I get a "Performance function replaced with squared error performance" warning when trying to set 'crossentropy' as the performance function.

If I run the following code:

[x,t] = house_dataset;
net = fitnet(10);
net.performFcn = 'crossentropy';
[net,tr] = train(net,x,t);

I get this warning:

Warning: Performance function replaced with squared error performance. 
> In trainlm>formatNet (line 155)
  In trainlm (line 65)
  In nntraining.setup (line 14)
  In network/train (line 335)

How can I use 'crossentropy' as the performance function then?

Regards

[Greg Heath]

1. Where did you find house_dataset?

It does not come with MATLAB17a

2. For classification

 help patternnet

 doc patternnet

Hope this helps

Greg

Answer 1

[Greg Heath]

Crossentropy is, theoretically, not appropriate for regression.

Classically, it is only used for classification and pattern-recognition. It's definition involves probability distribution functions and their logarithms.

That is not to say that it will not yield good answers for regression problems. Obviously I have never tried it and, one day when I get bored, I might tinker around with it.

Hope this helps.

If you think this answer is worth accepting, THANKS!

Greg

Answer 2

[Greg Heath]

If you insist on using CROSSENTROPY, try PATTERNNET.

Hope this helps.

Thank you for formally accepting my answer

Greg

Answer 3

[Alex]

It worked for me after adding

net.performParam.regularization = 0.1;
net.performParam.normalization = 'none';

Seems to be necessary when using cross entropy.

Answer 4

[Chinmay Maheshwari]

Any updates on it as i am also facing the same trouble. I am not to customize my performance function because of this Thanks in advance