Why do I get a dimension mismatch error when training my neural network using constant input data in Neural Network Toolbox 6.0 (R2008a)?

There are 2 datasets in the attached MAT files : data1.mat , data2.mat. Creating and training the network using the first dataset works fine, however using the second dataset (containing constant inputs) leads to an error:

 ??? Error using ==> plus

 Matrix dimensions must agree.

 Error in ==> calcperf2 at 163

         N{i,ts} = N{i,ts} + Z{k};

 Error in ==> trainlm at 253

 [perf,El,trainV.Y,Ac,N,Zb,Zi,Zl] = calcperf2(net,X,trainV.Pd,trainV.Tl,trainV.Ai,Q,TS);

 Error in ==> network.train at 219

   [net,tr] = feval(net.trainFcn,net,tr,trainV,valV,testV);

 Error in ==> netztraining at 63

     Netz = train(Netz,[p;t],t,pi);

To reproduce this behavior, run the attached sample script 'nettraining.m' after loading one of the MAT files.

Answer 1

[MathWorks Support Team]

This problem occurs because of some automated input

preprocessing that is done by Neural Network Toolbox 6.0 (R2008a). The current default preprocessing functions are:

 fixunknowns, removeconstantrows, and mapminmax

In some of the data (attached example data1.mat) there are elements of the

input vector that are constant over time. For network training,

these constant rows do not provide any useful information, since they

can be replaced with an adjustment of the bias in the final layer of

the network. For this reason, the toolbox removes any constant rows,

if the default settings are used.

In the second data set (data2.mat), all of the rows of

the input vector were constant, so they were all removed. This

produced an input with dimension zero. Even in the first data set,

there were some constant elements of the input vector that were removed.

To override the default settings, the following three lines can be

added to the attached example script:

Netz.inputs{1}.processFcns = {'mapminmax'};

Netz.inputs{2}.processFcns = {'mapminmax'};

Netz.outputs{2}.processFcns = {'mapminmax'};

These new lines should follow the existing line

Netz = newnarxsp(P1,T1,DI,DO,15);

In this way, the network will still normalize the inputs and targets,

which is good practice for network training, but it will not remove

the constant rows. If you do not want to normalize the

data, then you can remove " 'mapminmax' " from the above lines. (If

a row is constant, and mapminmax is used, the constant value will be

replaced by -1.)

As a side note – generally you may want to use a data set where all elements of

the input vector change. If any of the elements of the input vector

are constant throughout the data set, then the bias in the last layer

can compensate for the effect of those input elements. This means

that the weights that multiply the constant inputs will be effectively random.