Output details

Online learning and generalization of parts-based image representations by non-negative sparse autoencoders

<24> The paper introduces an incremental online-algorithm to obtain non-negative sparse factorizations for image data through decomposition in overcomplete basis functions. It addresses the problem that standard factorization methods require computationally costly iterative re-optimization when generalizing to unseen data. It outperforms standard offline methods in benchmarks. The method has been picked up in an entirely different community working on graph-segmentation in a paper in Chaos. It is shown that it can segment weighted graphs into communities and that it even produces soft assignments that perfectly match ground truth on standard benchmarks in that domain.