← Projects
mechanistic interpretability
Reading how neural networks represent concepts internally, through feature discovery, probing, and activation patching.
- Synthesized 30+ papers on modern mechanistic interpretability
- Feature discovery, probing, activation patching and weight-based analysis
- NCSSM Research in Computational Science program
The problem
A trained network is a working system nobody can read. It is possible to know exactly what a model outputs and have no account of why, which makes it hard to trust, debug, or steer.
What I built
Feature discovery, probing, activation patching and weight-based analysis, focused on representation geometry, sparse feature directions, and toy models that expose mechanisms like superposition and grokking.
What happened
Synthesized 30+ papers on modern mechanistic interpretability through the NCSSM Research in Computational Science program.