Department of Modular Computing and Artificial Intelligence (SCCMI)

New mathematical models for training neural network structures are being developed, including deep, quantum, pulse and complex-valued neural networks. To improve the known approaches, promising mathematical concepts of information geometry, optimization theory, fractional differential calculus and the theory of finite algebraic structures are used. The approbation of the developed artificial intelligence systems is carried out on the actual tasks of agriculture and medicine.

The properties of modular computing based on a residual class system for the development of high-performance digital signal and image processing devices are investigated. The key idea is the ability to quickly and parallel organize arithmetic operations of addition, subtraction and multiplication in a system of residual classes. The problematic operations of this system include non-modular operations of dividing numbers, restoring the binary form of a number, determining the sign, comparing and some others. The residual class system is a promising tool for improving the performance of digital filters and artificial neural networks, since in these applications the main share of calculations is accounted for by addition and multiplication. A comprehensive study of promising digital device architectures includes mathematical and software modeling, as well as hardware modeling of microelectronic devices on FPGA and ASIC.