퓨샷 학습

Few-shot is used across methods that supply only a small number of task examples, often after substantial prior learning. In an N-way K-shot classification episode, N counts classes and K counts labeled support examples per class. That notation describes the task’s support data rather than the model’s total pretraining corpus. In-context few-shot prompting supplies demonstrations in the input of an already trained model. The cited language-model study evaluates this mode without task-specific gradient updates during the prompt; examples condition the prediction rather than rewrite weights. Fine-tuning on a small dataset is a different procedure even when both are called few-shot adaptation. Prior training can make representations or patterns useful for a new task, but a few examples do not guarantee that the model infers the intended labels or rules. Prototypical Networks studies a metric-based alternative. A learned embedding maps labeled support items into a space, and each class prototype is the mean of its support embeddings. A query is classified using distances to those prototypes. If one class has support vectors (0,0) and (2,0), its arithmetic-mean prototype is (1,0). The support items used to build the prototype and the query items used to assess classification have different roles. Training episodes in the paper keep those sampled sets separate. For in-context models, demonstration order can also affect results: Fantastically Ordered Prompts studies large variation among permutations in particular text-classification setups. Selecting a favorable order should not be confused with proof of generalization on an untouched evaluation set.