Finn Chelsea Abbeel Pieter Levine Sergey

Abstract

We propose an algorithm for meta-learning that is model-agnostic, in thesense that it is compatible with any model trained with gradient descent andapplicable to a variety of different learning problems, includingclassification, regression, and reinforcement learning. The goal ofmeta-learning is to train a model on a variety of learning tasks, such that itcan solve new learning tasks using only a small number of training samples. Inour approach, the parameters of the model are explicitly trained such that asmall number of gradient steps with a small amount of training data from a newtask will produce good generalization performance on that task. In effect, ourmethod trains the model to be easy to fine-tune. We demonstrate that thisapproach leads to state-of-the-art performance on two few-shot imageclassification benchmarks, produces good results on few-shot regression, andaccelerates fine-tuning for policy gradient reinforcement learning with neuralnetwork policies.