Few-shot and zero-shot learning share hidden trade-offs visible only through optimization.

Computer vision models often struggle when they must recognize items with few or no examples. People usually treat few-shot learning and zero-shot learning as completely separate tasks. This review looks at both as a single math problem: making predictions with missing or scarce class evidence. Instead of focusing only on model architectures, the authors break methods down into four steps. They track the available information, the estimated variables, the mathematical goals, and the solvers that calculate the answers. They found that no single method works best everywhere. Every choice involves a trade-off. For example, flexible updates cause unstable estimates, and relying on large pretrained systems makes it difficult to trace where evidence comes from. Ultimately, both fields are connected estimation problems. Future progress requires safer use of uncertain prior knowledge, label-free model selection, and stable, efficient adaptation.

Why it matters

Researchers can now see why one method fails while another succeeds, by tracking the same four optimization coordinates across both few-shot and zero-shot settings. This means future systems won't just try random algorithms — they'll choose based on known trade-offs like estimator variance and representation bias. For anyone building AI that must learn from just a handful of images or unfamiliar categories, this framework offers a systematic way to diagnose and compare approaches.

Questions answered