On First-Order Meta-Learning Algorithms

TLDR

Meta‑learning seeks to train an agent that can quickly adapt to new tasks drawn from a task distribution, and this work builds on and generalizes first‑order MAML, extending prior results from Finn et al. The study analyzes a family of first‑order meta‑learning algorithms that learn a parameter initialization enabling rapid fine‑tuning on new tasks. The algorithms comprise first‑order MAML and the newly introduced Reptile, which iteratively samples tasks, trains on them, and shifts the initialization toward the resulting task‑specific weights. First‑order meta‑learning algorithms achieve strong performance on established few‑shot classification benchmarks, and the authors offer theoretical insights into their effectiveness.

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.