Numerical methods of interval analysis in learning neural network

The paper is devoted to the development and examination of the interval analysis-based numerical methods of the guaranteed learning of neural direct-propagation networks. Developed were contractive operators that allow for the singularities of the problem of learning (quadratic learning performance functional and superpositional weight-linear/nonlinear structure of the neural networks) and are used in the numerical methods of learning. The results of computer-aided experiments studying effectiveness of the developed methods were presented. The method of learning based on the algorithm of inverse error propagation and the method of weight shaking for determination of the global optimum were compared.