Probabilistic and unsupervised machine learning for auditory data and pattern recognition

摘要

I will introduce and discuss probabilistic data models and their applications to auditory data and to general pattern recognition tasks. The introduction will reflect the view that powerful data models are able to extract the true compositional nature of data, which allows for a decomposition into their structural primitives. Probabilistic sparse coding and probabilistic version of non-negative matrix factorization (NMF) will be the first concrete models that are introduced and discussed. The state-of-the-art of these models will then be used to point to recent generalization directions. Two of these, translation invariant versions and deep generalizations, will be discussed in more detail. I will work out the benefits and challenges such novel approaches face, and I will discuss their crucial differences compared to supervised deep neural networks. Finally, I briefly discuss semi-supervised approaches, a field where modern unsupervised and modern supervised Machine Learning algorithms come together, compete and where they are combined.