Meta-SGD: Learning to Learn Quickly for Few-Shot Learning

TLDR

Few‑shot learning is difficult for algorithms that learn each task from scratch, whereas meta‑learning trains a meta‑learner on many related tasks to enable rapid, accurate learning of new tasks with few examples, making the choice of meta‑learner critical. The paper proposes Meta‑SGD, an SGD‑like meta‑learner that can initialize and adapt any differentiable learner in a single step for supervised and reinforcement learning. The method trains a meta‑learner that, for any differentiable learner, outputs an initialization and a one‑step update rule, enabling rapid adaptation in supervised and reinforcement learning settings. Meta‑SGD outperforms LSTM and MAML, being simpler, more efficient, and having higher capacity by jointly learning initialization, update direction, and learning rate, and achieves competitive performance on regression, classification, and reinforcement learning few‑shot tasks.

Abstract

Few-shot learning is challenging for learning algorithms that learn each task in isolation and from scratch. In contrast, meta-learning learns from many related tasks a meta-learner that can learn a new task more accurately and faster with fewer examples, where the choice of meta-learners is crucial. In this paper, we develop Meta-SGD, an SGD-like, easily trainable meta-learner that can initialize and adapt any differentiable learner in just one step, on both supervised learning and reinforcement learning. Compared to the popular meta-learner LSTM, Meta-SGD is conceptually simpler, easier to implement, and can be learned more efficiently. Compared to the latest meta-learner MAML, Meta-SGD has a much higher capacity by learning to learn not just the learner initialization, but also the learner update direction and learning rate, all in a single meta-learning process. Meta-SGD shows highly competitive performance for few-shot learning on regression, classification, and reinforcement learning.