I have a mse mistake in Neural Network

Hello I have a question I need to know how to avoid local minima in Neural Networks, or I don't know why after I train the model and test it once, I have a mse 0.015, and once again and test I got 0.012, each time I train once I got other mse value. Thank you so much for your help

Answer 1

[Greg Heath]

You are confused. The task is to find a local minimum that is sufficiently low. I use the goal

 mse(target-output) <= 0.01*mean(var(target',1)).

This yields an Rsquare >= 0.99 % Look up Rsquare in wikipedia

Furthermore, because of the output's invariance to the shuffling of hidden node order and weight signs there are 2^H * H! equivalent nets with the same number of hidden nodes and weight absolute values. For the default value H = 10, 2^H *H! = 3.7159e+09

Furthermore there are often more than one set of orders and absolute values that will cause an acceptable Rsquare.

The differences you see can be caused by different initial random weight values.

Hope this helps.

Thank you for formally accepting my answer

Greg

[Andres Romero]

thank you so much

[Andres Romero]

Thanks for help me with my question, I got other one and maybe you can help me, I need a good book to show me how to choose a neural network and the steps to improve it, I mean to know when my net is better. Thank you so much Greg

[Greg Heath]

I haven't read any recent NN books.

Try taking a look at some of my tutorial type posts in both NEWSGROUP and ANSWERS

 greg tutorial neural