Implementation of BrainProp, a biologically plausible learning rule that can train deep neural networks on image-classification tasks (MNIST, CIFAR10, CIFAR100, Tiny ImageNet).

This repository is the official implementation of "Attention-Gated Brain Propagation: How the brain can implement reward-based error backpropagation", NeurIPS 2020 proceedings.

In the paper we show that by training only one output unit at a time we obtain a biologically plausible learning rule able to train deep neural networks on state-of-the-art machine learning classification tasks. The architectures used range from 3 to 8 hidden layers.

The current version of the code requires a recent (as of June 2020) version of tensorflow-gpu, CUDA and cuDNN and it was specifically tested on the following versions of the packages:

• Python 3.6.6
• pip 20.1.1
• CUDA 10.1.243
• cuDNN 7.6.5.32

To install the required libraries and modules (after having created a virtual environment with the versions of Python and pip indicated above):

pip install -r requirements.txt

• MNIST, CIFAR10 and CIFAR100 are automatically available through keras.
• Tiny ImageNet can be downloaded from the official page of the challenge or extracted by running:

python tinyimagenet.py

        in the directory where the file "tiny-imagenet-200.zip" is located.

To train the model(s) in the paper, run this command:

python main.py <dataset> <architecture> <algorithm>

the training will stop when the validation accuracy has not increased for 45 epochs, otherwise until 500 epochs are reached.

The possible <dataset> - <architecture> combinations are:

• MNIST - {dense, loccon, conv}

• CIFAR10 - {loccon, conv, deep}

• CIFAR100 - {loccon, conv, deep}

• TinyImageNet - deep

  For the details of the architectures, please refer to the paper.

For <algorithm>, set BrainProp for BrainProp or EBP for error-backpropagation.

Add the flag -s to save a plot of the accuracy, the trained weights (at the best validation accuracy) and the history file of the training.

To load and evaluate a saved model:

python main.py <dataset> <architecture> <algorithm> -l <weightfile.h5>

Three pre-trained models (on the deep network with BrainProp) on CIFAR10 (CIFAR10_BrainProp_weights.h5), CIFAR100 (CIFAR100_BrainProp_weights.h5) and Tiny ImageNet (TIN_BrainProp_weights.h5) are included.

All the hyperparameters (as specified in the paper) are automatically set depending on which architecture is chosen.

All the experiments ran on one node with a NVIDIA GeForce 1080Ti card.

Our algorithm achieved the following performances (averaged over 10 different seeds, the mean and standard deviation are indicated):

For the dense and conv simulations the speed was 3s/epoch, while for loccon the speed ranged between 45- and 60s/epoch.

For the complete tables and figures, please refer to the paper.