brendel-group / imi

Official repository for the paper "Scale Alone Does not Improve Mechanistic Interpretability in Vision Models" and the IMI dataset.

Project Page • Paper • Dataset • Citation

Scale Alone Does not Improve Mechanistic Interpretability in Vision Models

This repository contains code to reproduce the experiments described in the NeurIPS 2023 Spotlight paper Scale Alone Does not Improve Mechanistic Interpretability in Vision Models by Roland S. Zimmermann*, Thomas Klein*, Wieland Brendel'. If you have any questions, please reach out via email or create an issue here on GitHub and we'll try to answer it.

Structure

The mturk folder contains the implementation of the experiments' UI. Tools to host this on a web server can be found in the server directory. To generate the stimuli used in the experiments, look at the tools/data-generation folder. For performing the experiment using AWS Mechanical Turk, use the tools proved in tools/mturk. Finally, to evaluate the data and re-create the figures from the paper, use the notebooks provided in tools/data-analysis.

Citation

@inproceedings{zimmermann2023scale,
  title={Scale Alone Does not Improve Mechanistic Interpretability in Vision Models},
  author={Zimmermann, Roland S and Klein, Thomas and Brendel, Wieland},
  booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
  year={2023}
}