Learning Sparse Disentangled Representations for Multimodal Exclusion Retrieval

Multimodal representations are essential for cross-modal retrieval, but they often lack interpretability, making it difficult to understand the reasoning behind retrieved results. Sparse disentangled representations offer a promising solution; however, existing methods rely heavily on text tokens, resulting in high-dimensional embeddings. In this work, we propose a novel approach that generates compact, fixed-size embeddings that maintain disentanglement while providing greater control over retrieval tasks. We evaluate our method on challenging exclusion queries using the MSCOCO and Conceptual Captions benchmarks, demonstrating notable improvements over dense models like CLIP, BLIP, and VISTA (with gains of up to 11% in AP@10), as well as over sparse disentangled models like VDR (achieving up to 21% gains in AP@10). Furthermore, we present qualitative results that emphasize the enhanced interpretability of our disentangled representations.