Self-supervised visual learning in the low-data regime: a comparative evaluation

2024-04-26

Neurocomputing

AI Summary

This study investigates few-shot visual learning under extreme label scarcity (<1% ImageNet labels), systematically evaluating the generalization and robustness of self-supervised learning (SSL) methods. Within a unified benchmark, it conducts the first horizontal comparison of prominent SSL frameworks—including SimCLR, BYOL, DINO, and MAE—under linear probing, fine-tuning, and semi-supervised transfer protocols. Methodologically, it rigorously controls for architecture, data augmentation, and downstream evaluation to isolate SSL-specific effects. Key findings reveal that contrastive approaches significantly outperform generative ones; teacher-student-based methods exhibit superior resilience to label scarcity, yielding up to +12.3% absolute improvement in downstream classification accuracy. The work empirically delineates performance boundaries and failure modes of SSL under ultra-low-data regimes, uncovering critical trade-offs between representation quality and label efficiency. These results provide evidence-based guidance for SSL method selection and architectural design in few-shot settings, offering novel insights into the interplay between self-supervision paradigms and data-limited generalization.

Technology Category

• Machine Learning: Unsupervised & Self-Supervised Learning
• Computer Vision: Representation Learning for Vision
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Web Mining and Content Analysis: Robustness and generalizability of Web mining methods
• Security and Privacy: Large-scale security measurements

Problem

• Self-Supervised Learning
• Effectiveness Evaluation
• Limited Data

Innovation

• Self-Supervised Learning
• Limited Data
• Domain-specific Performance