Deep learning systems do not only fit input-output mappings. They also learn internal features: structured representations that make a task easier to solve and, in some cases, easier to generalize.
This project studies how features emerge during training, how they become organized across layers, and how their structure relates to generalization, interpretability, and robustness. The goal is to understand feature learning not as a vague byproduct of optimization, but as a central mechanism through which neural networks acquire usable structure.
More broadly, the project asks what kinds of features deep networks tend to discover, when those features become stable, and how we can analyze or guide that process.