Abstract

The field of ML has advanced considerably in recent years, but mostly in well-defined domains using huge amounts of human-labeled training data. Machines can recognize objects in images and translate text, but they must be trained with more images and text than a person can see in nearly a lifetime. Humans, on the other hand, can learn from far fewer examples, generalize well across tasks and modalities, and perform better than machines at most tasks, especially in complex and unpredictable situations. To address this gap, this talk focuses on active ML that close the loop on machine learning, sensing and data collection, and human labeling. Standard (passive) machine learning involves designing a classification rule based on a randomly selected training dataset. Active machine learning algorithms automatically and adaptively select the most informative data for labeling so that human time is not wasted labeling irrelevant or trivial examples. The aim is to make ML as efficient and robust as possible, with a minimal amount of human supervision and assistance. This talk describes ongoing theoretical and experimental work in several areas of active learning

Short biography: Rob is the McFarland-Bascom Professor in Engineering at the University of Wisconsin-Madison, where his research focuses on signal processing, machine learning, optimization, and statistics.