Research Thrusts

Our work centers on three complementary directions: representation, generalization, and efficiency.

Representation

We study how deep models organize high-dimensional data into structured, low-dimensional features that support their functionality.

Reading List Neural Collapse and Representation Learning

Representative Papers

Generalization

We aim to explain when deep models learn underlying distributions instead of memorizing samples, and how this transition depends on model architecture, optimization, and data.

Reading List Memorization and Generalization of Diffusion Models

Representative Papers

Efficiency

We develop low-complexity principles to help deep models train and run efficiently.

Representative Paper

Compressible Dynamics in Deep Overparameterized Low-Rank Learning & Adaptation

Can Yaras, Peng Wang, Laura Balzano, Qing Qu.ICML'24

Oral Presentation (top 1.5%); Best Poster Award at MMLS'24

Representation

Understanding Deep Representation Learning via Layerwise Feature Compression and Discrimination

A Geometric Analysis of Neural Collapse with Unconstrained Features

Generalization

A Closer Look at Model Collapse: From a Generalization-to-Memorization Perspective

Diffusion Models Learn Low-Dimensional Distributions via Subspace Clustering

The Emergence of Reproducibility and Generalizability in Diffusion Models

Efficiency

Compressible Dynamics in Deep Overparameterized Low-Rank Learning & Adaptation

Inverse Problems