Source-linked AI summary
Bayesian Model-Agnostic Meta-Learning
Taesup Kim, Jaesik Yoon, Ousmane Dia, Sungwoong Kim, Yoshua Bengio, Sungjin Ahn
TL;DR
Few-shot task learning creates substantial uncertainty, and Gaussian approximations may not adequately represent the resulting task posterior. The paper proposes Bayesian MAML, combining gradient-based meta-learning with nonparametric variational inference, and reports efficient, accurate, robust performance across several task types while acknowledging practical limits on posterior coverage.
Problem
Few-shot learning induces substantial uncertainty, motivating methods that can represent task posteriors beyond Gaussian approximations for robust meta-learning.
Method
Bayesian MAML combines Bayesian fast adaptation using nonparametric variational inference with gradient-based meta-learning and a Chaser loss for meta-update.
Results
Experiments report that the method is efficient, accurate, robust, and applicable to sinusoidal regression, image classification, reinforcement learning, and active learning.
Takeaways & Limitations
The model-agnostic method provides flexible uncertainty modeling while retaining gradient-based meta-learning and applicability across supervised, active-learning, and reinforcement-learning tasks.
Takeaways & Limitations
With finite particles and update steps, the method does not fully cover complex multimodal task posteriors, and full coverage is not achievable or desired in practice.
Abstract
from arXiv · showhide
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.
1 Introduction
Few-shot learning requires extracting statistical regularities from past tasks because learning from very few examples creates uncertainty and can lead neural networks to overfit. Bayesian MAML addresses this by combining Bayesian fast adaptation with gradient-based meta-learning, supporting flexible uncertainty modeling across diverse tasks.
- Few-shot learning aims to transfer statistical regularities from past tasks so new tasks do not begin entirely from scratch.
- Few-shot task learning induces substantial uncertainty, while existing few-shot algorithms can overfit.
- Bayesian MAML learns an approximate posterior for an unseen task during fast adaptation, providing access to uncertainty.
- The method combines nonparametric variational inference with gradient-based meta-learning to capture complex uncertainty structures while remaining simple to implement.
- Experiments evaluate the method on sinusoidal regression, image classification, reinforcement learning, and active learning, reporting efficiency, accuracy, and robustness.
2 Preliminaries
The preliminaries formalize meta-learning over related few-shot tasks and review MAML's gradient-based task adaptation and meta-update procedure. They also introduce SVGD, a nonparametric inference method whose particles represent flexible approximate distributions and interact through kernel-weighted gradients and repulsion.
- Meta-learning samples related tasks from a shared task distribution and divides task data into training, validation, and test subsets.
- MAML is model-agnostic because it relies on gradient optimization without requiring additional parameters or model modifications.
- MAML adapts each task from a generic initialization using a small number of gradient steps, then updates that initialization using task-validation losses.
- SVGD approximates a target distribution with particles without restricting the approximation to a tractable parametric family.
- Each SVGD particle uses kernel-weighted gradients from other particles, while a repulsive term prevents particle collapse.
- With one particle, SVGD reduces to standard gradient ascent, while multiple particles provide a distributional approximation applicable to reinforcement learning.
3 Proposed Method
The method combines gradient-based meta-learning with nonparametric variational inference to rapidly approximate task posteriors and represent complex uncertainty. A Chaser loss extends Bayesian treatment to the meta-update to retain uncertainty and reduce meta-level overfitting.
- Bayesian Fast Adaptation: Bayesian MAML learns to quickly obtain an approximate posterior for an unseen task while retaining efficient gradient-based meta-learning.The approach maintains initial particles and adapts them with SVGD, producing task-wise posterior samples from a shared initialization.
- Bayesian Fast Adaptation: SVGD-based Bayesian fast adaptation captures task-posterior uncertainty beyond a point estimate or simple Gaussian approximation.The nonparametric particle representation can model complex uncertainty structure while remaining compatible with gradient backpropagation.
