Purpose: The proliferation of massive, unlabeled multimodal datasets presents a significant opportunity and a fundamental challenge for modern artificial intelligence. Supervised learning methods, which depend on costly and often scarce human-annotated labels, are ill-suited for this reality. This article provides a comprehensive review of contrastive learning, a dominant self-supervised paradigm, as a powerful solution for learning rich feature representations from unlabeled multimodal data.

Approach: We survey the landscape of contrastive learning, beginning with the foundational principles and seminal unimodal architectures that established the field, including Momentum Contrast (MoCo) and SimCLR. We then conduct a detailed examination of the extension of these principles into the more complex multimodal domain. Key architectures are systematically categorized and analyzed, including pioneering vision-language models like CLIP and FLAVA, audio-visual systems, and applications to other data types like time series. The review synthesizes architectural innovations, theoretical underpinnings, and strategies for handling both aligned and unaligned data sources.

Findings: Multimodal contrastive learning has proven exceptionally effective at creating semantically rich, unified embedding spaces where different data modalities can be compared and aligned. By training models to distinguish between corresponding (positive) and non-corresponding (negative) pairs of data from different modalities, these systems learn transferable representations that excel at zero-shot, few-shot, and transfer learning tasks. These methods effectively bypass the need for explicit labels, instead leveraging the natural co-occurrence of information across modalities as a supervisory signal.

Conclusion: While transformative, significant challenges remain in computational scalability, robust negative sampling, and standardized evaluation. Future research will likely focus on developing more computationally efficient architectures, improving robustness to noisy data, and extending these powerful methods to a wider array of scientific and industrial domains.