PyTorch implementation of linear and convolutional layers with fixed, random feedback weights. Training a network that uses these layers relies on the feedback alignment effect described in "Random synaptic feedback weights support error backpropagation for deep learning" by Lillicrap et al., 2016.
Includes a sample convolutional network and code to train the network on CIFAR-10.
utils.py and cifar10_example.py adapted from:
https://github.com/kuangliu/pytorch-cifar