




The unit in Bp contains a bias weight and the various derivative terms
that are accumulated during the backpropagation phase:
BpCon bias
-
Contains the bias weight and its associated derivative and weight
change values. The bias weight is controlled by the
bias_spec
on
the BpUnitSpec.
float err
-
Contains the actual error or cost associated with the unit. It is
a function of the difference between the activation and the target
value, so it only shows up for those units that have target activation
values. It is not to be confused with the derivative of the activation
with respect to the error, which is
dEdA
.
float dEdA
-
The derivative of the error with respect to the activation of
the unit. For output units using the squared-error function, it is
simply
(targ - act)
. For hidden units, it is the accumulation of
the backpropagated dEdNet
values times the intervening weights
from the units to which the unit sends activation.
float dEdNet
-
The derivative of the error with respect to the net input of the
unit. It is simply the
dEdA
times the derivative of the
activation function, which is act * (1 - act)
for standard
sigmoidal units.
The unit specifications for Bp control what kind of error function is
being used, the parameters of the activation function, and the functions
on the spec orchestrate the computation of the activation and error
backpropagation phases:
SigmoidSpec sig
-
These are the parameters of the sigmoidal activation function. The
actual range of this activations are determined by the
act_range
parameters, and the sig
parameters determine the gain and any
fixed offset of the function (the offset is like a fixed bias term).
float err_tol
-
The error tolerance allows activation values that are sufficiently close
to the target activation to be treated as though they were equal to the
target value. Reasonable values of this parameter are from .02 to .1,
and its use prevents the large accumulation of weight values that
happens when the unit keeps trying to get closer and closer to an
activation of 1 (for example), which is impossible.
BpConSpec_SPtr bias_spec
-
A pointer to a BpConSpec that controls the updating of
the unit's bias weight. Typically, this points to the same
BpConSpec that updates the rest of the weights in the network,
but it is possible to have special BpConSpec's that do different
things to the bias weights, like initialize them to moderate negative
values, for example.
err_fun
-
A pointer to the error function to use in computing error for
output units that have target values. The function computes both
err
and dEdA
values (the former typically being the square
of the latter). While the user can define additional error functions,
the two that come with the standard distribution are
Bp_Squared_Error
and Bp_CrossEnt_Error
, which compute the
squared error and cross-entropy error functions, respectively.




