BYY learning, regularized implementation, and model selection on modular networks with one hidden layer of binary units

摘要

The BYY learning has been extended to a modular system, with developments on not only regularized implementation via either normalization or data smoothing, but also the least complexity based model selection. Moreover, both unsupervised and supervised learning have been specifically investigated on networks with one hidden layer of binary units. Adaptive EM-like learning algorithms are provided for implementing regularized learning with either automatic model selection during parameter learning or post-learning selection criteria for both the number of individual nets and the number of hidden units. Furthermore, discussions are made on application of rule extraction for tackling the paradox between conflict and redundancy.