svm - Moving from support vector machine to neural network (Back propagation) -

i'm working text recognition , i'm using support vector machine method. try neural network also. read few documents how neural network works, theory quite heavy , don't know how apply case. if can me make clear, neural network's architecture.

• currently, in svm, have 200 features (divided 4 main categories), used recognize text. if move neural network, 200 features, mean have 200 neutrons in input layer?
• with 200 features, how result in architecture of neural network (in term of numbers layers (hidden layer) , neutrons)?
• in svm, have 1 class classification (basically, true , false) , multi-class classification (labels), how difference apply output layer of neural networks?

and have few general questions :

• what decide number of hidden layers , number of neutrons inside each hidden layer?
• does number of hidden layer relate accuracy ?

i'm new neural network great if can explain me in understable way. :)
thank much.

point 1 - 200 input neurons, each neuron fed binary number, or float (preferably normalized in range -1 1).

point 2/4 - majority of problems solved single hidden layer. if starting out neural networks should stick 1 hidden layer. suggest starting less 200 input neurons, try 5 or 10. multiple hidden layers used in complex problems, example, first hidden layer learns macro features dog, cat, horse , next hidden layer learns finer features eyes, nose, ears etc.

there no definite procedure deciding number of hidden neurons. more complex problem, in theory more hidden neurons needs. if have 10 input neurons, start 20 hidden neurons. if doesn't work, wrong elsewhere. if work, can reduce number of hidden neurons until fails.
can start low , work up.

point 3 - true , false classification, use single output neuron, , train 0 or 1. n classes, use 1 of n encoding.

point 5 - no. accuracy measured how network can generalize - ie., how performs on data never seen before. in general, more training data = more accurate.