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
T2 - 14th International Conference on Neural Information Processing, ICONIP 2007
Y2 - 13 November 2007 through 16 November 2007
ER -