VQ 8-10 - Vector Quantization (14 min)

Vector Quantization - Sec. 8-10 (14 min) LECTURE VIDEO 3 of Lecture Videos on Vector Quantization Prof. Laurenz Wiskott Institut für Neuroinformatik Ruhr-Universität Bochum, Germany, EU Vector quantization is a method from machine learning to represent high dimensional data in a more compact form, thereby saving transmission or storage capacity. Data points are represented by reference vectors, such that they can be indexed and do not have to be fully represented. The main problem in vector quantization is to place the reference vectors such that the reconstruction error is minimal. This set of videos gives a basic introduction into the subject. ==================================================================== LECTURE [57 min] ---------------------------------------------------------------------- VIDEO 1 (21 min, on YouTube) 1 General idea 2 Error function 3 Optimal assignment given the reference vectors 3.1 Voronoi tesselation 3.2 Delaunay triangulation 4 Optimal reference vectors given an assignment 5 A simple vector quantization algorithm ---------------------------------------------------------------------- VIDEO 2 (22 min, on YouTube) 6 An online learning rule for vector quantization 7 Learning rates 7.1 Constant learning rate 7.2 1/t-decaying learning rate 7.3 Exponentially decaying learning rate ---------------------------------------------------------------------- VIDEO 3 (14 min, on YouTube) 8 Soft competitive learning 9 Growing networks + 10 Towards a neural implementation + 10.1 Euclidian distance and inner product 10.2 Winner takes all 10.3 (Adaptive resonance theory) 11 (Demo) ---------------------------------------------------------------------- ==================================================================== This and more material is also available at https://www.ini.rub.de/PEOPLE/wiskott/Teaching/Material/index.html