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

By Aniruddh Raghu, Maithra Raghu, Samy Bengio, Oriol Vinyals. 2019

756 citations recorded by the source corpus.

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.

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Representation & Self-Supervised Learning

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Representation & Self-Supervised Learning

Citations in the corpus

Rethinking Few-Shot Image Classification: a Good Embedding Is All You Need? (2020, methodology): “Unlike other methods [9, 39], we do not ﬁne-tune the embedding model f φ during the meta-testing stage.”
On the Outsized Importance of Learning Rates in Local Update Methods (2020, methodology): “The core idea has inspired a number of extensions (Antoniou et al., 2019; Nichol et al., 2018; Rusu et al., 2019; Grant et al., 2018; Rajeswaran et al., 2019; Raghu et al., 2020), which broadly use a two-level optimization structure to perform meta-learning.”
Disentangling Transfer in Continual Reinforcement Learning (2022, background): “Additionally, we show that the concept of feature reuse which is often utilized to explain supervised transfer learning [28, 33] might not be directly applicable in RL.”
Meta-Learning with Less Forgetting on Large-Scale Non-Stationary Task Distributions (2022, positive): “Supervised meta-learning methods without unlabeled data , including (1) gradient-based meta-learning ANIL [45], which is a simplified model of MAML [22]; (2) metric-based meta-learning Proto-typical Network ( ProtoNet ) [54].”
Generalized and Incremental Few-Shot Learning by Explicit Learning and Calibration without Forgetting (2021, background): “In FSL [9, 34] and IL [26, 36] multiple works show that adaptive representations can be beneficial for learning novel classes, specifically to fine-tune the parameters of the representation.”
What is being transferred in transfer learning? (2020, background): “[35] evaluated the role of feature reusing in meta-learning.”
Meta-TTS: Meta-Learning for Few-Shot Speaker Adaptive Text-to-Speech (2021, methodology): “The concept of ANIL is “feature reuse”.”
Online Fast Adaptation and Knowledge Accumulation: a New Approach to Continual Learning (2020, background): “Instead of adapting all the network layers, ANIL adapts only the network head towards the new task.”
Few-Shot Learning via Learning the Representation, Provably (2020, background): “Recent work Raghu et al. (2019) empirically suggested that the effectiveness of the popular meta-learning algorithm Model Agnostic Meta-Learning (MAML) is due to its ability to learn a useful representation.”
Bilevel Optimization: Convergence Analysis and Enhanced Design (2020, methodology): “…(AID-BiO with a constant number of inner-loop steps as in our analysis), AID-BiO-increasing (AID-BiO with an increasing number of inner-loop steps under analysis in Ghadimi & Wang 2018), and two popular meta-learning algorithms MAML 2 (Finn et al., 2017) and ANIL 3 (Raghu et al., 2019).”
FedBABU: Towards Enhanced Representation for Federated Image Classification (2021, methodology): “…accuracy is calculated by replacing the trained head with the nearest template following previous studies (Kang et al., 2019; Snell et al., 2017; Raghu et al., 2019; Oh et al., 2021): (1) Using the trained global model and training data set on each client, the representations of each class…”
Transformers learn in-context by gradient descent (2022, methodology): “learning a linear model on a meta-learned deep data representations (Finn et al., 2017; Finn & Levine, 2018; Gordon et al., 2019; Lee et al., 2019; Rusu et al., 2019; Raghu et al., 2020; von Oswald et al., 2021).”
A Comprehensive Survey of Few-shot Learning: Evolution, Applications, Challenges, and Opportunities (2022, neutral): “Such as MAML++ [96], First-order MAML (FOMAML) [95], Meta-SGD [97], TAML [98], iMAML [99], iTMAML [100].”
Meta-Learning Requires Meta-Augmentation (2020, background): “The large interest in the ﬁeld has spurred the creation of meta-learning benchmarks [20, 27, 34, 36, 38], investigations into tuning few-shot models [4], and analysis of what these models learn [26].”
On Modulating the Gradient for Meta-learning (2020, methodology): “As paper [30] suggests that the representations in last layers undergo significant changes during adaptation, we apply updates to the last two convolutional layers of the base network.”
Bootstrapped Meta-Learning (2021, methodology): “Standard MG does not have this property, but several ﬁrst-order approximations have been proposed: we allow the MG approach to switch from a full meta-gradient to either the FOMAML approximation (Finn et al., 2017) or the ANIL approximation (Raghu et al., 2020).”