MATLAB: What does “Performance” mean in NN Toolbox’s training record window

matlab guineural networkperformancetoolbox

I'm testing a custom neural network and after training, this window appears telling me that the performance is 0.26. What does that mean? Is an high value a good or a bad result ? How is it calculated?

Thank you in advance.

Best Answer

It depends on the training function. The defaults for fitnet and patternnet are mean-square-error and crossentropy, respectively.

In general, if you omit the trailing semicolon, you can see which training function and performance measure is being used

 net = fitnet % no semicolon

For details, use the help and doc commands, e.g.,

 help crossentropy
 doc crossenrtopy

Hope this helps.

Thank you for formally accepting my answer

Greg

Related Solutions

MATLAB: NN Toolbox: What is the difference between “performance” and “plotperform”

I understand that you are a little confused as to how "perform" and "plotperform" differ.

From what I can infer, there is not much difference between them, both "plotperform" and "perform" work on "Mean Sqaured Error" by default. But "plotperform" actually plots these mean squared errors for their corresponding epochs where as the "perform" function calculates the performance of the network based on two property values:

performFcn : which is "mse"(mean squared error) by default but can be set to any of these as well:

mae - Mean absolute error performance function.
mse - Mean squared error performance function.
sae - Sum absolute error performance function.
sse - Sum squared error performance function.
crossentropy - Cross-entropy performance.
msesparse - Mean squared error performance function with L2 weight and sparsity regularizers.

performParam: This is set to "regularization" and "normalization" by default.

The value returned by the "perform" function is a double value and the closer this value gets to zero, the better the network.

Also, check out these links for more information:

MATLAB: How to choose the number of neurons in the hidden layer of neural network

Use trial and error with the training subset goal

 MSEtrngoal = 0.01*var(trntarget,1)        % 1-D target
 or
 MSEtrngoal = 0.01*mean(var(trntarget',1)) % Otherwise

Start with the MATLAB default H = 10 and design ~10 nets for each setting of H. Each of the 10 is initialized with a different setting of random initial weights.

Hope this helps