Kevin Miller - Ensuring Exploration and Exploitation in Graph-Based Active Learning

Abstract: Semi-supervised learning methods leverage both labeled and unlabeled data to achieve an accurate classification with significantly fewer training points as compared to supervised learning methods that only use labeled data. Simultaneously, the choice of training points can significantly affect classifier performance, especially due to the limited size of the training set of labeled data in semi-supervised learning. Active learning seeks to judiciously select a limited number of query points from the unlabeled data that will inform the machine learning task at hand. These points are then labeled by an expert, or human in the loop, with the aim of significantly improving the classifier performance.

Uncertainty sampling has traditionally been the de facto, simplest acquisition function for active learning in semi-supervised learning. Comparatively cheap to compute and straightforward to interpret, uncertainty sampling has been known to suffer from myopic sampling bias that fails to properly explore the extent of geometric structure of the dataset prior to exploiting learning decision boundaries. As such, most work in active learning for graph-based learning has focused on the design of more intricate acquisition functions that are explorative in nature, though are almost always more costly to compute and therefore do not scale well to larger datasets. We will present two different uncertainty sampling methods for graph-based active learning that (1) have provable guarantees for exploration and exploitation in graph-based semi-supervised learning given an assumption on the clustering structure of the data and (2) are efficient to compute to allow for scalability in downstream applications.

Видео Kevin Miller - Ensuring Exploration and Exploitation in Graph-Based Active Learning канала One world theoretical machine learning