Harvard Machine Learning Theory

We are a research group focused on building towards a theory of modern machine learning. We are interested in both experimental and theoretical approaches that advance our understanding.

Key topics include: generalization, over-parameterization, robustness, dynamics of SGD, and relations to kernel methods.

We also run a research-level seminar series on recent advances in the field. Join the seminar mailing list for talk announcements. In Spring 2021, Professor Boaz Barak will be teaching Harvard CS 229br, a graduate level course on recent advances and open questions in the theory of machine learning and specifically deep learning.

People

Researchers

Avatar

Boaz Barak

Faculty

Avatar

Preetum Nakkiran

PhD Student

Avatar

Gal Kaplun

PhD Student

Avatar

Yamini Bansal

PhD Student

Avatar

Nikhil Vyas

PhD Student

Avatar

Sharon Qian

PhD Student

Avatar

Ben Edelman

PhD Student

Avatar

Yonadav Shavit

PhD Student

Avatar

Fred Zhang

PhD Student

Avatar

Tristan Yang

Undergraduate

Affiliated

Recent Publications

By our group and its members.

For Self-supervised Learning, Rationality Implies Generalization, Provably

The Deep Bootstrap: Good Online Learners are Good Offline Generalizers

Distributional Generalization: A New Kind of Generalization

Learning From Strategic Agents: Accuracy, Improvement, and Causality

Deep Double Descent: Where Bigger Models and More Data Hurt

SGD on Neural Networks Learns Functions of Increasing Complexity

More Data Can Hurt for Linear Regression: Sample-wise Double Descent

Computational Limitations in Robust Classification and Win-Win Results

Minnorm training: an algorithm for training over-parameterized deep neural networks

Adversarial Robustness May Be at Odds With Simplicity

On the Information Bottleneck Theory of Deep Learning

Recent & Upcoming Talks

Standard machine learning produces models that are accurate on average but degrade dramatically on when the test distribution of …

As datasets continue to grow in size, in many settings the focus of data collection has shifted away from testing pre-specified …

Deep learning seeks to discover universal models that work across all modalities and tasks. While self-attention has enhanced the …

Many supervised learning methods are naturally cast as optimization problems. For prediction models which are linear in their …

Gradient descent algorithms and their noisy variants, such as the Langevin dynamics or multi-pass SGD, are at the center of attention …

Seminar Calendar

Join the mailing list for talk announcements.