Delving into Inter-Image Invariance for Unsupervised Visual Representations

Abstract: Contrastive learning has recently shown immense potential in unsupervised visual representation learning. Existing studies in this track mainly focus on intra-image invariance learning. The learning typically uses rich intra-image transformations to construct positive pairs and then maximizes agreement using a contrastive loss. The merits of inter-image invariance, conversely, remain much less explored. One major obstacle to exploit inter-image invariance is that it is unclear how to reliably construct inter-image positive pairs, and further derive effective supervision from them since no pair annotations are available. In this work, we present a comprehensive empirical study to better understand the role of inter-image invariance learning from three main constituting components: pseudo-label maintenance, sampling strategy, and decision boundary design. To facilitate the study, we introduce a unified and generic framework that supports the integration of unsupervised intra- and inter-image invariance learning. Through carefully-designed comparisons and analysis, multiple valuable observations are revealed: 1) online labels converge faster and perform better than offline labels; 2) semi-hard negative samples are more reliable and unbiased than hard negative samples; 3) a less stringent decision boundary is more favorable for inter-image invariance learning. With all the obtained recipes, our final model, namely InterCLR, shows consistent improvements over state-of-the-art intra-image invariance learning methods on multiple standard benchmarks. We hope this work will provide useful experience for devising effective unsupervised inter-image invariance learning. Code: https://github.com/open-mmlab/mmselfsup.

• The paper presents InterCLR, an unsupervised framework that integrates intra- and inter-image invariance learning to enhance visual representations.
• It introduces an online mini-batch k-means clustering method for dynamic pseudo-label maintenance and a semi-hard negative sampling strategy to balance challenging negatives.
• Empirical results demonstrate substantial gains on tasks like image classification and object detection, underscoring the framework's efficiency and scalability.

Delving into Inter-Image Invariance for Unsupervised Visual Representations

The paper "Delving into Inter-Image Invariance for Unsupervised Visual Representations" explores the relatively under-investigated domain of inter-image invariance learning as applied to unsupervised contrastive learning. Aimed at addressing the challenges of exploiting inter-image representations without explicit pair annotations, the authors propose a unified framework dubbed InterCLR, which integrates both intra-image and inter-image invariance learning strategies.

Key Contributions

The study systematically investigates inter-image invariance through three core components: pseudo-label maintenance, sampling strategy, and decision boundary design. Each aspect is meticulously studied to enhance inter-image learning's effectiveness in contrastive frameworks.

1. Pseudo-Label Maintenance: Traditional methods predominantly employ offline cluster assignments that are computationally demanding, resulting in stale labels. The authors propose an online mini-batch $k$-means clustering method, allowing labels to be updated iteratively and synchronously with network training, which helps maintain label reliability and convergence speed.
2. Sampling Strategy: Recognizing the challenge of sampling for unsupervised learning, the study introduces a semi-hard negative sampling strategy. This approach balances the presence of misleading hard negatives and the ineffectiveness of easy negatives, promoting more reliable and unbiased learning processes.
3. Decision Boundary Design: The study highlights the importance of margin settings in loss functions, adapting MarginNCE to suit both intra- and inter-image consistency branches. They find that intra-image learning benefits from stricter margins, whereas inter-image learning, given its inherent noise, requires more lenient margins.

Empirical Validation

InterCLR exhibits substantial improvements over established intra-image invariance learning methods when transferred to various downstream tasks, such as image classification, low-shot classification, semi-supervised learning, and object detection. Notably, it outperforms in efficiency and effectiveness compared to its baselines, pre-trained within a constrained computational budget.

Implications and Future Directions

The research underlines the potential of leveraging inter-image invariance in advancing unsupervised representation learning. By effectively navigating challenges around label dynamics and sampling strategy, it sets the stage for future explorations in:

• Extending InterCLR's framework to other modalities or dense prediction tasks.
• Investigating more sophisticated clustering techniques or adaptive margin strategies.
• Evaluating performance under varying resource constraints, particularly in batch size and epochs, which is crucial for scalability and generalization.

The findings provide pivotal insights into refining contrastive learning approaches, ensuring a balanced integration of intra- and inter-image statistics to foster robust and generalizable representations. Given the rapid advancements in AI, such comprehensive studies form the foundation for more resilient and efficient unsupervised learning frameworks in practice.