Elastic Feature Consolidation for Cold Start Exemplar-Free Incremental Learning

TL;DR

This paper introduces Elastic Feature Consolidation, a method for exemplar-free incremental learning that effectively manages feature drift and task bias, especially in cold start scenarios with limited initial data, outperforming existing methods.

Contribution

It proposes a novel regularization approach using a second-order approximation of feature drift and prototypes to improve exemplar-free incremental learning in cold start conditions.

Findings

Outperforms state-of-the-art on multiple datasets.

Effectively balances plasticity and stability in learning.

Reduces task-recency bias with prototype-based methods.

Abstract

Exemplar-Free Class Incremental Learning (EFCIL) aims to learn from a sequence of tasks without having access to previous task data. In this paper, we consider the challenging Cold Start scenario in which insufficient data is available in the first task to learn a high-quality backbone. This is especially challenging for EFCIL since it requires high plasticity, which results in feature drift which is difficult to compensate for in the exemplar-free setting. To address this problem, we propose a simple and effective approach that consolidates feature representations by regularizing drift in directions highly relevant to previous tasks and employs prototypes to reduce task-recency bias. Our method, called Elastic Feature Consolidation (EFC), exploits a tractable second-order approximation of feature drift based on an Empirical Feature Matrix (EFM). The EFM induces a pseudo-metric in feature space which we use to regularize feature drift in important directions and to update Gaussian prototypes used in a novel asymmetric cross entropy loss which effectively balances prototype rehearsal with data from new tasks. Experimental results on CIFAR-100, Tiny-ImageNet, ImageNet-Subset and ImageNet-1K demonstrate that Elastic Feature Consolidation is better able to learn new tasks by maintaining model plasticity and significantly outperform the state-of-the-art.