CULT: Continual Unsupervised Learning with Typicality-Based Environment Detection

TL;DR

CULT is a novel continual unsupervised learning algorithm that employs a typicality metric in VAE latent space to detect environment shifts, effectively reducing catastrophic forgetting through generative replay and environment classification.

Contribution

The paper introduces CULT, a new method combining typicality-based environment detection with generative replay for continual unsupervised learning with VAEs.

Findings

CULT significantly outperforms baseline methods in experiments.

The typicality metric effectively detects environmental shifts.

CULT reduces catastrophic forgetting in unsupervised learning.

Abstract

We introduce CULT (Continual Unsupervised Representation Learning with Typicality-Based Environment Detection), a new algorithm for continual unsupervised learning with variational auto-encoders. CULT uses a simple typicality metric in the latent space of a VAE to detect distributional shifts in the environment, which is used in conjunction with generative replay and an auxiliary environmental classifier to limit catastrophic forgetting in unsupervised representation learning. In our experiments, CULT significantly outperforms baseline continual unsupervised learning approaches. Code for this paper can be found here: https://github.com/oliveradk/cult