Addressing data scarcity with Few-Shot Learning: a systematic literature review

Abstract

Recent advancements in Machine Learning and Deep Learning have yielded remarkable achievements across a wide range of fields, including object detection, image classification, and Natural Language Processing (NLP). This progress has been largely driven by the availability of large-scale annotated datasets, which enable high accuracy and strong generalization. However, such performance levels remain out of reach in settings characterized by limited data and high annotation costs. This challenge has motivated the emergence of Few-Shot Learning (FSL), a novel branch of artificial intelligence that seeks to replicate the human brain’s capacity to learn effectively from minimal examples. This systematic literature review offers a comprehensive examination of FSL techniques, drawing on relevant research, spanning from 2015 to 2025. The reviewed works are organized according to their learning strategies for data-limited settings, covering meta-learning models, hybrid models, and non-meta learning approaches, alongside a survey of notable real-world applications where FSL has been successfully deployed. Beyond exploring the progress within this field of research, this review advances several critical and underexplored findings. First, the evidence accumulated across the reviewed literature suggests that representation quality consistently outweighs algorithmic sophistication as the primary driver of performance gains, a finding with direct implications for how future FSL research should prioritize its efforts. Second, the episodic training paradigm, which underlies a large body of FSL methodology, exhibits structural misalignments with real-world deployment conditions. Third, the field exhibits a systematic bias toward natural image benchmarks, which constrains the transferability of algorithmic advances to other critical domains that operate on different data modalities. Fourth, the dominant evaluation practice of averaged accuracy across tasks is shown to be an inadequate proxy for practical deployment performance. Together, these limitations reveal important structural gaps in the current FSL landscape. In light of these findings, the review identifies concrete avenues for improvement aimed at bridging these gaps, enhancing model robustness, and ultimately enabling more reliable FSL deployment in real-world applications.