Support-set bottlenecks for video-text representation learning

TL;DR

This paper introduces a novel approach to video-text representation learning that relaxes the strict dissimilarity enforcement of contrastive learning by using a generative model to promote semantic sharing among related samples, improving retrieval performance.

Contribution

It proposes a generative-based method that encourages semantic similarity among related samples, addressing limitations of traditional contrastive learning in video-text tasks.

Findings

Outperforms existing methods on MSR-VTT, VATEX, ActivityNet, and MSVD datasets.

Achieves significant improvements in video-to-text and text-to-video retrieval tasks.

Promotes more generalizable and semantically rich representations.

Abstract

The dominant paradigm for learning video-text representations -- noise contrastive learning -- increases the similarity of the representations of pairs of samples that are known to be related, such as text and video from the same sample, and pushes away the representations of all other pairs. We posit that this last behaviour is too strict, enforcing dissimilar representations even for samples that are semantically-related -- for example, visually similar videos or ones that share the same depicted action. In this paper, we propose a novel method that alleviates this by leveraging a generative model to naturally push these related samples together: each sample's caption must be reconstructed as a weighted combination of other support samples' visual representations. This simple idea ensures that representations are not overly-specialized to individual samples, are reusable across the dataset, and results in representations that explicitly encode semantics shared between samples, unlike noise contrastive learning. Our proposed method outperforms others by a large margin on MSR-VTT, VATEX and ActivityNet, and MSVD for video-to-text and text-to-video retrieval.