Efficient scalable algorithms for adaptive data collection

In many applications, data-driven discovery is limited by the rate of data collection: the skilled labor it takes to operate a pipette, the time to execute a long-running physics simulation, the patience of an infant to remain still in an MRI, or the cost of labeling large corpuses of complex images. A powerful paradigm to extract the most information with such limited resources is active learning, or adaptive data collection, which leverages already-collected data to guide future measurements in a closed loop. But being convinced that data-collection should be adaptive is not the same thing as knowing how to adapt in a way that is both sample efficient and reliable. In this talk, I will present several examples of my provably reliable — and practical — adaptive data collection algorithms being applied in the real-world. In particular, I will show how my adaptive algorithms are used each week to crowd-source the winner of the New Yorker Magazine Cartoon Caption Contest. I will also discuss my application of adaptive learning concepts at Google to accelerate the tuning of deep networks in a highly parallelized environment of thousands of GPUs.
Kevin Jamieson is a postdoctoral researcher working with Professor Benjamin Recht in the Department of Electrical Engineering and Computer Sciences at the University of California, Berkeley. He is interested in the theory and practice of machine learning algorithms that sequentially collect data using an adaptive strategy. This includes active learning, multi-armed bandit problems, and stochastic optimization. Kevin received his Ph.D. from the University of Wisconsin – Madison under the advisement of Robert Nowak. Prior to his doctoral work, Kevin received his B.S. from the University of Washington, and an M.S. from Columbia University, both in electrical engineering.