Probabilistically grounded unsupervised training of neural networks

The chapter is a survey of probabilistic interpretations of artificial neural networks (ANN) along with the corresponding unsupervised learning algorithms. ANNs for estimating probability density functions (pdf) are reviewed first, including parametric estimation via constrained radial basis functions and nonparametric estimation via multilayer perceptrons. The approaches overcome the limitations of traditional statistical estimation methods, possibly leading to improved pdf models. The focus is then moved from pdf estimation to online neural clustering, relying on maximum-likelihood training. Finally, extension of the techniques to the unsupervised training of generative probabilistic hybrid paradigms for sequences of random observations is discussed