conv_emerge

This respository contains code to accompany the paper "Data-driven emergence of convolutional structure in neural networks" [arXiv][https://arxiv.org/abs/2202.00565] by A. Ingrosso and S. Goldt.

Overview

Usage

This package contains a utilities:

• inputs.py defines various input models, such as Gaussian process, non-linear Gaussian process, etc.
• tasks.py defines various tasks, for example mixture classification tasks like the ones analysed in the paper
• network.py defines the neural network models that we train
• utils.py contains some helper functions

Simulations of online learning, like the ones shown in the figure above, can be obtained by simulating online learning with conv_emerge_online.py.

To get an overview of the different parameters, run

python conv_emerge_online.py --help

Requirements

To run the code, you will need up-to-date versions of

• pyTorch
• numpy