Adaptive Shortcut Debiasing for Online Continual Learning

TL;DR

This paper introduces DropTop, an adaptive framework for online continual learning that effectively suppresses shortcut bias without prior knowledge, significantly improving accuracy and reducing forgetting across multiple benchmarks.

Contribution

DropTop is a novel, adaptive debiasing framework that automatically adjusts the suppression of shortcut bias in online continual learning environments.

Findings

Increases average accuracy by up to 10.4%.

Decreases forgetting by up to 63.2%.

Effective across five benchmark datasets.

Abstract

We propose a novel framework DropTop that suppresses the shortcut bias in online continual learning (OCL) while being adaptive to the varying degree of the shortcut bias incurred by continuously changing environment. By the observed high-attention property of the shortcut bias, highly-activated features are considered candidates for debiasing. More importantly, resolving the limitation of the online environment where prior knowledge and auxiliary data are not ready, two novel techniques -- feature map fusion and adaptive intensity shifting -- enable us to automatically determine the appropriate level and proportion of the candidate shortcut features to be dropped. Extensive experiments on five benchmark datasets demonstrate that, when combined with various OCL algorithms, DropTop increases the average accuracy by up to 10.4% and decreases the forgetting by up to 63.2%.