~/projects/ff-algorithm-pytorch-implementation/

Forward-Forward Algorithm in PyTorch

May 3, 2023 #pytorch#deep-learning#research

An unofficial PyTorch implementation of Geoffrey Hinton's Forward-Forward algorithm, a backprop-free way to train neural networks.

Purpose

This repository is an unofficial PyTorch implementation exploring Geoffrey Hinton’s Forward-Forward (FF) approach. Its notebooks study replacing backpropagation’s forward and backward passes with two forward passes: one for positive examples and one for negative examples.

Method

Each layer tries to raise a “goodness” score for positive examples and lower it for negative ones. MLP and CNN experiments use this local objective; signals are evaluated per layer rather than propagating standard gradients backward through the network.

Status

The project aims to follow the paper but is not Hinton’s official repository. The README notes unfinished parts. The notebooks are for studying the method and experiments, not a finished product or benchmark result.

← cd ..