Tyumen Researcher Improves Reliability of Memristor-Based Neural Networks
Alexander Neustroev, a researcher at Tyumen State University, is developing ways to improve the reliability of hardware neural networks built on memristors.
He has created a mathematical model of non-ideal memristor behavior and a methodology for modeling a neural network layer.
In hardware neural networks, computations are performed by specialized electronic components. A memristor can change and retain resistance, so systems similar to biological neural networks are built on its basis. Real components may behave differently from what calculations assume, and the neural network does not work as intended. Neustroev is looking for ways to predict such deviations in advance and reduce them.
He studied undesirable effects in circuits with memristors and Zener diodes and proposed a way to reduce their impact. The methodology for modeling a neural network layer will help test circuits on a computer before testing them on real devices. The researcher is currently working on a program for modeling a hardware neural network layer and comparing calculations with experiments. Such research is needed for specialized processors that perform neural network computations in electronic circuits. In the future, the technologies are applicable in robotics and control systems.
