This repository hosts code to reproduce all results in the paper "An analytic theory of shallow networks dynamics for hinge loss classification" by F. Pellegrini and G. Biroli, which was accepted as a poster to NeurIPS 2020, and can be found here.
The code consists of a single Jupyter notebook, based on Python 3 and requiring libraries numpy, scipy, tensorflow (1.xx), and matplotlib. All examples can be run in a few minutes on a moderately powerful machine. For more information please see the paper and comments in the notebook.
Additionally, we include a short animation showing the training and validation error and parameters evolution (again, see the paper for more details).
