Assessment of stability learning algorithms large artificial neural networks of biometric application

The problem of increasing the speed, stability and efficiency of known and emerging algorithms for training large artificial neural networks is considered. To solve this problem it is proposed to create the most balanced multi-criteria performance comparison of different learning algorithms, enter a numerical estimate of the learning parameters and systematically carry out work to improve them. It is argued that resistance training of neural networks is an analogue of the assessment procedure of the condition number of matrices of linear algebra. We propose a measure of stability training, given the connection between this indicator with the computational complexity of learning algorithms, as well as with a number of examples on which the algorithm is efficient. We discuss the effect of increasing the stability by supplementing the training sample synthetic image examples of “your”. The optimum amount of additional synthetic examples, the addition of which can reduce the number of examples in the training set is provided. A block diagram of a promising means of training neural network converters, which introduced the third machine reproduction examples image of “Its” is given. To control the number of synthetic examples, it is proposed to create a table of restrictions for the first machines based on expert opinion.