Introduction to active learning for classification

Active learning is, in essence, about maximizing the performance gain (of an underlying, supervised machine learning algorithm) per label. Some possible (query) strategies and the core problem emerging, being sampling bias, are discussed. Then an effective way of circumventing it by balancing exploring and exploiting accordingly is given via Thompson sampling. Lastly active learning is put to use for time series classification. Concepts are outlined in a heuristic manner.

Speaker: Simon Bachhuber is currently writing his master's thesis in physics at the University of Regensburg. The topic is active learning and its practical implementations. The ulterior motive is to (ultimately) implement active learning in collaboration with Syskron to improve label efficiency.