Classification for HC.

There are 1,700 objects, including 130 objects of class A, 1570 objects of class B. each object of this 130 characteristics, by screening for multicollinearity (Tau Kendall more than 0,7) and using genetic algorithms to probabilistic networks (statistica 6.1) was selected 50 significant characteristics. Continue in the same package want to run a mlp to classify these objects, but I can only submit 260 (130 per class), because otherwise NS a priori to include all objects of class B, but I read that the number of parameters (weights?) in the na should be 10 times smaller than the sample. Obviously, if you follow this rule, then the hidden layer will be a pair of neurons, and this, in theory, is not enough. Need to increase these 130 pieces of class A. Thoughts go in the direction of reproduction by adding random noise for each characteristic, but this is not accurate. And yet, perhaps we should take away from the characteristics of only those with a normal distribution and then noise to add or on the basis of empirical characteristics to finish somehow.

Programming languages don't know it, please tell me a software product with the realized increase in the sample size or other ways to solve this problem, is also preferably implemented programmatically :)