Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks

The paper proposes an algorithm for meta-learning that is model-agnostic, meaning it is compatible with any model trained with gradient descent and applicable to a variety of learning problems including classification, regression, and reinforcement learning. The parameters of the model are explicitly trained such that a small number of gradient steps with a small amount of training data from a new task will produce good generalization performance on that task; in effect, the method trains the model to be easy to fine-tune.

The approach leads to state-of-the-art performance on two few-shot image classification benchmarks, produces good results on few-shot regression, and accelerates fine-tuning for policy gradient reinforcement learning with neural network policies. This addresses the central goal of meta-learning, training a model on a variety of learning tasks so that it can solve new tasks using only a small number of training samples, across diverse problem settings.