An empirical study of unsupervised few-shot learning that utilizes self-supervised representation learning

초록

Unsupervised Few-Shot Learning (UFSL) aims to learn feature representations from unlabeled data for rapid adaptation to new tasks. Within this field, methods based on Self-Supervised Learning (SSL) have become the state-of-the-art approach, driving recent progress by learning feature representations directly from unlabeled data. However, the selection of a specific SSL method is often treated as an arbitrary design choice, and an extensive analysis of their performance is notably absent. This issue is exacerbated as most SSL methods are benchmarked on large-scale datasets, a setting that fundamentally differs from the data-scarce conditions inherent to UFSL. This paper addresses these deficiencies through this extensive empirical study examining how different SSL training dynamics affect performance on diverse UFSL conditions. To assess their performance on small-scale data, we evaluate seven representative SSL methods from four diverse learning paradigms, many of which are foundational to existing state-of-the-art UFSL models. The evaluation spans nine few-shot benchmark datasets, covering both in-domain and cross-domain scenarios. Furthermore, we introduce a novel evaluation protocol using systematically scaled-down training sets to assess model robustness in data-scarce conditions more severe than those of conventional few-shot benchmarks. Through extensive experimentation with statistical validation, this study reveals that the performance of SSL methods is dictated by their training dynamics and the target task's characteristics; no single approach is universally superior. This research establishes a benchmark to clarify these performance differences and provides empirical guidelines for SSL method selection in UFSL applications.

키워드