Implementation of Elastic Weight Consolidation / Fisher information matrix

Hey guys,

I was wondering if anyone has implemented Elastic Weight Consolidation (EWC) as outlined in this paper? This algorithm allows for sequential/continuous learning without the model encountering catastrophic forgetting.

The main part of implementing this is calculating the Fisher information matrix. If anyone has any code they can share on this, that’d be great. Otherwise I’m happy to attempt it and share my code here.

Found a tensorflow implementation here: https://github.com/stokesj/EWC which we can use for reference.

2 Likes

Did you find/do anything on PyTorch?

These two repos might have what you’re looking for:


1 Like

Hello. Are you that egg fried rice guy?

2 Likes

The original EWC requires you to compute the importance for each weight based on an additional pass over the training set. The importance is the squared gradient averaged over each minibatch. Anyway, you can take a look at the implementation available in the ContinualAI notebooks. It is an association for Continual Learning :slight_smile:

Disclaimer: I am part of ContinualAI.

1 Like