Demo: Feedback Alignment for a Slightly More Bio-Plausible Backprop
This repo contains short demos based on the following papers:
- Random feedback weights support learning in deep neural networks
- Direct Feedback Alignment Provides Learning in Deep Neural Networks
They explore the idea of unpairing the forwards-pass weights of a neural network and the backwards gradients used for training. Instead, a different, random matrix is used for the backwards pass.