I’ve to create a neural network for classifying 20 objects. My input matrix is 100 by 400, and target is 1 by 400. Each 20 columns of input input matrix belong to same class, like this:
P(1:20) belongs to class A
P(21:40) belongs to class B
.
.
.
P(381:400) belongs to Class T
Target vector contains 10 to 200, twenty 10 for class A, twenty 20 for class B…twenty 200 for class T, like input matrix.
I put 15 sample of input class:
Now I’ve some questions:
1) Are these inputs good enough for classification?
2) 20 samples for each class is enough?
3) Is Feed-forward back-propagation network type is suitable for this network?
4) How many hidden layers should I use? and how many neurons for each layer (approximately)
I put .mat file for input matrix and target vector in skydrive website: input.mat, target.mat
I’ll be appreciated for your help.
The answer to all you questions is “try and see what happens”. With neural networks, it is very hard to say how well they will work until you’ve run an experiment. My intuition is 15 training examples per class are not enough for 100-dimensional problem, but I suppose you cannot get any more and have to work with what you’ve got. Look into dimensionality reduction.