On the Bayesian Interpretation of Robust Regression Neural Networks

The aim of this work is to search for intuitive interpretations of regularized regression procedures within the framework of Bayesian inference. First, the paper considers Bayesian estimation of parameters of the linear regression model. Second, regularized neural networks are explained to correspond to the Bayesian approach obtained under specific assumptions. The contribution is a unique compact look on training neural networks with available prior information, i.e. a likelihood-based perspective of training neural networks. Attention is also paid to very recently proposed regularized versions of robust neural networks; as a novelty, these are expressed by means of quasi-likelihood and their connection to Bayesian reasoning is discussed as well.