Rapid Learning or Feature Reuse? Towards Understanding the Effectiveness of MAML

Abstract

An important research direction in machine learning has centered around developing meta-learning algorithms to tackle few-shot learning. An especially successful algorithm has been Model Agnostic Meta-Learning (MAML), a method that consists of two optimization loops, with the outer loop finding a meta-initialization, from which the inner loop can efficiently learn new tasks. Despite MAML's popularity, a fundamental open question remains -- is the effectiveness of MAML due to the meta-initialization being primed for rapid learning (large, efficient changes in the representations) or due to feature reuse, with the meta initialization already containing high quality features? We investigate this question, via ablation studies and analysis of the latent representations, finding that feature reuse is the dominant factor. This leads to the ANIL (Almost No Inner Loop) algorithm, a simplification of MAML where we remove the inner loop for all but the (task-specific) head of a MAML-trained network. ANIL matches MAML's performance on benchmark few-shot image classification and RL and offers computational improvements over MAML. We further study the precise contributions of the head and body of the network, showing that performance on the test tasks is entirely determined by the quality of the learned features, and we can remove even the head of the network (the NIL algorithm). We conclude with a discussion of the rapid learning vs feature reuse question for meta-learning algorithms more broadly.

arXiv commentICLR 2020

References in corpus (20)

• Model-Agnostic Meta-Learning for Fast Adaptation of Deep Networks
• Prototypical Networks for Few-shot Learning
• Transfusion: Understanding Transfer Learning for Medical Imaging
• Similarity of Neural Network Representations Revisited
• SVCCA: Singular Vector Canonical Correlation Analysis for Deep Learning Dynamics and Interpretability
• Meta-Learning with Latent Embedding Optimization
• A Closer Look at Few-shot Classification
• Fast Context Adaptation via Meta-Learning
• Recasting Gradient-Based Meta-Learning as Hierarchical Bayes
• Unsupervised Learning via Meta-Learning
• Convergent Learning: Do different neural networks learn the same representations?
• Deep Meta-Learning: Learning to Learn in the Concept Space
• Meta-Learning and Universality: Deep Representations and Gradient Descent can Approximate any Learning Algorithm
• How to train your MAML
• A Closer Look at Deep Learning Heuristics: Learning rate restarts, Warmup and Distillation
• Insights on representational similarity in neural networks with canonical correlation
• Few-shot Text Classification with Distributional Signatures
• Universality and individuality in neural dynamics across large populations of recurrent networks
• Identifying and Controlling Important Neurons in Neural Machine Translation
• Meta-Learning Representations for Continual Learning

Cited by in corpus (29)

• Rethinking Few-Shot Image Classification: a Good Embedding Is All You Need?
• Multimodal Few-Shot Learning with Frozen Language Models
• Exploiting Shared Representations for Personalized Federated Learning
• On the Theory of Transfer Learning: The Importance of Task Diversity
• Inexact-ADMM Based Federated Meta-Learning for Fast and Continual Edge Learning
• learn2learn: A Library for Meta-Learning Research
• Probing transfer learning with a model of synthetic correlated datasets
• On Fast Adversarial Robustness Adaptation in Model-Agnostic Meta-Learning
• Making Graph Neural Networks Worth It for Low-Data Molecular Machine Learning
• Context Meta-Reinforcement Learning via Neuromodulation
• Learning Adaptive Exploration Strategies in Dynamic Environments Through Informed Policy Regularization
• Modeling and Optimization Trade-off in Meta-learning
• Memory-Based Optimization Methods for Model-Agnostic Meta-Learning and Personalized Federated Learning
• ARCADe: A Rapid Continual Anomaly Detector
• Modular Meta-Learning with Shrinkage
• Learning to Learn with Feedback and Local Plasticity
• A Sample Complexity Separation between Non-Convex and Convex Meta-Learning
• Meta Learning for Few-Shot One-class Classification
• Sign-MAML: Efficient Model-Agnostic Meta-Learning by SignSGD
• One Shot Learning for Speech Separation
• Impact of Representation Learning in Linear Bandits
• Information-Theoretic Analysis of Epistemic Uncertainty in Bayesian Meta-learning
• Provable Representation Learning for Imitation Learning via Bi-level Optimization
• HetMAML: Task-Heterogeneous Model-Agnostic Meta-Learning for Few-Shot Learning Across Modalities
• Model-Agnostic Meta-Learning using Runge-Kutta Methods
• Uniform Priors for Data-Efficient Transfer
• Sample Efficient Linear Meta-Learning by Alternating Minimization
• DropClass and DropAdapt: Dropping classes for deep speaker representation learning
• Meta Learning in the Continuous Time Limit