New Prior Model for Bayesian Neural Networks Learning and Application to Classification of Tissues in Mammographic Images

摘要

The Multilayer Perceptron (MLP) is the most useful artificial neural network to estimate the functional structure in the non- linear systems, but the determination of its architecture, weights and hyperparameters is a fundamental problem due to their direct impact on the network generalization ability and convergence. The Bayesian approach provides a naturel way to adjust the weights decay parameters automatically that's give the best generalization. This paper presents an improvement of a prior model to construct a new objective function for learning neural network in Bayesian perspectives with the Hybrid Monte Carlo (HMC) algorithm. The proposed model is applied to classification of Normal, Benign and Malignant Tissues in Mammographic images. Compared to the other regularization model the numerical results illustrate the advantages of our approach.