TY - GEN

T1 - Learning of Bayesian discriminant functions by a layered neural network

AU - Ito, Yoshifusa

AU - Srinivasan, Cidambi

AU - Izumi, Hiroyuki

PY - 2008

Y1 - 2008

N2 - Learning of Bayesian discriminant functions is a difficult task for ordinary one-hidden-layer neural networks, because the teacher signals are dichotomic random samples. When the neural network is trained, the parameters, the weights and thresholds, are usually all supposed to be optimized. However, those included in the activation functions of the hidden-layer units are optimized at the second step of the BP learning. We often experience difficulty in training such 'inner' parameters when teacher signals are dichotomic. To overcome this difficulty, we construct one-hidden-layer neural networks with a smaller number of the inner parameters to be optimized, fixing some components of the parameters. This inevitably causes increment of the hidden-layer units, but the network learns the Bayesian discriminant function better than ordinary neural networks.

AB - Learning of Bayesian discriminant functions is a difficult task for ordinary one-hidden-layer neural networks, because the teacher signals are dichotomic random samples. When the neural network is trained, the parameters, the weights and thresholds, are usually all supposed to be optimized. However, those included in the activation functions of the hidden-layer units are optimized at the second step of the BP learning. We often experience difficulty in training such 'inner' parameters when teacher signals are dichotomic. To overcome this difficulty, we construct one-hidden-layer neural networks with a smaller number of the inner parameters to be optimized, fixing some components of the parameters. This inevitably causes increment of the hidden-layer units, but the network learns the Bayesian discriminant function better than ordinary neural networks.

KW - Bayesian

KW - Layered neural network

KW - Learning

KW - Quadratic form

UR - http://www.scopus.com/inward/record.url?scp=54249110427&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=54249110427&partnerID=8YFLogxK

U2 - 10.1007/978-3-540-69158-7_26

DO - 10.1007/978-3-540-69158-7_26

M3 - Conference contribution

AN - SCOPUS:54249110427

SN - 3540691545

SN - 9783540691549

T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

SP - 238

EP - 247

BT - Neural Information Processing - 14th International Conference, ICONIP 2007, Revised Selected Papers

Y2 - 13 November 2007 through 16 November 2007

ER -