Revisiting Catastrophic Forgetting in Continual Learning via Functional Similarity

This work is under review.

Requirements

This project uses the preset environment in Google Colab.

Usage

Please open the example file (.ipynb) in colab and copy the preparation files(func_simMNIST or func_simCIFAR10) into /home(in your Colab virtual host) as required by the notes in the example file.

Note: Please do not open the example file derictly in your own Google Drive. As the data exchange between Colab virtual host and your own Google Drive, it will make the data generation module run extremely slow .

Getting simple

We conduct experiments on two datasets, SplitMNIST and SplitCIFAR10, respectively.

SplitMNIST experiment:

SplitCIFAR10 experiment:

Acknowledgment

Our datasets are based on the continual learning community project: avalanche