
Hybrid Networks
235
tion. Most canned routines for training feedforward networks assume
that the same input values are applied to all hidden neurons. This is
no longer the case. But the exact same formulas are still appropriate.
The subroutine that computes the correction terms must be modified
to use the inputs to each individual hidden neuron when updating the
weights for that neuron. Also, if the n's are different, every routine
that references the number of inputs must be modified to use the
correct n.
Fast Bayesian Confidences
In order to compute Bayesian confidence levels for classification
decisions, one must be able to compute estimate ...