Meta-Consolidation for Continual Learning

TL;DR

This paper introduces MERLIN, a novel meta-consolidation approach for continual learning that incrementally learns a meta-distribution of neural network weights, enabling models to adapt to new tasks without forgetting previous ones.

Contribution

The paper proposes MERLIN, a new meta-consolidation method that learns and consolidates a meta-distribution of weights for continual learning in an online setting.

Findings

Consistent improvement over five baselines on multiple benchmarks.

Effective in online continual learning where data points are seen only once.

Outperforms recent state-of-the-art methods.

Abstract

The ability to continuously learn and adapt itself to new tasks, without losing grasp of already acquired knowledge is a hallmark of biological learning systems, which current deep learning systems fall short of. In this work, we present a novel methodology for continual learning called MERLIN: Meta-Consolidation for Continual Learning. We assume that weights of a neural network $\boldsymbol \psi$, for solving task $\boldsymbol t$, come from a meta-distribution $p(\boldsymbol{\psi|t})$. This meta-distribution is learned and consolidated incrementally. We operate in the challenging online continual learning setting, where a data point is seen by the model only once. Our experiments with continual learning benchmarks of MNIST, CIFAR-10, CIFAR-100 and Mini-ImageNet datasets show consistent improvement over five baselines, including a recent state-of-the-art, corroborating the promise of MERLIN.