Hypno: Evading the Catastrophic Forgetting Problem in Block I.I.D. Data with Sleep

In this project we propose an algorithm that attempts to mitigate the issue of catastrophic forgetting when learning on data that cyclicly changes distributions.

Pre-requisites

Install the requirements with

pip install -r requirements.txt

and by installing jax from https://github.com/google/jax#installation

Just running the algorithm

To simply demonstrate the algorithm run the demo code with

python3 main.py

Optional commands and documentation can be found with

python3 main.py -h

Seeing a Full Explanation

To see a full explanation of the code and some experiments open the hypno.ipynb file directly here in github (https://github.com/codymlewis/hypno/blob/main/hypno.ipynb) or locally with jupyter, note that in the latter case you will need to install jupyter from pip first.