Hi there! This repo contains the experimentation systems and analysis for the paper "Generalizing Backpropagation for Gradient-Based Interpretability" by Du et al., published at ACL 2023.

Hi there! Please note that while this repo contains all preprocessing, experimentation, and analysis code to fully reproduce the results in the paper (as of July 2023), refactoring it to be easily accessible is still a work-in-progress. Thanks for your patience, and please open an issue on Github if you have any questions or concerns. I will soon be opening Github issues to track the specific tasks required to make this code more easily accessible and runnable.

Current priorities include:

• Refactoring the experiment code and renaming scripts for consistency and clarity
• Improved documentation describing the entry points for each experiment and any manual steps
• Documenting which experiment code corresponds to which results in the paper
• Dockerizing appropriate experiments

1. Make sure you have an editable install of brunoflow (https://github.com/kdu4108/brunoflow). This library implements backpropagation such that it is easily extensible for other semirings.
2. Make sure you have an editable install of the brunoflow-modified transformers (https://github.com/kdu4108/transformers). This is necessary because transformers does not support brunoflow models and therefore I needed to reimplement brunoflow-based BERT models in this fork of the transformers library.
3. conda install wandb. We use Weights & Biases (https://wandb.ai/site) for experiment tracking. You will probably need to make an account in order to run these scripts (or else you'll need to comment out wandb-related code in the scripts/notebooks.)
4. conda install papermill. Since some of the experimentation code is written in Jupyter Notebooks, we use papermill (https://www.google.com/search?q=papermill&oq=papermill&aqs=chrome.0.69i59j0i512l2j0i10i512j0i512l2j0i10i30j0i30l3.831j0j7&sourceid=chrome&ie=UTF-8) as a programmatic executor for the notebooks. This enables us to run parameterized Jupyter Notebooks from top-to-bottom (like scripts) and saves them for provenance.
5. pre-commit install for some nice auto-linting/formatting upon committing.

For this paper, we constructed various synthetic datasets in addition to standard NLP datasets. For the implementation of synthetic datasets, see the preprocessing/datasets.py. For the subject-verb agreement (SVA) dataset, see https://github.com/yoavg/bert-syntax and our additional preprocessing/subsampling code in preprocessing/lgd_sva/subsample_lgd_data.ipynb. Note that you will need to download the dataset from the repo linked above into the path data/lgd_sva/lgd_dataset.tsv for the subsampling script to work out of the box.

Validating top gradient path on a synthetic task (section 5.1)

• Task: FirstTokenRepeatedOnce

• Experiment code entry point: analyze_kvq_synthetic.py

• Analysis code: analysis/synthetic_bert_max_grad.ipynb

Analyzing top gradient path of BERT on SVA (section 5.2)

• Task: subject-verb Agreement (SVA)

• Experiment code entry point: analyze_kvq.py

• Analysis code: analysis/sva_max_grad.ipynb

Comparing gradient graph entropy vs task difficulty (section 5.3)

• Task: various synthetic tasks (see Appendix A of the paper).

• Experiment code entry point: compute_mdl.py

• Analysis code: analysis/analyze_entropy_vs_mdl_synthetic.ipynb

Comparing gradient graph entropy vs example difficulty (appendix B.2)

• Task: subject-verb Agreement (SVA)

• Experiment code entry point: eval_bert.py

• Analysis code: analysis/sva_entropy.ipynb

Sanity check for gradient graph entropy vs model complexity (appendix B.1)

• Task: FFOOM

• Experiment code entry point: python pm_train_mlp.py -p train_ffoom_mlp -hs 64 -s 6 -d FFOOM -k '{"num_points": 10000}'

• Analysis code: analysis/analyze_ffoom_entropy.ipynb

• Since brunoflow is not optimized for execution speed, it faces practical computational and memory constraints. We aim to recruit assistance to reimplement semiring-backpropagation in a more popular, GPU-accelerated library such as Pytorch or JAX to promote more practical use-cases of this method.
• Again due to computational constraints from brunoflow's implementation, we run some experiments distributed over our slurm cluster. For those who are interested, we share those scripts in the slurm/ directory.