- Bayesian Fast Adaptation: The initial particles are jointly optimized so that few SVGD steps produce task posteriors quickly and with a small number of samples.This procedure is called Bayesian Fast Adaptation and can be viewed as an interacting Bayesian ensemble.
- Bayesian Meta-Learning with Chaser Loss: The Chaser loss minimizes dissimilarity between task-train and further-updated particle sets to shape an initialization close to the task posterior.The method uses chaser particles after n updates and leader particles after s additional updates, with gradients stopped through the leaders.
- Bayesian Meta-Learning with Chaser Loss: Because the leader converges toward the posterior rather than empirical risk minimization, the Chaser loss retains uncertainty during meta-update and prevents meta-level overfitting.The method requires maintaining M model instances, although parameter sharing is introduced for large models.
4 Related Works
Prior work cast meta-learning in probabilistic terms, while Bayesian MAML addresses limitations of existing posterior approximations with a more flexible and scalable approach.
- Probabilistic formulations of meta-learning predate recent deep-learning and few-shot-learning advances.The paper situates its approach within earlier probabilistic modeling studies and renewed interest driven by neural networks and few-shot datasets.
- Grant et al. reinterpret MAML hierarchically but use local Laplace inference, K-FAC-based second-order optimization, and a fixed-variance isotropic Gaussian for fast adaptation.
- The paper motivates Bayesian MAML as a response to limitations of simple Gaussian approximations for skewed few-shot posteriors.
5 Experiments
Experiments evaluate BMAML across regression, classification, active learning, and reinforcement learning, generally showing greater robustness or performance than EMAML, especially under uncertainty.
- Experimental setup: BMAML is evaluated on sinusoidal regression, miniImagenet classification, active learning, and reinforcement learning against EMAML with matched architectures.
- Sinusoidal regression: BMAML provides significantly better robustness and accuracy than MAML and EMAML across sinusoidal-regression settings with varying task counts and K-shot sizes.MAML and EMAML severely overfit in high-uncertainty settings, while additional BMAML particles tend to improve performance.
- Classification: BMAML provides more accurate miniImagenet predictions than EMAML with both M = 5 and M = 10 particles.Both methods perform slightly worse with 10 particles than with 5, and BMAML remains robust when training tasks decrease from 800K to 10K.
- Active learning: Active learning with Bayesian fast adaptation consistently outperforms EMAML, with the performance gap increasing as more examples are added.BMAML selects unlabeled examples by maximum predictive entropy to reduce uncertainty.
- Reinforcement learning: BMAML generally outperforms EMAML in locomotion and 2D Navigation, with particularly consistent gains under TRPO meta-updates and on goal-direction tasks.The authors associate goal-direction gains with improved exploration, while goal-velocity gains are smaller because exploration becomes less useful after reaching the target velocity.
- Reinforcement learning: More particles do not necessarily improve reinforcement-learning performance, consistent with reported instability in SVGD.
6 Discussions
The discussion explains why BMAML uses SVGD, parameter-space Chaser loss, and few adaptation steps, while clarifying that full posterior coverage is neither practically achievable nor always desired.
- Inference design: BMAML could use differentiable-sample inference methods beyond SVGD, but SVGD is chosen because jointly updating particles captures distributions quickly with few steps.The discussion contrasts this with MCMC, which requires more iterations for mixing and longer backpropagation through the chain.
- Optimization order: When the inner update is included, TRPO, Chaser, and Reptile are respectively third-, second-, and first-order methods.
- Meta-update design: The Chaser loss uses parameter-space distance because it reduces prediction-forward computation and performs better empirically in reinforcement learning.Prediction-space distance remains an alternative and performs similarly on other tasks.
- Posterior approximation: Small SVGD step counts are intended to locate broadly useful regions of task-posterior space rather than fully represent each complex posterior.The discussion states that complete coverage would require infinitely many particles and update steps and is not desired in the small-data, large-network setting.
