Source-linked AI summary
Meta-SGD: Learning to Learn Quickly for Few-Shot Learning
Zhenguo Li, Fengwei Zhou, Fei Chen, Hang Li
TL;DR
Few-shot learning requires accurate, rapid adaptation from limited data, while conventional deep learning is data-hungry and slow to adapt. The paper introduces Meta-SGD, an SGD-like meta-learner that learns initialization, update direction, and learning rate for one-step adaptation. Across regression, classification, and reinforcement learning, Meta-SGD achieves highly competitive or state-of-the-art few-shot results, while requiring substantial computation for meta-training.
Problem
Few-shot learning needs a principled way to determine initialization, update direction, and learning rate so learners generalize from very few examples and adapt quickly.
Method
Meta-SGD meta-learns an initialization θ and element-wise parameters α that jointly determine a learner's update direction and learning rate for one-step adaptation.
Results
Meta-SGD achieves highly competitive or state-of-the-art few-shot performance across regression, classification, and reinforcement learning, including one-step adaptation.
Takeaways & Limitations
Learning all optimizer ingredients across related tasks gives Meta-SGD higher capacity than methods that learn only initialization or follow a fixed gradient-based update.
Takeaways & Limitations
Meta-learning can require far more computation than traditional learning, especially when each task has substantially more than a few examples or requires a large learner.
Abstract
from arXiv · showhide
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.
1 Introduction
Few-shot learning seeks data-efficient, rapid adaptation, motivating meta-learning approaches that learn learning strategies across related tasks. Meta-SGD extends optimizer-based meta-learning by learning initialization, update direction, and learning rate while adapting learners quickly.
- Motivation: Few-shot learning addresses the data hunger and slow adaptation of deep learning systems trained on isolated tasks from scratch.The motivation is especially strong when data are limited or environments change dynamically.
- Meta-learning: Meta-learning trains a meta-learner across related tasks so it can guide task-specific learners toward stronger generalization.Tasks provide separate training and testing sets, allowing the meta-learner to optimize expected generalization across the task space.
- Meta-SGD: Figure 1 separates gradual meta-learning across tasks in (θ, α) from rapid task-specific learning in learner space θ.The two levels operate on different time scales.
- Meta-learner design: Meta-learners include recurrent models, metrics, and optimizers, with optimizers offering a route to learning explicit learner-update procedures.Recurrent meta-learners are versatile but less comprehensible, while metrics are mainly suited to non-parametric learners.
- Meta-SGD: Meta-SGD learns initialization, update direction, and learning rate through meta-learning while retaining an SGD-like adaptation process.Its update direction and learning rate are learned rather than manually chosen or constrained to the ordinary gradient direction.
2 Related Work
Related work approaches few-shot learning through generative models, meta-learned metrics, and recurrent or optimizer-like meta-learners. These methods differ in their use of prior structure, learner type, and learned update strategy.
- Generative models: Generative few-shot methods can use compositional knowledge of related concepts to learn new concepts from a single example.The cited approach represents handwritten-character concepts with probabilistic programs and models pen-stroke composition.
- Metric-based methods: Metric-based meta-learning modifies distances between examples and is suited to non-parametric learners such as k-nearest-neighbor classifiers.Matching training and testing conditions is identified as important for few-shot learning.
- Recurrent optimizers: Recurrent meta-learners can model adaptive optimization and have been used as generic SGD-like optimizers for updating learners such as CNNs.LSTM is reported as the strongest recurrent architecture among those compared in the cited work.
3 Meta-SGD
Meta-SGD is a meta-learned optimizer that initializes and adapts a learner in one step by learning initialization, update direction, and learning rate across related tasks. Its objective minimizes expected testing loss after adapting from each task’s training set.
- Scope: The proposed meta-learner applies to supervised learning and reinforcement learning.The section introduces supervised learning as the running case and discusses reinforcement learning separately.
- Meta-Learner: Meta-SGD is designed to address few-shot settings where choosing initialization, update direction, learning rate, and stopping is difficult because of overfitting.The paper motivates learning these factors for generalization rather than merely fitting the small task dataset.
- Meta-Learner: Meta-SGD learns an optimizer’s initialization, update direction, and learning rate end to end for few-shot learning.These are the three key optimizer ingredients, and Meta-SGD determines them through meta-learning rather than hand-designed rules.
