Section: New Software and Platforms
SketchMLBox
Keyword: Clustering
Scientific Description: The SketchMLbox is a Matlab toolbox for fitting mixture models to large collections of training vectors using sketching techniques. The collection is first compressed into a vector called sketch, then a mixture model (e.g. a Gaussian Mixture Model) is estimated from this sketch using greedy algorithms typical of sparse recovery. The size of the sketch does not depend on the number of elements in the collection, but rather on the complexity of the problem at hand [2,3]. Its computation can be massively parallelized and distributed over several units. It can also be maintained in an online setting at low cost. Mixtures of Diracs ("K-means") and Gaussian Mixture Models with diagonal covariance are currently available, the toolbox is structured so that new mixture models can be easily implemented
Functional Description: Matlab toolbox for fitting mixture models to large databases using sketching techniques.
-
Authors: Nicolas Keriven, Nicolas Tremblay and Rémi Gribonval
-
Publications: Sketching for Large-Scale Learning of Mixture Models - Compressive K-means - Spikes super-resolution with random Fourier sampling - Sketching for large-scale learning of mixture models - Blind Source Separation Using Mixtures of Alpha-Stable Distributions - Sketching for Large-Scale Learning of Mixture Models - Compressive Gaussian Mixture Estimation by Orthogonal Matching Pursuit with Replacement