lowshot

Low-shot learning is a subfield of machine learning focused on enabling models to recognize and adapt to new concepts from a small number of labeled examples. In practice, low-shot is often discussed in the same regime as few-shot learning, typically involving one to five labeled examples per class, though the exact definitions vary. The goal is to generalize beyond seen classes to unseen classes with limited supervision, contrasting with traditional supervised learning that relies on large labeled datasets.

Approaches in low-shot learning commonly combine meta-learning, metric learning, and transfer learning. Meta-learning methods aim to learn how to adapt quickly to new tasks, using episodic training that mimics few-shot evaluation. Metric-based methods compute a discriminative space where new class representations can be formed from few examples, e.g., prototypes or relation scores. Transfer learning leverages rich pretraining on large datasets and adapts the model with minimal fine-tuning. Data augmentation, self-supervised pretraining, and synthetic data generation via GANs or diffusion models are also used to supplement scarce labels. Some methods integrate semi-supervised components to exploit unlabeled data from the target domain.

Common benchmarks include image classification datasets such as MiniImageNet, tieredImageNet, CIFAR-FS, and FC-100, as well as Omniglot for character recognition tasks. Researchers also apply low-shot techniques to natural language processing, robotics, and other modalities. Challenges include domain shift, long-tail distributions, and scalability to many classes or tasks. Evaluation typically reports accuracy under various K-way N-shot settings to quantify rapid adaptation with minimal labeled data.