- Meta-Learner: Given a few examples, Meta-SGD initializes the learner with θ and adapts it to θ′ in one step using α ◦∇L_T(θ).The learned adaptation direction can differ from the ordinary gradient, while α implicitly determines the learning rate.
- Meta-training: Meta-training samples related tasks, adapts learners on each training set, and minimizes their expected loss on corresponding testing sets.The resulting objective is differentiable with respect to θ and α and can be optimized with SGD.
10: end
For reinforcement learning, Meta-SGD treats each task as a Markov decision process and adapts a stochastic policy using a policy-gradient update. Meta-training evaluates the adapted policy on new trajectories, while comparison with MAML and Meta-LSTM frames its capacity and complexity trade-offs.
- Reinforcement Learning: In reinforcement learning, each task is a Markov decision process and the learner is a stochastic policy f_θ.The task includes states, actions, transition and initial-state distributions, horizon, reward, and discount factor.
- Reinforcement Learning: Meta-SGD samples N1 trajectories, computes an empirical policy gradient, updates the policy, then measures generalization loss on N2 new trajectories.This mirrors the supervised-learning procedure while using policy-gradient methods for adaptation.
- Reinforcement Learning: The reinforcement-learning objective minimizes expected task loss after updating θ to θ − α ◦∇L_T(θ).The optimization is expressed over tasks sampled from p(T).
- Comparison: Compared with MAML, Meta-SGD learns update direction and learning rate in addition to initialization, whereas Meta-LSTM learns the same ingredients with higher complexity.The comparison presents Meta-SGD as higher-capacity than MAML and simpler than Meta-LSTM.
4 Experimental Results
Meta-SGD is evaluated on few-shot regression, classification, and 2D navigation, using one-step adaptation and comparisons with MAML and other state-of-the-art models. Across these experiments, it achieves strong results while learning initialization, update direction, and learning rate jointly.
- Regression: Meta-SGD consistently outperforms MAML on K-shot sine-curve regression across the evaluated cases.It learns initialization, update direction, and learning rate simultaneously, whereas MAML primarily learns initialization and treats the learning rate as a hyper-parameter.
- Regression: After one update using only five examples, Meta-SGD adapts more quickly to the target sine curve and continues improving with additional meta-testing examples.The comparison remains favorable even when the training examples occupy only one half of the input range.
- Experimental Setup: The classification and navigation experiments evaluate one-step adaptation across unseen classes or sampled tasks using task-specific training data and subsequent testing.Classification uses Omniglot and MiniImagenet benchmarks, while navigation updates policy networks with vanilla policy gradients and the meta-learner.
- Classification: Meta-SGD is slightly better than state-of-the-art models on all Omniglot classification tasks.For 5-shot meta-testing, the reported results were obtained using 1-shot meta-training.
- Classification: Meta-SGD outperforms all other evaluated models on every MiniImagenet classification case.The paper highlights that this performance is achieved while adapting the learner in just one step.
- Reinforcement Learning: Meta-SGD obtains relatively higher returns than MAML on both 2D navigation task families.The tasks vary either the goal position alone or both the start and goal positions, and evaluation averages returns over 600 sampled tasks with 95% confidence intervals.
5 Conclusions
Meta-SGD is an easily trainable SGD-like meta-learner that learns initialization, update direction, and learning rate end-to-end, enabling one-step adaptation with state-of-the-art few-shot results. Future work must address the computational demands of large-scale meta-learning and generalization to unseen setups and domains.
- Meta-SGD learns initialization, update direction, and learning rate end-to-end, giving it higher capacity than optimizer-like meta-learners.These optimizer components are learned in a single meta-learning process.
- One-step adaptation with Meta-SGD achieves new state-of-the-art results on few-shot regression, classification, and reinforcement learning.
- Large-scale meta-learning remains a future challenge because training many learners creates substantially greater computational demand than traditional learning.The burden is especially pronounced when each task has far more than a few examples and requires a large learner.
- Meta-learning also needs greater versatility for unseen problem setups, new task domains, and multi-tasking meta-learners.The paper identifies this generalization capacity as an important problem for expanding practical value.