Task Agnostic Continual Learning Using Online Variational Bayes

TL;DR

This paper introduces a novel task-agnostic continual learning method using online variational Bayes to prevent catastrophic forgetting without prior knowledge of task boundaries.

Contribution

It proposes the first approach for task-agnostic continual learning that effectively mitigates forgetting without requiring known task boundaries during training.

Findings

The method successfully prevents catastrophic forgetting in unknown task boundary scenarios.

The approach outperforms existing methods in continual learning benchmarks.

Code implementation is publicly available for reproducibility.

Abstract

Catastrophic forgetting is the notorious vulnerability of neural networks to the change of the data distribution while learning. This phenomenon has long been considered a major obstacle for allowing the use of learning agents in realistic continual learning settings. A large body of continual learning research assumes that task boundaries are known during training. However, research for scenarios in which task boundaries are unknown during training has been lacking. In this paper we present, for the first time, a method for preventing catastrophic forgetting (BGD) for scenarios with task boundaries that are unknown during training --- task-agnostic continual learning. Code of our algorithm is available at https://github.com/igolan/bgd.