Improving and Understanding Variational Continual Learning

TL;DR

This paper enhances the Variational Continual Learning framework to better address catastrophic forgetting, model efficiency, and transfer learning in continual learning, demonstrating improved results and providing insights into its functioning.

Contribution

It introduces improved best practices for mean-field variational Bayesian neural networks within VCL and offers a detailed analysis of why VCL performs effectively.

Findings

Significant performance improvements on split and permuted MNIST benchmarks.

Enhanced understanding of VCL's mechanisms and its comparison to ideal solutions.

Establishment of new best practices for variational Bayesian neural networks.

Abstract

In the continual learning setting, tasks are encountered sequentially. The goal is to learn whilst i) avoiding catastrophic forgetting, ii) efficiently using model capacity, and iii) employing forward and backward transfer learning. In this paper, we explore how the Variational Continual Learning (VCL) framework achieves these desiderata on two benchmarks in continual learning: split MNIST and permuted MNIST. We first report significantly improved results on what was already a competitive approach. The improvements are achieved by establishing a new best practice approach to mean-field variational Bayesian neural networks. We then look at the solutions in detail. This allows us to obtain an understanding of why VCL performs as it does, and we compare the solution to what an `ideal' continual learning solution might be.