[raf]

How do I design in words ( step by step)  neural network that learns number [0-9]? Could someone explain?

[Bragi]

Learns what about numbers? Recognize written numbers on paper or something else?

[raf]

Learns numbers from a tab6x6 ( tab contains  0 and 1. 0 is an empty field,  1s make up a number). What would look the process of designin such a network (step by step including calculating neuron, layers etc)? Thanks in advanced.

[Bragi]

I'm a little confused by the 6*6 grid. Do you mean something like this, which would be the nr 3:

Code

1111
0001
1111
0001
1111

I'd build a decision tree, like so: http://bragisoft.com/2013/02/creating-and-using-decision-trees/

[raf]

Yeah, that what I mean, But the grid is 6X6 not 4x5. The problem is that I am really new in this and I am not familiar enough with decision tree to give it a try. So lets focus on the grid. To be honest that was an exam question to desing a NN that recognizes numbers.  If you could help me, how would you answer the question ( how would you desing the network)?

[Bragi]

Well, how far did you get?

[raf]

Well for me its obv that this network should have 36 inputs , 10 outputs ( numbers from 0 to 9)  and 2 hidden layers ( guessing). And it could use Sigmoid funcion for calculating the neurons output. Trained with back propagation rules. Thats all i can write;

How do i know kow many neurons should I have in every layer?
What can i write more?

[Bragi]

Ahh, that's the tricky bit. Trial and error, training functions,..

[raf]

could you answer in more detail?

[Ben.F.Rayfield]

You should be able to learn bit patterns using SimpleRBM which is where I started learning about Boltzmann Machines and then wrote in enough technical detail here http://aidreams.co.uk/forum/index.php?topic=5990.0 ("Ben F Rayfield - My bizarre research toward networking minds together") how you can build one from scratch, train, and use it in realtime.