CULT: Continual Unsupervised Learning with aTpyicality-based Evironmental Selection

Inspired by Aachille et al. 2018, we implement a simple continual learning framework with variational autoencoders and generative replay, where new environments are detected using the atypicality score of the latents, where atypicality is given by:

{% raw %} $$\alpha = D_{KL}(\mathcal{N}(\bar{\mu}_{z}, \bar{\sigma}_z)||\mathcal{N}(0,1))$$ {% endraw %}

Installation

1. Clone this repository
2. Within the cloned repository, run pip install -e .

Experiments

The primary experiment(s) are run in 03a_experiments.cult_experiments.ipynb

Results are visualized in 04_visualize_data.ipynb

Nbdev

This repository uses nbdev. As a result, there are some superflous files for unused functionality.