Bayesian Model-Agnostic Meta-Learning

Abstract

Due to the inherent model uncertainty, learning to infer Bayesian posterior from a few-shot dataset is an important step towards robust meta-learning. In this paper, we propose a novel Bayesian model-agnostic meta-learning method. The proposed method combines efficient gradient-based meta-learning with nonparametric variational inference in a principled probabilistic framework. Unlike previous methods, during fast adaptation, the method is capable of learning complex uncertainty structure beyond a simple Gaussian approximation, and during meta-update, a novel Bayesian mechanism prevents meta-level overfitting. Remaining a gradientbased method, it is also the first Bayesian model-agnostic meta-learning method applicable to various tasks including reinforcement learning. Experiment results show the accuracy and robustness of the proposed method in sinusoidal regression, image classification, active learning, and reinforcement learning.

Efficiently Identifying Task Groupings for Multi-Task Learning

“…"Lookahead" methods in deep learning can often be characterized by saving the current state of the model, applying one or more gradient updates to a subset of the parameters, reloading the saved state, and then leveraging the information learned from the future state to modify the current set of parameters. This approach has been used extensively in the metalearning [16,42,7,17,29], optimization [41,21,55,23,24], and recently auxiliary task learning domains [34]. Unlike the above mentioned methods which look into the future to modify optimization processes, our work adapts this central concept to the multi-task learning domain to characterize task interactions and assign tasks to groups of networks.…”

Meta-learning Amidst Heterogeneity and Ambiguity

“…Many variants have emerged to balance generalization and customization in a task-adaptive manner. To begin with a generalization perspective, [13,27] suggested probabilistic extensions through the hierarchical Bayesian model and Stein variational gradient descent (SVGD) [37]. In addition, [49] conducted the inner-loop on the low-dimensional latent embedding space, and [70] proposed the meta-regularization that was built on information theory.…”

Transfer Bayesian Meta-Learning Via Weighted Free Energy Minimization

“…(17) Finally, the hyper-posterior q W F EM −GP (θ|D 1:N ) is used in lieu of the corresponding PACOH-GP hyperposterior in (14) in order to define the predictive distribution.…”

“…In practice, for both PACOH-GP and WFEM-GP, the expectations in the predictive distributions (14) need to be approximated. As detailed in the supplementary materials A, this can be done by evaluating the maximum of the hyperposteriors, and by plugging this value into the predictive distribution p θ (t(x) = t|D) -an approach we refer to as maximum a posteriori (MAP).…”

“…Bayesian approaches to meta-learning have become increasingly popular in the recent years due to their important advantages in quantifying uncertainty and model selection [14,15]. In this context, both parametric methods such as Bayesian neural networks and non-parametric methods like GPs have both been successfully applied to real-world applications.…”