Overcoming Catastrophic Forgetting by Incremental Moment Matching

TL;DR

This paper introduces Incremental Moment Matching (IMM), a novel method to prevent catastrophic forgetting in neural networks by incrementally aligning the posterior distributions of sequential tasks, improving continual learning performance.

Contribution

IMM is a new approach that matches the moments of posterior distributions to retain knowledge across tasks, enhanced by transfer learning techniques for smoother optimization.

Findings

IMM achieves state-of-the-art results on multiple datasets.

IMM effectively balances old and new task information.

The method reduces catastrophic forgetting in neural networks.

Abstract

Catastrophic forgetting is a problem of neural networks that loses the information of the first task after training the second task. Here, we propose a method, i.e. incremental moment matching (IMM), to resolve this problem. IMM incrementally matches the moment of the posterior distribution of the neural network which is trained on the first and the second task, respectively. To make the search space of posterior parameter smooth, the IMM procedure is complemented by various transfer learning techniques including weight transfer, L2-norm of the old and the new parameter, and a variant of dropout with the old parameter. We analyze our approach on a variety of datasets including the MNIST, CIFAR-10, Caltech-UCSD-Birds, and Lifelog datasets. The experimental results show that IMM achieves state-of-the-art performance by balancing the information between an old and a new network.