MoE-Adapters++: Toward More Efficient Continual Learning of Vision-Language Models Via Dynamic Mixture-of-Experts Adapters.

Abstract

In this paper, we first propose MoE-Adapters, a parameter-efficient training framework to alleviate long-term forgetting issues in incremental learning with Vision-Language Models (VLM). Our MoE-Adapters leverages incrementally added routers to activate and integrate exclusive expert adapters from a pre-defined static expert set, enabling the pre-trained CLIP to efficiently adapt to new tasks. To preserve the zero-shot capability of VLM, a Distribution Discriminative Auto-Selector (DDAS) is introduced that automatically routes in-distribution and out-of-distribution inputs to the MoE-Adapters and the original CLIP, respectively. However, relying on a static expert set and a separate distribution selector can lead to parameter redundancy and increased training complexity. In response, we further extend an MoE-Adapters++ framework by introducing dynamic MoE-adapters, which allows experts to be adaptively involved during the continual learning process. Additionally, a Latent Embedding Auto-Selector (LEAS) is proposed that incorporates distribution selection within CLIP to create a more unified architecture. Extensive experiments across diverse settings demonstrate that the proposed method consistently surpasses previous state-of-the-art approaches while concurrently improving training efficiency.