Construction of efficient neural networks: Algorithms and tests

In this paper the problem considered is: how to obtain a maximum of skills with minimum number of connections between neurons. Under consideration are the learnable neural nets. Training was done by minimizing the estimation function using the single-step quasinewtonian method (BFGS-formula). At the beginning of training the net features a maximum number of connections. In the course of training the connections are eliminated with minimum effect on the estimation of the net operation. Several computational experiments are described.