Data-Efficient Multimodal Fusion on a Single GPU

TL;DR

FuseMix enables efficient multimodal fusion by leveraging pre-trained unimodal encoders, achieving competitive results with significantly less compute and data, and can adapt generative models across modalities.

Contribution

Proposes FuseMix, a novel multimodal augmentation method that operates on latent spaces of pre-trained unimodal encoders, reducing training costs while maintaining high performance.

Findings

Outperforms state-of-the-art methods in image-text and audio-text retrieval.

Requires approximately 600 times fewer GPU days and 80 times less data than CLIP.

Can convert pre-trained text-to-image models into audio-to-image models.

Abstract

The goal of multimodal alignment is to learn a single latent space that is shared between multimodal inputs. The most powerful models in this space have been trained using massive datasets of paired inputs and large-scale computational resources, making them prohibitively expensive to train in many practical scenarios. We surmise that existing unimodal encoders pre-trained on large amounts of unimodal data should provide an effective bootstrap to create multimodal models from unimodal ones at much lower costs. We therefore propose FuseMix, a multimodal augmentation scheme that operates on the latent spaces of arbitrary pre-trained unimodal encoders. Using FuseMix for multimodal alignment, we achieve competitive performance -- and in certain cases outperform state-of-the art methods -- in both image-text and audio-text retrieval, with orders of magnitude less compute and data: for example, we outperform CLIP on the Flickr30K text-to-image retrieval task with $\sim \! 600\times$ fewer GPU days and $\sim \! 80\times$ fewer image-text pairs. Additionally, we show how our method can be applied to convert pre-trained text-to-image generative models into audio-to-image ones. Code is available at: https://github.com/layer6ai-labs/fusemix.