Probabilistic Latent Variable Decompositions for Image and Audio Analysis

Опубликовано 6 сентября 2016, 6:39

In this talk we present a model which can decompose probability densities into sets of shift invariant components. We will show how this model is very well suited for audio and image problems and will demonstrate its applications on complex data sets. We will also explore the relationship of this model to well known decomposition methods (such as PCA/ICA/NMF/PARAFAC) and argue about its appropriateness when dealing with density-like representations. We will also present various extensions to this model to allow it to perform sparse coding, discover Markovian components, model a priori dispositions and temporal relationships through standard statistical models.