Continual Learning on Noisy Data Streams via Self-Purified Replay

TL;DR

This paper introduces a novel replay-based continual learning framework that effectively handles noisy labels and mitigates catastrophic forgetting by maintaining a purified replay buffer through self-supervised techniques and graph-based filtering.

Contribution

It proposes the first combined approach addressing noisy labels and forgetting in continual learning using self-replay and a self-centered filter for buffer purification.

Findings

Outperforms state-of-the-art methods on noisy datasets

Maintains high purity of replay buffer during streaming

Effectively mitigates catastrophic forgetting with noisy labels

Abstract

Continually learning in the real world must overcome many challenges, among which noisy labels are a common and inevitable issue. In this work, we present a repla-ybased continual learning framework that simultaneously addresses both catastrophic forgetting and noisy labels for the first time. Our solution is based on two observations; (i) forgetting can be mitigated even with noisy labels via self-supervised learning, and (ii) the purity of the replay buffer is crucial. Building on this regard, we propose two key components of our method: (i) a self-supervised replay technique named Self-Replay which can circumvent erroneous training signals arising from noisy labeled data, and (ii) the Self-Centered filter that maintains a purified replay buffer via centrality-based stochastic graph ensembles. The empirical results on MNIST, CIFAR-10, CIFAR-100, and WebVision with real-world noise demonstrate that our framework can maintain a highly pure replay buffer amidst noisy streamed data while greatly outperforming the combinations of the state-of-the-art continual learning and noisy label learning methods. The source code is available at http://vision.snu.ac.kr/projects/SPR