Experiments with Direct Feedback Alignment (https://arxiv.org/abs/1609.01596) and comparison to Backpropagation.
Small DNN framework with:
- conv-, fully-connected-, dropout- and max-pool-layers
- implementation of backward pass with dfa and backpropagation
- customizable loss-function and optimizers (e.g. gd, momentum)
Reproducible experiments and comparisons
python setup.py