Few-Shot Learning
Training ML models to generalize accurately from only a handful of labeled examples.
Few-shot learning is a machine learning paradigm in which models recognize patterns and make accurate predictions from only a small number of labeled training examples, typically between one and five per class. Conventional supervised learning, by contrast, demands thousands or millions of examples to achieve reliable performance. The underlying challenge is bridging the gap between human learning, where a child can identify a new animal from a single picture, and the data hunger of standard deep learning systems.
The dominant approaches to few-shot learning fall into three broad families. Meta-learning (or "learning to learn") trains a model across many related tasks so it develops an inductive bias that allows rapid adaptation to new tasks with minimal data. MAML (Model-Agnostic Meta-Learning) is a canonical example. Metric-based methods such as Siamese Networks, Matching Networks, and Prototypical Networks learn an embedding space where examples from the same class cluster together, enabling classification by nearest-neighbor comparison. Transfer learning approaches fine-tune large pretrained models on small target datasets, reusing representations already learned from massive corpora.
Few-shot learning gained traction in the mid-2010s alongside advances in meta-learning and the proliferation of large pretrained models. The introduction of benchmark datasets like Omniglot and miniImageNet gave researchers standardized evaluation grounds, accelerating progress. More recently, large language models such as GPT-3 demonstrated few-shot capabilities through in-context learning, adapting to new tasks from a few prompt examples without weight updates, which changed how the community thinks about the concept.
The practical importance of few-shot learning is considerable. In medicine, rare disease classification may yield only dozens of confirmed cases. In linguistics, thousands of languages lack sufficient digital text for standard training. In personalization, individual user data is sparse by nature. By enabling models to generalize from limited signal, few-shot learning extends machine learning into domains where large-scale data collection is economically, ethically, or logistically difficult.
Sources
- Matching Networks for One Shot Learning
arXiv · Jan 1, 2016
Research this in Signals
Signals turns a topic into a sourced research record you can inspect and rerun. Your first scan is free, and this one starts with Few-Shot Learning already loaded, so edit it or scan as is.
