ann_FF_INT — internal implementation of feedforward nets.
This man page describes the internals of implementation, it is of interest only to those wanting to modify/adapt the functions provided (in case they do not fit the need :-)
The variables described here are used internally only in the functions body. They are: err_dz the partial derivative of error with respect to neuronal outputs, on current layer. err_dz_deriv_af the element-wise product between err_dz and deriv_af. deriv_af the derivative of activation function for current layer. grad_E gradient of error, same hypermatrix structure as W. E.g. grad_E(n,i,l) destined to hold the partial derivative of error with respect to W(n,i,l). grad_E_mod similar to grad_E but it is not the real grad_E because is modified in some significant way. l, ll current layer number (different from l(1), l(2)) (if l=[l(1),l(2)] is used as parameter then ll is used as layer counter) L total number of layers, including input. p current pattern number. P total number of patterns. y holds the network output(s), one per column. z matrix of neuronal outputs, one column for each layer, e.g. z(1:N(1),1) contains the inputs z(1:N(2),2) contains the outputs of first hidden layer z(1:N(L),L) contains the output BIAS Bias is simulated trough an additional neuron with constant output "1". The "bias neuron" is present on all layers, except output, as the first neuron.