Page 1 of 1

### Backpropagation / who can help me?

Posted: Fri May 25, 2018 5:13 am
Hello!

I tried to create a neuronal net to recognize alphabetical and decimal characters! I hope that someone can give me some advice..

After Training my Network what to do with the Output??

What do I get with .compute() function?

I´ve my Project added to this post, if someone can help me please!?

### Re: Backpropagation / who can help me?

Posted: Sun May 27, 2018 8:44 am
OK,
Maybe a few Questions for my understanding!

I´ve bitmaps with measurements like 100x100px equals 10000 input neurons!
Each bitmap represents a digit between 0 up to 9!

So only config I get Network to learn is to set it up like this:

Dim Network As ActivationNetwork = New ActivationNetwork(New BipolarSigmoidFunction(4), 10000, 1000, 100, 100, 10, 2, 1)
Dim _teacher As BackPropagationLearning = New BackPropagationLearning(Network)

or:

Dim Network As ActivationNetwork = New ActivationNetwork(New SigmoidFunction(2), 10000, 1000, 100, 100, 10, 2, 1)
Dim _teacher As BackPropagationLearning = New BackPropagationLearning(Network)

So I need 6 Layers, with fewer layers network error rate won´t change!

I´ve checked my Input data..they´re correct..
..so, If I get my network to learn and I´ll save the network how can I recall or recognize any digit with it?

Now I´m doing it this way:

Dim _sol() As Double = Network.Compute(_double)

but what to do with the Output-double?? On every digit I get same double as output!
I don´t think that Input neurons are overdimensioned..

That I´m setting output layer to 1 is correct, because I want it to learn that specific digit, right?
Sould I config the output Network to output layer to 10 and double with highest number is chosen one?!?

### Re: Backpropagation / who can help me?

Posted: Mon May 28, 2018 1:44 pm
Hello,

There is a lot written about digit/character recognition on the web. You may need to start reading some basics.

Here is something I did long time ago - should give you some hints. But you can find more by googling.