On August 5, 2015, Dr. Georg Krempl (Otto-von-Guericke University of Magdeburg) will give a talk about "Probabilistic Active Learning in Classification" in the context of the SFB 901.
Abstract:
Facing ever increasing volumes of data but limited human supervision and annotation capacities, active learning approaches gain in importance, as they allocate these capacities such that the most insightful information is acquired. This talk presents the probabilistic active learning approach, with particular focus on a novel fast, closed-form solution for cost-sensitive classification tasks. This active learning approach combines the fast asymptotic runtime of popular heuristics like uncertainty sampling with the direct optimisation of the expected performance gain known from expected error reduction approaches. This is done by modelling for each given labelling candidate both its label's realisation and the true posterior in its neighbourhood as random variables. PAL then computes the expected performance gain not only over the label realisation, but also over the true posterior, thereby accounting for the quantity of already obtained similar information. The experimental evaluation of this approach on several synthetic and real-world data sets, in combination with different classifier technologies, shows promising results, which are further discussed.