Code for Local Signal Adaptivity: Provable Feature Learning in Neural Networks Beyond Kernels

by Stefani Karp, Ezra Winston, Yuanzhi Li, Aarti Singh

NeurIPS 2021

Setup

Python requirements are given in requirements.txt.

To install jax for gpu usage, see instructions at github.com/google/jax.

Contents

The file train.py runs the CIFAR-variant experiments, and the file synthetic_data_sim.py runs the synthetic-data experiments with the model used in the theory.

Examples of both types of runs are given in main_paper_experiments.sh.

Calling train.py allows for specification of the image dataset, noise type, image placement, etc.

Examples of various data variants are given in data_examples.ipynb.