Generalizable interactions

This repository contains the Python implementation of the paper "Defining and extracting generalizable interaction primitives from DNNs", ICLR 2024. The paper develops a new method to extract generalizable interactions shared by different DNNs (see papers for details and citations).

Install

Generalizable interactions shared by different DNNs can be installed in the Python 3 environment:

pip3 install git+https://github.com/sjtu-xai-lab/generalizable-interaction

How to use

Given two DNNs trained for the same task

Task 1: Sentiment classification with language models

Directly access the pre-trained BERT-base and BERT-large models, which were fine-tined on the SST-2 dataset in Google Drive. You can download pretrained_model.zip and unzip it into path like ./pretrained_model/task1/....

Computing generalizable interactions extracted from DNNs

To compute generalizable interactions extracted from two DNNs, you can use codes like the following:

python eval_interaction_text.py  --interaction='generalizable'  --model_path_1='BERT-base.pth.tar' --model_path_2='BERT-large.pth.tar'

then you can extract AND-OR interactions from the first DNN at paths ./output/task1/generalizable/Iand_1.npy and ./output/task1/generalizable/Ior_1.npy, and AND-OR interactions from the second DNN at paths ./output/task1/generalizable/Iand_2.npy and ./output/task1/generalizable/Ior_2.npy.

More details

In contrast, to compute traditional interactions extracted from a DNN (github), you can use codes like the following:

python eval_interaction_text.py  --interaction='traditional'  --model_path='BERT-base.pth.tar' --sparsify-lr=1e-6

then you can extract AND-OR interactions from the DNN at paths ./output/task1/traditional/{model}/Iand.npy and ./output/task1/traditional/{model}/Ior.npy

Sample notebooks

We provide Jupyter notebooks for all tasks to compare generalizable interactions with traditional interactions in the folder notebooks.

Citations

@InProceedings{chen2024defining,
  title = {Defining and extracting generalizable interaction primitives from DNNs},
  author = {Lu, Chen and Siyu, Lou and Benhao, Huang and Quanshi, Zhang},
  booktitle = {The Twelfth International Conference on Learning Representations},
  year = {2024}
}