The Clock and the Pizza: Two Stories in Mechanistic Explanation of Neural Networks

This is the code repository for our paper The Clock and the Pizza: Two Stories in Mechanistic Explanation of Neural Networks, accepted by NeurIPS 2023 (oral).

How to Replicate

Some trained models are available in code/save/, including ModelA and ModelB in the paper. Use code/eval-main.ipynb (for transformers) and code/eval-linear.ipynb (for linear models) to calculate the metrics of the trained models. To train models from scratch, use code/train-main.py for transformers and code/train-linear.py for linear models.

The collection of all the trained models are not released for size concerns, but two tables of stats collected across all 1-layer runs are available as result/1l_fixedwidth.csv and result/1l_varywidth.csv. The former is for the fixed-width experiments and the latter is for the varying-width experiments. Note that the two tables also contain runs with slightly different settings: runs with 1 in diff_vocab and eqn_sign columns correspond to the runs in the appendix. You may also want to filter out non-circular runs (circularity column).

Citation

@article{zhong2023clock,
  title={The clock and the pizza: Two stories in mechanistic explanation of neural networks},
  author={Zhong, Ziqian and Liu, Ziming and Tegmark, Max and Andreas, Jacob},
  journal={arXiv preprint arXiv:2306.17844},
  year={2023}
}