The neural network stops before it starts? (minimum gradient reached ) error
6 Ansichten (letzte 30 Tage)
Ältere Kommentare anzeigen
Abdullah
am 30 Mai 2014
Bearbeitet: Greg Heath
am 30 Sep. 2015
Hello every body. I am using Neural network in my research graduation project " Pattern Identification ". I have built a custom neural network, and I used the function 'trainlm' .
The Problem is that, once I start the training process, the training window appears and tells me that the minimum gradient is reached and the neural network stops even before it starts !
( The architecture of my NN is complicated, and I'm using three inputs with 13 layers each of them carries different neurons ! )
I wish I can find someone helping me
Thanks in advance !
2 Kommentare
EanX
am 30 Sep. 2015
I have a similar problem in designing a NARNET.
Currently best performance was obtained with 15 hidden nodes (1 hidden layer), mapminmax for processing function for inputs (ymin=0, ymax=10), tansig as transfer function for layer 1 and poslin for output layer.
I noticed that if I change output transfer function to the more commonly employed purelin I can avoid "minimum gradient reached" issue, but I used poslin to obtain forecasts always positive.
Can anyone enlighten me on this topic?
Thanks in advance !
Greg Heath
am 30 Sep. 2015
Bearbeitet: Greg Heath
am 30 Sep. 2015
Do not change defaults until you have obtained the best default result.
Are you looping over 10 or more weight initializations and getting premature MinGrad reached on all of them?
Akzeptierte Antwort
Greg Heath
am 30 Mai 2014
The best way to design a neural net is to first find the simplest architecture that yields a good solution.
My advice is to start with a single hidden layer and find the smallest number of hidden neurons that will yield a solution that cannot be significantly improved by using more neurons.
The best targets for patternnet are unit column vectors with a single unity component. The best training algorithm for that type of target is trainscg.
Hope this helps.
Greg
Weitere Antworten (0)
Siehe auch
Kategorien
Mehr zu Define Shallow Neural Network Architectures finden Sie in Help Center und File Exchange
Produkte
Community Treasure Hunt
Find the treasures in MATLAB Central and discover how the community can help you!
Start Hunting!