[Paper Review] Meta-SGD: Learning to Learn Quickly for Few-Shot Learning

Zhenguo Li, Fengwei Zhou+2 more|arXiv (Cornell University)|Jul 31, 2017

Domain Adaptation and Few-Shot LearningComputer Science26 references839 citations

TL;DR

Meta-SGD learns an SGD-like meta-learner that initializes and adapts any differentiable learner in one step by learning initialization, update direction, and per-parameter learning rates, achieving strong results across regression, classification, and reinforcement learning.

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.

Motivation & Objective

• Motivate fast adaptation from limited data in few-shot learning using a learnable optimization strategy.
• Propose an SGD-like meta-learner that jointly learns initialization and a per-parameter update rule.
• Demonstrate that learning update direction and learning rates increases capacity over prior meta-learners.
• Show competitive or state-of-the-art performance across regression, classification, and reinforcement learning with one-step adaptation.

Proposed method

• Introduce Meta-SGD, an optimizer-like meta-learner that updates the learner as θ' = θ − α ∘ ∇Ltrain(θ) with α learned alongside θ.
• Learn both the initialization θ and the update scaling α end-to-end across many related tasks sampled from a task distribution p(T).
• Train the meta-learner by minimizing the expected generalization loss on test sets after one-step adaptation across tasks (meta-training).
• Apply the framework to supervised learning (regression and classification) and reinforcement learning, using SGD-like updates in all cases.
• Compare Meta-SGD to Meta-LSTM and MAML, highlighting higher capacity and simpler, more efficient training.

Experimental results

Research questions

• RQ1 How can a meta-learner be designed to initialize and adapt a learner for new tasks from few-shot data?
• RQ2 Does learning the update direction and per-parameter learning rates improve few-shot generalization beyond learning only initialization?
• RQ3 Can a single forward pass adaptation (one-step) achieve competitive or superior performance across regression, classification, and reinforcement learning tasks?
• RQ4 How does Meta-SGD compare to existing meta-learners (MAML, Meta-LSTM) in terms of training efficiency and performance?

Key findings

• Meta-SGD achieves highly competitive results with one-step adaptation across regression, classification, and reinforcement learning.
• In regression, Meta-SGD outperforms MAML on sine-curve few-shot tasks across 5-, 10-, and 20-shot settings.
• In Omniglot and MiniImagenet classification, Meta-SGD attains state-of-the-art or competitive accuracy, often surpassing MAML and Meta-LSTM across multiple settings.
• In reinforcement learning, Meta-SGD yields higher returns than MAML on 2D navigation tasks with both fixed and varying starts.
• Meta-SGD trains and adapts faster than Meta-LSTM and MAML due to learning initialization, update direction, and learning rates within a single meta-learning process.