The Weizmann Institute of Science Faculty of Mathematics and Computer Science Vision and Robotics Seminar Fred Hamprecht Heidelberg University will speak on Active Learning and Large-Scale Segmentation for Connectomics Lecture Hall, Room 1, Ziskind Building on Thursday, April 8, 2010 12:00 - 13:00 Abstract: In the first part of this talk (1), I will present a method based on the ``random forest" ensemble classifier that allows to estimate not only the posterior probability of each class at a given query point in feature space, but the full distribution of that posterior probability. This added information allows to raise an alarm when the query lies outside the region that is sampled well by the training set, thus making for a natural fusion of outlier detection and classification. In addition, I will point to how such a distributional estimate can be used in active learning schemes. ``Connectomics" has the aim of extracting a wiring diagram from very high resolution 3D microscopic images of a nervous system. The intricate geometry, imperfect SNR and large amounts of data (10 billion voxels per image) make for a difficult segmentation problem. In the second part of the talk (2), I will explain our use of probabilistic graphical models in a hierarchical segmentation scheme, and give details on the required geometry extraction which we perform with log-linear complexity. (1) Joint work with Jens Roeder (HCI), Boaz Nadler (Weizmann Institute) (2) Joint work with Bjoern Andres, Thorben Kroeger, Ullrich Koethe (HCI) and Winfried Denk (MPI Heidelberg) --------------------------------------------------------- Technion Math Net-2 (TECHMATH2) Editor: Gershon Wolansky Announcement from: Diana Mandelik