rodsveiga / pyscm

(Online) learning in two-layer neural networks

(Online) Learning in two-layer neural networks

This small package provides utilities to analyse online learning in neural networks with two fully connected networks, which are also known as Soft Committee Machines in the physics literature - or SCMs for short [1-3].

This package was used to conduct the simulations of our recent paper on the dynamics of stochastic gradient descent for two-layer neural networks in the teacher-student setup [4].

Contents

The package contains two parts:

1. A functional implementation of (online) learning in soft committee machines, i.e.* neural networks with a single fully connected hidden layer and scalar output [1-3].
2. An implementation of the ordinary differential equations (ODE) that describe online learning in soft committee machines, an appraoch that was pioneered in [1-3].

Install

To install locally, simply type

python setup.py install --user

Tools

This package comes with two ready-to-use tools:

1. examples/scm_online.py runs of learning in SCMs in the teacher-student setup; adding --help displays the available options.
2. examples/scm_ode.py runs an integrator of the ODEs that describe online learning in this setup [2,3]. Again, run scm_ode.py --help to see all available options.

Tests

There is a set of tests included with this package; to run them, simply type

nose2

in the top-level directory.

Requirements

• The SCM implementation for simulations is plain NumPy.
• The ODE integrator for SCMs with erf activation functions needs Cython

References

• [1] M. Biehl and H. Schwarze, J. Phys. A. Math. Gen. 28, 643 (1995).
• [2] D. Saad and S. A. Solla, Phys. Rev. Lett. 74, 4337 (1995)
• [3] D. Saad and S. A. Solla, Phys. Rev. E 52, 4225 (1995).
• [4] S. Goldt, M.S. Advani, A.M. Saxe, F. Krzakala, L. Zdeborová, NeurIPS 2019 (forthcoming) arXiv:1906.08632