On the Importance of Contrastive Loss in Multimodal Learning

TL;DR

This paper analyzes the role of contrastive loss in multimodal learning, revealing how positive and negative pairs influence the alignment and conditioning of learned representations, especially in non-isotropic data.

Contribution

It provides a theoretical analysis of contrastive learning dynamics, highlighting the importance of positive and negative pairs in balancing representations.

Findings

Positive pairs promote alignment but increase condition number.

Negative pairs reduce the condition number, balancing representations.

Contrastive pairs are crucial for efficient multimodal representation learning.

Abstract

Recently, contrastive learning approaches (e.g., CLIP (Radford et al., 2021)) have received huge success in multimodal learning, where the model tries to minimize the distance between the representations of different views (e.g., image and its caption) of the same data point while keeping the representations of different data points away from each other. However, from a theoretical perspective, it is unclear how contrastive learning can learn the representations from different views efficiently, especially when the data is not isotropic. In this work, we analyze the training dynamics of a simple multimodal contrastive learning model and show that contrastive pairs are important for the model to efficiently balance the learned representations. In particular, we show that the positive pairs will drive the model to align the representations at the cost of increasing the condition number, while the negative pairs will reduce the condition number, keeping the learned representations balanced.