[Paper Review] Bayesian Model-Agnostic Meta-Learning

Taesup Kim, Jaesik Yoon+4 more|arXiv (Cornell University)|Jun 11, 2018

Domain Adaptation and Few-Shot Learning26 references163 citations

TL;DR

Proposes BMAML, a Bayesian extension of MAML that uses Stein Variational Gradient Descent to obtain a flexible task-posterior during fast adaptation and introduces a Chaser loss to mitigate meta-overfitting, applicable to supervised, active learning, and reinforcement learning tasks.

ABSTRACT

Learning to infer Bayesian posterior from a few-shot dataset is an important step towards robust meta-learning due to the model uncertainty inherent in the problem. In this paper, we propose a novel Bayesian model-agnostic meta-learning method. The proposed method combines scalable gradient-based meta-learning with nonparametric variational inference in a principled probabilistic framework. During fast adaptation, the method is capable of learning complex uncertainty structure beyond a point estimate or a simple Gaussian approximation. In addition, a robust Bayesian meta-update mechanism with a new meta-loss prevents overfitting during meta-update. Remaining an efficient gradient-based meta-learner, the method is also model-agnostic and simple to implement. Experiment results show the accuracy and robustness of the proposed method in various tasks: sinusoidal regression, image classification, active learning, and reinforcement learning.

Motivation & Objective

• Motivate robust few-shot learning by addressing model uncertainty via Bayesian inference.
• Develop a gradient-based meta-learning method that can capture non-Gaussian task posterior uncertainty.
• Enable efficient fast adaptation and principled meta-update to avoid meta-overfitting.
• Demonstrate applicability of Bayesian MAML to supervised, active learning, and reinforcement learning tasks.

Proposed method

• Introduce Bayesian Fast Adaptation (BFA) using SVGD to sample from p(θτ|Dτtrn,Θ0) with multiple particles Θ0.
• Use SVGD to propagate particles through task-train data to form task-specific posteriors Θτ(Θ0).
• Define a meta-loss based on the disparity between the fast-adapted posterior and a higher-fidelity posterior (leader) obtained via additional SVGD steps with augmented data.
• Propose the Chaser Loss to guide Θ0 so that the chaser quickly follows the leader, reducing meta-overfitting.
• Share parameters across particles to reduce space complexity in large networks (e.g., shared feature extractor, per-particle classifier).
• Demonstrate applicability to supervised learning, active learning, and reinforcement learning through experiments.

Experimental results

Research questions

• RQ1 Can a gradient-based meta-learning method capture complex, non-Gaussian task posterior uncertainty beyond a simple Gaussian approximation?
• RQ2 Does a Bayesian meta-learning framework with an efficient fast adaptation and a principled meta-update reduce meta-level overfitting and improve robustness across tasks?
• RQ3 How does a Bayesian ensemble via SVGD perform in sinusoidal regression, image classification, active learning, and reinforcement learning settings compared to standard MAML?
• RQ4 Can parameter sharing among particles make Bayesian MAML scalable to large networks without sacrificing performance?

Key findings

• BMAML outperforms EMAML and standard MAML in sinusoidal regression, especially under higher uncertainty and with fewer training tasks or shots.
• In mini-Imagenet classification, BMAML with shared feature extractors outperforms EMAML across multiple particle counts, and shows robustness with fewer meta-training tasks.
• BMAML enables active learning by selecting high-uncertainty samples, yielding better performance than EMAML.
• In reinforcement learning, BMAML with SVPG-based updates achieves superior exploration and performance over EMAML, particularly with TRPO as meta-updater.
• Parameter sharing across particles mitigates space complexity while maintaining or improving performance.
• BMAML provides improved prediction accuracy, robustness to overfitting, and efficient exploration across tasks.