, 2021 - Parthe Pandit (PhD Defense): Exact Analysis of Inverse Problems in High Dimensions with Applications to Machine Learning
• Nov 23, 2021 - Gil Kur: On the Minimax Sub-Optimality of Least Squares Estimator in the non-Donsker Regime
• Dec 9, 2021 - Nikhil Ghosh: The Three Stages of Learning Dynamics in High-Dimensional Kernel Methods

Events

Some Recorded Talks: