This code accompanies our paper:
Alleviating catastrophic forgetting using context-dependent gating and synaptic stabilization
Nicolas Y. Masse, Gregory D. Grant, David J. Freedman
https://www.pnas.org/content/115/44/E10467
The recurrent network model can be found in the repository:
Dependencies:
Python 3
TensorFlow 1+
In the paper, the model is tested on the following datasets:
MNIST Dataset
https://github.com/mrgloom/MNIST-dataset-in-different-formats
The dataset folder for MNIST is extracted and placed in './mnist/', so accessing the data from stimulus.py will be, for example, './mnist/data/original/train-images-idx3-ubyte'
CIFAR Dataset
https://www.cs.toronto.edu/~kriz/cifar.html
The dataset folders for CIFAR-10 and CIFAR-100 are extracted and placed separately in './cifar/', so accessing the data from stimulus.py will be, for example, './cifar/cifar-10-python/data_batch_1' or './cifar/cifar-100-python/test'
ImageNet Dataset
The dataset files for ImageNet are extracted and placed into './ImageNet', so accessing the data from stimulus.py will be, for example, './ImageNet/train_data_batch_1'