Sounds like you want to train a pattern recognition neural network. To train a neural network you need to provide a set of training vectors and a set of target vectors.
Explaining what these are requires knowing the format of the NN input/output. The default settings for a pattern recognition NN takes in an input vector (ex: vector describes an image) and outputs an n-dimensional target vector with each element falling in the range (0,1). Each index of this output vector is associated with a class, and the index with the highest value is the network's decision for which class the input falls into. For example, if A2 matches B13, then ideally your network produces a vector t of length 20 where t(13) near 1 and all other t(i) near 0 for i ~= 13.
To train the NN you need to provide a matrix of example input vectors and a matrix of their corresponding ideal target vectors by concatenating several cases. Ideal target vectors are binary with the index of the correct class being 1 and all others being 0. You'll also need to decide the number of nodes in your hidden layer. A higher number of hidden layer nodes increases the network's ability to recognize classes at the cost of longer training/computation time and risk of being unable to generalize to inputs outside of the training set. I don't know of a definite way to find the ideal number, so I tend to go with trial and error on this.
During training Matlab will randomly separate your training vectors into training, testing, and validation sets. The training set is used to train the NN. The validation set is used to check if the NN is too well fitted to the training set, which is done by stopping training after performance on the validation set gets worse 6 consecutive times in a row. Performance on the test set is the best indication of your network's performance.
After a network is trained, you can give use it by giving it other inputs as if it were a function, like:
for k = 1:N
t = net(pictures(k));
[~,I] = max(t);
fprintf(file,'A%iB%I', k, I);
end
