xqding / Context-Dependent-Gating-RNN

Algorithm to alleviate catastrophic forgetting in neural networks by gating hidden units, in RNNs

Context-Dependent Gating for RNNs

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://arxiv.org/abs/1802.01569

The feed-forward network model can be found in the repository:

https://github.com/nmasse/Context-Dependent-Gating/

Dependencies:

Python 3
TensorFlow 1+

In our paper, the model is tested on a set of cognitive tasks proposed by Yang et al.

https://www.biorxiv.org/content/early/2017/09/01/183632

Our tasks are set up in the stimulus.py file, and are cycled through automatically with the current configuration. Note that our model uses one-hot outputs as opposed to distributions, as originally described by Yang et al.