Mean Field ICA

Kaare Brandt Petersen

    Research output: Book/ReportPh.D. thesisResearch

    130 Downloads (Pure)

    Abstract

    This thesis describes investigations and improvements of a technique for Independent Component Analysis (ICA), called "Mean Field ICA". The main focus of the thesis is the optimization part of the algorithm, the so-called "EM algorithm". Using different approaches it is demonstrated that the EM algorithm is inefficient and therefore an improper choice for a certain class of models. As an alternative, the so-called "Easy Gradient Algorithm" is presented, which makes it possible to apply advanced optimization techniques using the computationally simple E-step and M-step of the EM algorithm. The Easy Gradient Recipe is applicable to a wide selection of models. Furthermore, the Mean Field ICA model is extended to incorporate ltering over time in a so-called "convolutive ICA" model. Finally, by using mixture of Gaussians as source priors, the generative and ltering approach to ICA is compared in the overcomplete setting, i.e. the situation in which there is more sources than sensors.
    Original languageEnglish
    Place of PublicationKgs. Lyngby
    PublisherTechnical University of Denmark
    Publication statusPublished - Jan 2006
    SeriesIMM-PHD-2005-148

    Fingerprint Dive into the research topics of 'Mean Field ICA'. Together they form a unique fingerprint.

    Cite this