Abstract
Learning to infer Bayesian posterior from a few-shot dataset is an importantstep towards robust meta-learning due to the model uncertainty inherent in theproblem. In this paper, we propose a novel Bayesian model-agnosticmeta-learning method. The proposed method combines scalable gradient-basedmeta-learning with nonparametric variational inference in a principledprobabilistic framework. During fast adaptation, the method is capable oflearning complex uncertainty structure beyond a point estimate or a simpleGaussian approximation. In addition, a robust Bayesian meta-update mechanismwith a new meta-loss prevents overfitting during meta-update. Remaining anefficient gradient-based meta-learner, the method is also model-agnostic andsimple to implement. Experiment results show the accuracy and robustness of theproposed method in various tasks: sinusoidal regression, image classification,active learning, and reinforcement learning.