O(I) current value of node I in the network, does not include the input unitsA(I) current value of the summed input
to node I
D(I) current value of the
at node I
T(I) current value of the threshold
(see eq. (
)) at node I
DT(I) current value of the update
for the threshold at node I
W(IW) current value for weight with index IW
DW(IW) current value of the update for weight with index IW
NSELF(IW) switches for updating weight with index IW
0
do not update
1
update
for self-organizing map, NSELF is equivalent to the vector NBHD
NTSELF(I) switches for updating threshold for node I.
0
do not update
1
update
G(IW) temporary weight and threshold vector used in CG, SCG and QP
ODW(IW) stores old weight gradient in CG, SCG, QP and Rprop
ODT(I) stores old threshold gradient in CG, SCG, QP and Rprop
ETAV(IW) individual learning rates used in Rprop
M(I) number of nodes in layer I ( I=0input layer)
MV0(I) offset index for node vectors - tells the index in the node vectors for the
last node in layer ( I-1)
MM0(I) offset index for weight vectors - tells the index in the weight vectors for
the last weight going from layer ( I-1)
NG(I) activation function
for layer I
NL number of layers except input layer
IPOTT switch for use of Potts units
IPOTT = dimension of Potts units
ER1, ER2 internal variables used for calculating PARJN(7-9)
SM(I) internal variable used for calculating saturation measures
ICPON switch for precision chopping
1
on
NXIN x-width of input field (if negativeThe geometry of the receptive fields is defined as follows: The input nodes are assumed to be organized in a plane of NXIN*NYIN nodes (with periodic boundaries if NXIN or NYIN are negative). Note that the coordinates (IX,IY) corresponds to input node number IN=(IX-1)*NYIN+IY. Each receptive field node in the first hidden layer scan an area of NXRF*NYRF input nodes.periodic boundary)
set from MSTJN(23)
NYIN y-height of input-field (if negative
periodic boundary)
set from MSTJN(24)
NXRF x-width of the receptive field -- set from MSTJN(25)
NYRF y-height of receptive field -- set from MSTJN(26)
NXHRF number of overlapping receptive field in x-direction
NYHRF number of overlapping receptive fields in y-direction
NHRF total number of receptive fields
NRFW number of weights per receptive field
NHPRF number of nodes per receptive field -- set from MSTJN(27)
If NHPRF is negative, weights from equivalent receptive field nodes in the
first hidden layer to each node in the second hidden layer are shared.
ILINON switch that tells the current status of the line search.0
off
1
on and trying to bracket the minimum
-1
on and looking for the minimum (knowing that its been bracketed)
NC number of steps taken in CG minimization
G2 squared magnitude of the vectors DW(IW) and DT(I)
NIT number of calls to the line search (along current direction)
ERRLN(I) stored errors used in line search
I=0
current error
I>0
previous errors
DERRLN gradient along search direction at initial point
STEPLN(I) coordinates, relative to current position, for previous points corresponding to ERRLN(I)
STEPMN position, relative to current point, of best minimum so far
ERRMN minimum error so far
IEVAL switch for the comparison parameter in SCG training
0
do not compute the comparison parameter
1
compute the comparison parameter
ISUCC switch that tells if a successful step was made in the SCG algorithm
successful step
ICURVE switch to tell whether curvature information exists or not -- used in the SCG algorithm.
0
curvature information does not exist
1
curvature information exists
NSC number of attempted steps in the SCG algorithm
GVEC2 squared magnitude of the vector G(IW)
D2E(IW,IW) the Hessian matrix
. They are computed when the function GJN is
called.
GPJN(I) first derivative
GPPJN(I) second derivative
