This is the github page for the results and code to reproduce the results for "Direct Feedback Alignment with Sparse Connections for Local Learning" (https://arxiv.org/abs/1903.02083). The main concept for this work is using Feedback Alignment (https://www.nature.com/articles/ncomms13276) and a extremely sparse matrix to reduce datamovement by orders of magnitude while enabling bio-plausible learning.
These instructions will get you a copy of the project up and running on your local machine for development and testing purposes.
What things you need to install the software and how to install them
tensorflow-gpu or tensorflow
ImageNet (http://www.image-net.org/)
git clone https://github.com/bcrafton/ssdfa
cd ssdfa
python mnist_fc.py --dfa 1 --sparse 1
To run imagenet, the training and test set can be acquired from: http://www.image-net.org/ The links must be changed inside of imagenet.py and imagenet_vgg.py
This code was run on 8 Nvidia Titan Xp 12GB GPUs. Only 1 was used per simulation (no multi-gpu simulations).
- tensorflow - The GPU framework used
- Brian Crafton
- Abhinav Parihar
- Evan Gebhardt
- Arijit Raychowdhury
Georgia Institute of Technology, ICSRL (http://icsrl.ece.gatech.edu/)
This project is licensed under the MIT License - see the LICENSE.md file for details