Neural Network Transfer Function

Hi friends,

I was using neural network to train a set of sensing data. The transfer function is designed to be 'logsig'. The input is a n*4 matrix values between 0 to 3, output is a n*3 matrix values between 0 to 10.

The training process is normal. But after we have the resultant network, when applying, output values are no less than 5 which is the average of my target output. That is to say, for all values supposed to be lower than 5, it returns 5.

Code will be the following: net1=feedforwardnet([80,80]); net1.trainParam.max_fail=10; net1.layers{3}.transferFcn='logsig'; net1.layers{1}.transferFcn='logsig'; net1.layers{2}.transferFcn='logsig'; net1=train(net1,transpose(det),transpose(pos));

Thank you all in advance.

Yichang

Answer 1

[Greg Heath]

There is no good reason to differ from the default

 help fitnet   % Form of feedforwardnet for regression & curvefitting)

 doc fitnet

 given matrix sizes

 [ I N ] = size(input)  % [ 4 N ]
 [ O N ] = size(target) % [ 3 N ]

Default number of training equations

 Ntrneq = 0.7*N*O = 2.1*N

and number of unknown weights for a single hidden layer with H nodes yielding Nw unknown weights and biases

 Nw = (I+1)*H+(H+1)*O = (4+1+3)*H + 3 = 8*H+3

For more training equations than unknowns

 Ntrneq >= Nw

or

 H <= (Ntrneq -O) /( I + O +1)=(2.1*N-3)/8

Therefore, the default value of H = 10 can be used if

 N >= 40 .

Hope this helps.

Thank you for formally accepting my answer

Greg

Answer 2

[Akshay Joshi]

Try the following:

 net1.layers{1}.transferFcn='logsig';
 net1.layers{2}.transferFcn='logsig';
 net1.layers{3}.transferFcn='purelin'

In many cases, using non-linear function for classification at hidden layer, and then using linear function at output layer yields a proper, classified output data.