Few-Shot Learning refers to the problem of training a model f_θ so that it can quickly solve new tasks T_j with only a few examples S_j = {(x_i, y_i)}_{i=1}^k, where k is small (e.g., 1 to 5).
Formally, this is often framed within meta-learning, where there is a distribution over tasks p(T) and the goal is to minimize the expected loss function over new tasks:
\[ \min_\theta \mathbb{E}_{T \sim p(T)} \left[ \mathcal{L}_{T} (f_\theta) \right] \]
Well-known algorithms include Model-Agnostic Meta-Learning (MAML), which generalizes to new tasks through a few gradient updates, as well as Prototypical Networks, which use representations as class centroids in the feature space.
Few-Shot Learning addresses central challenges of generalization and data efficiency in machine learning systems.
Definition:
“Few-Shot Learning is a learning paradigm in which models are trained to generalize to new tasks with only very few training examples.”
Source:
Wang, Y., Yao, Q., Kwok, J. T., & Ni, L. M. (2020). Generalizing from a Few Examples: A Survey on Few-Shot Learning. ACM Computing Surveys, 53(3), 63.