On First-Order Meta-Learning Algorithms

TL;DR

This paper analyzes and introduces first-order meta-learning algorithms, including a new method called Reptile, demonstrating their effectiveness and providing theoretical insights into their success in few-shot learning tasks.

Contribution

It introduces Reptile, a new first-order meta-learning algorithm, and offers theoretical analysis explaining why first-order methods perform well.

Findings

Reptile is effective for few-shot classification tasks.

First-order meta-learning algorithms perform comparably to more complex methods.

Theoretical analysis sheds light on the success of first-order algorithms.

Abstract

This paper considers meta-learning problems, where there is a distribution of tasks, and we would like to obtain an agent that performs well (i.e., learns quickly) when presented with a previously unseen task sampled from this distribution. We analyze a family of algorithms for learning a parameter initialization that can be fine-tuned quickly on a new task, using only first-order derivatives for the meta-learning updates. This family includes and generalizes first-order MAML, an approximation to MAML obtained by ignoring second-order derivatives. It also includes Reptile, a new algorithm that we introduce here, which works by repeatedly sampling a task, training on it, and moving the initialization towards the trained weights on that task. We expand on the results from Finn et al. showing that first-order meta-learning algorithms perform well on some well-established benchmarks for few-shot classification, and we provide theoretical analysis aimed at understanding why these algorithms work.