Determining Operation Tolerances of Memristor-Based Artificial Neural Networks

This article offers a general approach to developing methods of determining operation tolerances for the parameters' values of memristor-based artificial neural networks (ANNM), as a system that constitutes an united physical and informational object implemented by the hardware and software learning facilities. While looking for a solution to the issues of analysis and synthesis of this system's tolerances, the authors conducted its functional and structural decomposition with the introduction of several levels of hierarchy of the system, subsystems, functional links, and circuit components. The authors have researched the developed synthesis algorithm for the operation tolerances through the example of a two-layer feedforward neural network taught to detect the squitter of an info-communication signal when affected by noise, and implemented in MATLAB. The main parameters of neurons varied in the course of the research.