This repository contains code accompanying the Supervised Similarity Matching Equilibrium Propagation paper. The code in this repo modifies the code for Equilibrium Propagation, released by Scellier & Bengio 2017 Equilibrium Propagation: Bridging the Gap between Energy-Based Models and Backpropagation and is implemented in the Theano Deep Learning Framework.
Learning Rule Variants Implemented:
- EP, 'betasigned': Equilibrium Propagation without lateral connections, beta (the nudge parameter) is randomly assigned a positive or negative sign post each batch. In the other learning rules, beta is always taken to be positive.
- EP, 'betapos': Equilibrium Propagation without lateral connections, beta (the nudge parameter) is positive.
- EP, Lateral: Equilibrium Propagation with lateral connections.
- SMEP: Similarity Matching update for lateral connections, Equilibrium Propagation update for forward connections.
This directory is organized as follows:
- Results: Contains saved network model results for all runs discussed in paper / supplementary information.
- Structured: Code for networks with structured connectivity. (SMEP)
- Main: Code for networks without structured connectivity. (EP, EP+Lateral, SMEP)
See README file in each folder about the instruction of running the code.