RethinkingNNsWithBL

Code for the paper "Rethinking Neural Networks with Benford's Law" in NeurIPS 2021 Machine Learning for Physical Sciences Workshop.

Usage

• To reproduce Experiment in Table 2 and Fig. 4, run python3 run_experiments.py. This will train over 900 LeNet-like models, and will run for a very long time. The results would be collected as json files at ./stats/. Tensorboard logs will be generated at lightning_logs. We have provided experimental data at stats_fig4 for our run.

• Plots in Fig. 3 were plotted using early stopping results.ipynb

• Plots in Fig. 5 were plotted using plot_simulation.ipynb.

File Descriptions

experiments.py

• contains PyTorch code for conducting all of the experiments in the paper (except for synthetic datasets). run_experiments.py
• is a python script to run multiple "experiments" in parallel.
• Run python3 run_experiments.py to reproduce results for most of the experiments presented in the paper.

weight_hist.py

• contains code for computing MLH score defined in the paper.
• Initilization method definitions.
• Plotting Layerwise MLH for various models.

models.py

• contains model definitions for various experiments.
• Info on where each model is used is described in the paper.