The Benefits of Balance: From Information Projections to Variance Reduction

TL;DR

This paper reveals that data balancing across multiple modalities in foundation models reduces variance, providing a new theoretical understanding and practical insights for improving contrastive and self-supervised learning methods.

Contribution

It introduces a non-asymptotic statistical bound linking data balancing to variance reduction and offers a novel perspective to enhance multimodal learning techniques.

Findings

Data balancing reduces variance in foundation models.

A non-asymptotic bound quantifies variance reduction effects.

Insights improve contrastive and self-supervised learning methods.

Abstract

Data balancing across multiple modalities and sources appears in various forms in foundation models in machine learning and AI, e.g. in CLIP and DINO. We show that data balancing across modalities and sources actually offers an unsuspected benefit: variance reduction. We present a non-asymptotic statistical bound that quantifies this variance reduction effect and relates it to the eigenvalue decay of Markov operators. Furthermore, we describe how various forms of data balancing in contrastive multimodal learning and self-supervised clustering can be better understood, and even improved upon, owing to our variance reduction viewpoint.