7 Conclusion
The paper combines Stein Variational Gradient Descent with gradient-based meta-learning to build a model-agnostic Bayesian method evaluated across supervised, active, and reinforcement learning tasks. Its main limitations are ensemble-related computational costs and sensitivity to SVGD kernel parameters.
- The proposed Bayesian gradient-based meta-learning method combines Stein Variational Gradient Descent with gradient-based meta-learning in a probabilistic framework.
- Experiments evaluate the model-agnostic method on supervised learning, active learning, and reinforcement learning, reporting prediction accuracy, robustness to overfitting, and efficient exploration.
- Space/time complexity remains proportional to the number of particles, although parameter sharing mitigates this ensemble-method limitation.
- SVGD performance can be sensitive to kernel-function parameters, motivating future work on adapting those parameters during fast adaptation.
A.2 Classification
The classification setup uses few-shot 5-way tasks with one adaptation example per class and evaluates on unseen instances. Active learning repeatedly selects the pool example with maximum predictive entropy before updating the posterior particles.
- Classification: Each miniImagenet task randomly selects 5 classes, uses one instance per class for adaptation, and evaluates on unseen instances from those classes.
- Classification: The classification meta-validation dataset contains one example per class for the 5-way 1-shot setting.
- Classification: The classification posterior differs from regression because its data likelihood uses softmax.
- Active Learning: Active learning initializes particles from the meta-learned parameters and updates them with Stein Variational Gradient Descent using the labeled task data.
- Active Learning: At each active-learning iteration, the method selects the unlabeled example maximizing predictive entropy, requests its label, and adds it to the labeled dataset.
C.2 Used Methods
The reinforcement-learning methods use SVPG for Bayesian inner updates and compare TRPO- and chaser-loss-based meta-updates with corresponding variational-gradient baselines. The chaser loss keeps uncertainty while converting the meta-update toward matching a further-updated agent.
- Compared Methods: The experiments compare SVPG-TRPO with VPG-TRPO and SVPG-Chaser with VPG-Reptile; VPG-TRPO is MAML when M = 1.
- SVPG-TRPO: SVPG supplies the inner update, while TRPO performs the meta-update using few-shot trajectories and validation data.
- SVPG-Chaser: The chaser-loss method uses SVPG for the inner update and applies the chaser loss during meta-update to maintain uncertainty.
- SVPG-Chaser: In reinforcement learning, the chaser loss changes the meta-update from a policy-gradient problem to one resembling imitation learning.
- VPG-Reptile: The Reptile-style variant computes a chaser parameter and minimizes its Euclidean distance to the global parameter while avoiding gradients through the chaser.
C.3 Experimental Details
Across the reinforcement-learning experiments, inner updates use one step with learning rate 0.1, while meta-update rates and SVPG exploration strengths vary by task and method.
- Training Settings: Inner updates use learning rate 0.1 and one update across locomotion and 2D-Navigation experiments.Meta-update learning rates are 0.1 for ant goal direction and 0.01 for the other experiments.
- Training Settings: The SVPG exploration parameter η is set to 0.1 for several experiments and 1.0 for others.
C.4 Additional Experiment Results
The additional experiments cover 2D navigation and qualitative regression adaptation, illustrating performance across reinforcement-learning and few-shot regression settings.
- 2D Navigation: 2D navigation tasks require an agent to reach randomly assigned goals within a unit square using clipped velocity actions.Rewards are based on negative squared distance, with termination upon reaching the goal or after 100 timesteps.
- 2D Navigation: Figure 5 compares SVPG-TRPO with VPG-TRPO and SVPG-Chaser with VPG-Reptile across three random seeds.The two panels correspond to these method comparisons, respectively.
- Regression: Figure 6 presents qualitative regression examples using randomly sampled tasks with 10 examples and 10 gradient updates for adaptation.The figure focuses on adapted predictions under a fixed 10-shot, 10-update setting.