Dorogov A. Yu., Shestopalov M. Yu.

The methods of analytical and experimental estimation of separating capacity for weakly-connected feedforward neural networks is offered. It is shown that morphologic invariant of weakly-connected networks provides construction of analytical estimations. Direct and dual models of the neural networks are investigated. Freedom degrees of the neural networks are estimated through freedom degrees of neural module. The analytical dependence of amount the recognized images in general position from freedom degrees is obtained. The experimental results are represented.