Preprint
Computer Vision

A Comprehensive Survey of Few-shot Learning: Evolution, Applications, Challenges, and Opportunities

Yisheng Song(East China Normal University), Ting Wang(East China Normal University), Puyu Cai(Michigan State University), Subrota Kumar Mondal(Macau University of Science and Technology), Jyoti Prakash Sahoo(Siksha O Anusandhan University)
February 4, 2023ACM Computing Surveys716 citations
arXivPDFDOI

716

Citations

9

Influential Citations

ACM Computing Surveys

Venue

2023

Year

Abstract

Few-shot learning (FSL) has emerged as an effective learning method and shows great potential. Despite the recent creative works in tackling FSL tasks, learning valid information rapidly from just a few or even zero samples remains a serious challenge. In this context, we extensively investigated 200+ FSL papers published in top journals and conferences in the past three years, aiming to present a timely and comprehensive overview of the most recent advances in FSL with a fresh perspective and to provide an impartial comparison of the strengths and weaknesses of existing work. To avoid conceptual confusion, we first elaborate and contrast a set of relevant concepts including few-shot learning, transfer learning, and meta-learning. Then, we inventively extract prior knowledge related to few-shot learning in the form of a pyramid, which summarizes and classifies previous work in detail from the perspective of challenges. Furthermore, to enrich this survey, we present in-depth analysis and insightful discussions of recent advances in each subsection. What is more, taking computer vision as an example, we highlight the important application of FSL, covering various research hotspots. Finally, we conclude the survey with unique insights into technology trends and potential future research opportunities to guide FSL follow-up research.

Analysis

Why This Paper Matters

Few-shot learning (FSL) is a critical area in AI, aiming to enable models to learn from very few examples, akin to human learning. This survey is particularly significant because it synthesizes a large body of recent work (200+ papers) into a coherent framework, addressing the fragmentation in the field. By clarifying the often-confused concepts of few-shot learning, transfer learning, and meta-learning, the authors provide a solid conceptual foundation that is essential for both newcomers and experienced researchers.

The introduction of the 'knowledge pyramid' is a novel contribution that organizes FSL methods based on the type of prior knowledge used and the challenges they address. This framework not only helps in understanding existing approaches but also highlights gaps and opportunities for future research. The survey's focus on computer vision applications makes it highly relevant to a broad audience, as vision is a primary domain for FSL.

Technical Contributions

  • Conceptual Clarification: The paper meticulously distinguishes few-shot learning from transfer learning and meta-learning, reducing confusion and enabling more precise research.
  • Knowledge Pyramid: A novel taxonomy that categorizes FSL methods based on prior knowledge (e.g., data, model, algorithm) and challenges (e.g., sample insufficiency, domain shift).
  • Comprehensive Review: Covers a wide range of FSL techniques, including metric-based, optimization-based, and hallucination-based methods, with in-depth analysis.
  • Application Showcase: Highlights key computer vision applications such as image classification, object detection, and semantic segmentation, demonstrating FSL's practical utility.
  • Future Directions: Identifies promising research avenues, including cross-domain FSL, self-supervised FSL, and FSL in natural language processing.

Results

The survey does not present new experimental results but rather synthesizes existing literature. It provides a structured overview of the state of the art, noting that recent advances have significantly improved FSL performance, particularly in image classification benchmarks like miniImageNet and tieredImageNet. However, the paper emphasizes that challenges remain, such as handling domain shifts and scaling to more complex tasks. The authors also note that many methods achieve strong results on standard benchmarks but struggle in real-world scenarios.

Significance

This survey is a valuable resource for the AI community, offering a clear roadmap for FSL research. Its knowledge pyramid framework can serve as a common language for discussing FSL methods, facilitating collaboration and comparison. By highlighting open challenges and opportunities, the paper encourages further innovation, potentially accelerating progress in few-shot learning and its applications across various domains. The survey's impact is likely to be long-lasting, as it consolidates knowledge and guides future research directions.

few-shot-learningmeta-learningtransfer-learningcomputer-visionsurveyknowledge-pyramiddeep-learning
Few Shot LearningComputer Vision

Related Papers