Text-centric Alignment for Multi-Modality Learning
It addresses the challenge of dynamic modality availability in real-world applications, offering a flexible solution for multimodal systems.
This paper tackles the problem of modality mismatch in multimodal learning, where training and inference modalities differ, by proposing the Text-centric Alignment for Multi-Modality Learning (TAMML) approach, which uses Large Language Models and foundation models to improve generalizability, demonstrating significant performance gains in handling unseen modality combinations.
This research paper addresses the challenge of modality mismatch in multimodal learning, where the modalities available during inference differ from those available at training. We propose the Text-centric Alignment for Multi-Modality Learning (TAMML) approach, an innovative method that utilizes Large Language Models (LLMs) with in-context learning and foundation models to enhance the generalizability of multimodal systems under these conditions. By leveraging the unique properties of text as a unified semantic space, TAMML demonstrates significant improvements in handling unseen, diverse, and unpredictable modality combinations. TAMML not only adapts to varying modalities but also maintains robust performance, showcasing the potential of foundation models in overcoming the limitations of traditional fixed-modality frameworks in embedding representations. This study contributes to the field by offering a flexible, effective solution for real-world applications where modality availability is dynamic and uncertain.