Learning Quadratic Forms by Density Estimation and its Applications to Image Coding
Hauke Bartsch, Sepp Hochreiter, Klaus Obermayer · 2003
We develop a novel method for source separation and apply it to natural images. It is a specialization of independent factor analysis (IFA) but overcomes generic IFA problems and nds many independent sources in few observations. A fast and robust EM learning algorithm produces an over-complete basis. Compared to standard approaches our method generates superior codes in terms of population sparseness and dispersity. The algorithm learns features which possess properties that are observed in simple as well as complex cells found in V1.