Code for the paper Quantitative Propagation of Chaos for SGD in Wide Neural Networks by Valentin De Bortoli, Alain Durmus, Xavier Fontaine and Umut Şimşekli, in Advances in Neural Information Processing Systems, 2020.

See the paper for more details.

Run the run.py file by specifying which parameters to use (width, batch size, etc.) . It will produce a jobs file. Then run all the jobs of this file, for example with the parallel tool: parallel < jobs. This will create log files into a folder name width_exp_T100.

Once the networks are trained, the weights are saved into the width_exp_T100 folder. Run the ipython notebook visualize_results.ipynb to display the histograms.

The code is based on the following repository.