Adaptive Sensing, Active Learning, and Experimental Design

The fields of active learning, adaptive sensing and sequential experimental design have seen a growing interest over the last decades in a number of communities, ranging from machine learning and statistics to biology and computer vision. Broadly speaking, all active and adaptive approaches focus on closing the loop between data analysis and acquisition. Said in a different way the goal is to use information collected in past samples to adjust and improve the future sampling and learning processes, in the spirit of the twenty questions game. These fields typically address the problem in very diverse ways, and using different problem formulations. The main objective of this workshop is to bring these communities together, share ideas and knowledge, and cross-fertilize the various fields.

Most of the theoretical work in the area of adaptive sensing and active learning has remained quite distant from the realm of practical applications (with a few notable exceptions). In less-than-ideal settings, many modeling assumptions are only approximately true, and hence closed-loop (active) methods as described need to be very robust in other to: (i) guarantee consistency, in the sense that the proposed method must not fail dramatically; (ii) improve on the performance of open-loop (passive) procedures whenever favorable conditions are met. Due to the feedback nature of closed-loop procedures these are often prone to failure when modeling assumptions are only approximately met, and this has been observed by many when deploying practical algorithms. By bringing together both theoreticians and practitioners from the fields of computer vision and robotics, statistics, signal and information processing and machine learning it will be possible to identify promising directions for active learning at large, and address these points in a satisfactory way.