MoE-TTA: Enhancing Continual Test-Time Adaptation for Vision-Language Models through Mixture of Experts

Continual learning enables vision-language models to incrementally acquire new knowledge without relying on access to the entire historical dataset. This capability is crucial for adapting to evolving data distributions in real-world scenarios, where models must handle domain shifts and incorporate new information while retaining previously learned knowledge. However, maintaining performance in large-scale models remains challenging due to parameter shifts during learning and the significant computational costs associated with full-model updates. To address these challenges, this paper introduces a novel method for Continual Test-Time Domain Adaptation (CTTDA) on vision-language datasets, leveraging a MoE and adapter modules to optimize domain-specific adaptation while maintaining zero-shot classification capabilities. Utilising LoRA in MoE framework within transformer layers, the model efficiently updates a small set of parameters by dynamically selecting experts. This approach minimizes computational costs by activating only a portion of the model, avoiding the need for full-model updates during domain adaptation. During test-time adaptation, entropy loss is calculated without access to labels, improving the model’s fine-tuning and guiding towards confident predictions across domains. A contrastive warm-up phase further optimizes the adapter blocks by enhancing the differentiation of domain-specific and domain-invariant features, thereby establishing a strong foundation for effective test-time adaptation. The proposed MoE-TTA model achieves an average accuracy of 36.43% across diverse ImageNet datasets and 32.93% in fine-grained classification, demonstrating promising results, especially in datasets like EuroSAT (51.67%), while being lower than several competitors, with no doubt in its ability to capture domain shifts effectively.