REMIND Your Neural Network to Prevent Catastrophic Forgetting

TL;DR

REMIND is a brain-inspired, online learning method that uses compressed memory replay to prevent catastrophic forgetting in neural networks, outperforming existing approaches on image classification and extending to VQA.

Contribution

It introduces REMIND, a novel online, compressed replay approach inspired by neuroscience, for incremental learning in neural networks.

Findings

Outperforms existing methods on ImageNet incremental class learning.

Demonstrates robustness to data order schemes that induce forgetting.

Extends online learning to Visual Question Answering (VQA).

Abstract

People learn throughout life. However, incrementally updating conventional neural networks leads to catastrophic forgetting. A common remedy is replay, which is inspired by how the brain consolidates memory. Replay involves fine-tuning a network on a mixture of new and old instances. While there is neuroscientific evidence that the brain replays compressed memories, existing methods for convolutional networks replay raw images. Here, we propose REMIND, a brain-inspired approach that enables efficient replay with compressed representations. REMIND is trained in an online manner, meaning it learns one example at a time, which is closer to how humans learn. Under the same constraints, REMIND outperforms other methods for incremental class learning on the ImageNet ILSVRC-2012 dataset. We probe REMIND's robustness to data ordering schemes known to induce catastrophic forgetting. We demonstrate REMIND's generality by pioneering online learning for Visual Question Answering (VQA).