MultiWay-Adapater: Adapting large-scale multi-modal models for scalable image-text retrieval
This addresses the computational and memory demands of fine-tuning large multimodal models for scalable image-text retrieval, offering an incremental improvement over existing efficient adaptation methods.
The paper tackles the challenge of adapting large multimodal models to specialized tasks efficiently by introducing MultiWay-Adapter (MWA), which improves inter-modal alignment and reduces training time by up to 57% while adding only 2-3% more parameters.
As Multimodal Large Language Models (MLLMs) grow in size, adapting them to specialized tasks becomes increasingly challenging due to high computational and memory demands. Indeed, traditional fine-tuning methods are costly, due to the need for extensive, task-specific training. While efficient adaptation methods exist that aim to reduce these costs, in practice they suffer from shallow inter-modal alignment, which severely hurts model effectiveness. To tackle these computational challenges and improve inter-modal alignment, we introduce the MultiWay-Adapter (MWA), a novel framework featuring an 'Alignment Enhancer'. This enhancer deepens inter-modal alignment, enabling high transferability with minimal tuning effort. Our experiments show that unlike prior efficient tuning approaches, MWA maintains model effectiveness, while reducing training time by up-to 57%. MWA is also lightweight, increasing model size by only 2-3% (in terms of parameters) for state-of-the-art foundation models like BEiT-3 Large. These results demonstrate that MWA provides an efficient and effective adaptation method for MLLMs, significantly broadening their applicability.