Memory-Efficient Continual Learning with CLIP Models

TL;DR

This paper introduces a memory-efficient, distributionally robust method for continual learning with CLIP models, enabling quick adaptation and minimal forgetting even with limited memory buffers.

Contribution

It proposes a novel loss reweighting technique that improves CLIP's continual learning performance under small memory constraints.

Findings

Achieves effective class incremental learning on CIFAR-100 and ImageNet1K.

Maintains high performance with minimal memory buffer sizes.

Reduces catastrophic forgetting in domain incremental tasks.

Abstract

Contrastive Language-Image Pretraining (CLIP) models excel at understanding image-text relationships but struggle with adapting to new data without forgetting prior knowledge. To address this, models are typically fine-tuned using both new task data and a memory buffer of past tasks. However, CLIP's contrastive loss suffers when the memory buffer is small, leading to performance degradation on previous tasks. We propose a memory-efficient, distributionally robust method that dynamically reweights losses per class during training. Our approach, tested on class incremental settings (CIFAR-100, ImageNet1K) and a domain incremental setting (DomainNet) adapts CLIP models quickly while minimizing catastrophic forgetting, even with minimal memory usage.