CSA: Data-efficient Mapping of Unimodal Features to Multimodal Features

TL;DR

CSA is a novel method that efficiently maps unimodal features to multimodal space using limited data, outperforming existing methods without extensive training.

Contribution

Introduces CSA, a data-efficient approach for multimodal mapping that leverages unimodal encoders and matrix decomposition, reducing training data and computational requirements.

Findings

CSA outperforms CLIP with 50,000x less data

CSA surpasses state-of-the-art multimodal mapping methods

Effective for modalities beyond image and text

Abstract

Multimodal encoders like CLIP excel in tasks such as zero-shot image classification and cross-modal retrieval. However, they require excessive training data. We propose canonical similarity analysis (CSA), which uses two unimodal encoders to replicate multimodal encoders using limited data. CSA maps unimodal features into a multimodal space, using a new similarity score to retain only the multimodal information. CSA only involves the inference of unimodal encoders and a cubic-complexity matrix decomposition, eliminating the need for extensive GPU-based model training. Experiments show that CSA outperforms CLIP while requiring $50,000\times$ fewer multimodal data pairs to bridge the modalities given pre-trained unimodal encoders on ImageNet classification and misinformative news caption detection. CSA surpasses the state-of-the-art method to map unimodal features to multimodal features. We also demonstrate the ability of CSA with modalities beyond image and text, paving the way for future modality pairs with limited paired multimodal data but abundant unpaired unimodal data, such as lidar and text.