ABSTRACT
This talk expounds on the natural synergy of meta and active learning.

The general scheme of metalearning is to swap hard-coded procedures that output models for trainable models (that output models). For example, one can use many related classification problems to train a model which outputs a classifier given an input set of labeled examples. In effect, the trained "set2func" model replaces the role of, e.g., SGD in producing a classifier for each problem.

From a Bayesian perspective, the set2func model can be interpreted as encoding both a prior over classifiers and a procedure for performing inference w.r.t. the learned prior given some labeled data. Existing approaches to meta classification can be interpreted as learning a prior over classifiers, which is co-adapted with some mechanism for performing MAP inference. E.g., the "base" parameters learned by MAML give the mean of a Gaussian prior over parameterizations of the "base" model. Adapting the base parameters to a new problem instance via SGD can be interpreted as approximate posterior inference (see, e.g. "Early Stopping as Non-parametric Variational Inference").

Current metalearning methods perform MAP inference, which can be particularly limiting in the case of tiny sets of labeled data. Consider the Omniglot meta classification task. If we have unlabeled data for all classes in the current problem, but labeled data only for a subset of the classes, a MAP-based model won't explicitly represent the "cluster-based" label assignments on which this task is based. A "Bayesian" set2func model could express this property of the underlying task distribution through the classifiers assigned high likelihood by its cross-task prior and per-task posteriors.

The role of active learning, in the setting of Bayesian metalearning, is to collect labeled examples for the current problem in a way that maximally reduces entropy in the posterior distribution over classifiers. Having an explicit representation of the relevant posterior should permit more effective active learning, and adapting the prior to the task distribution should permit tighter, more-accurate per-task posteriors. Conversely, encouraging the set2func model to represent its prior and posteriors in a way that facilitates active learning may improve the quality of the learned prior and posteriors.

Developing models which perform well in the few-shot setting (metalearning), and which efficiently collect information to improve task performance (active learning), will be critical to success in settings like "life-long learning".

This talk discusses these themes, and presents some related concrete results.