Elastic Feature Consolidation for Cold Start Exemplar-free Incremental Learning

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. Deep neural networks achieve state-of-the-art performance on a broad range of visual recognition problems. However, the traditional supervised learning paradigm is limited in that it presumes all training data for all tasks is available in a single training session. The goal of Class-Incremental Learning (CIL) is to enable incremental integration of new classification tasks into already-trained models as they become available (Masana et al., 2022). Class-incremental learning entails balancing model plasticity (to allow learning of new tasks) against model stability (to avoid catastrophic forgetting of previous tasks) (McCloskey & Cohen, 1989). Exemplar-based approaches retain a small set of samples from previous tasks which are replayed to avoid forgetting, while exemplar-free methods retain no samples from previous tasks.