PyTorch implementation of the paper "Where We Have Arrived in Proving the Emergence of Sparse Interaction Primitives in AI Models" in ICLR 2024 (paper).
- Python 3.8.0
- pytorch 2.0.1
- CUDA 11.7
- numpy 1.24.4
- transformers 4.31.0
All models were tested on a single A100 GPU.
You can also try the following command to install dependencies:
conda env create -f environment.yml
To better reproduce our results in the paper, we suggest directly downloading all the raw interactions from this Google Drive.
To obtain the box diagram in Figure 4, run the following command. Fill the --model argument with one of [opt, llama, aquila].
python ./demo/plot_inter_strength_boxplot.py --model=opt
To visualize the monotonicity assumption in Figure 5(a), run the following command.
python ./demo/plot_monotonicity_examples.py --model=opt
To reproduce the statistics in Table 1, run the following command.
python ./demo/check_monotonicity.py --model=opt
python ./demo/count_salient_concepts.py --model=opt
To reproduce the statistics in Table 2 and Figure 5(b), run the following command.
python ./demo/compute_p_and_bound.py --model=opt
See our project page here!
@inproceedings{
ren2024where,
title={Where We Have Arrived in Proving the Emergence of Sparse Interaction Primitives in {DNN}s},
author={Qihan Ren and Jiayang Gao and Wen Shen and Quanshi Zhang},
booktitle={The Twelfth International Conference on Learning Representations},
year={2024}
}