SYNTHESIS OF RECURSIVE DEVICES FOR VERTICAL-GROUP CALCULATION OF BASIC MULTI-OPERAND NEUROOPERATIONS

Abstract

The operational basis of artificial neural networks has been determined, comprising groups of the following neurooperations: preprocessing, processing, and computation of transfer functions. A set of basic multi-operand neurooperations was selected for hardware implementation, including: finding maximum and minimum values in a one-dimensional data array, calculation of the sum of squared differences, scalar product calculation. The methods of vertical-group computation of basic multi-operand neurooperations (finding for maximum and minimum values in a one-dimensional array, calculation of the sum of squared differences, and scalar product calculation) have been improved. Using the selection of number of bits for operands group for single-cycle processing, these methods enable synchronization of data arrival time with calculation time and ensure high hardware utilization efficiency during the hardware implementation. It's proposed a recursive devices design for vertical-group computation of basic multi-operand neurooperations based on an integrated approach. This approach leverages the capabilities of modern element base, incorporates vertical methods, algorithms, and recursive device structures for implementing basic neurooperations and considers the requirements of specific applications. The principles for designing recursive devices for vertical-group calculation of basic multi-operand neurooperations have been chosen. These include: the use of a basis of elementary arithmetic operations and a multi-operand approach; modularity; pipelining and spatial parallelism; homogeneity and regularity of the structure; synchronization between data arrival time and neurooperation calculation time; specialization and adaptation of structure to specific application requirements. A format converter has been developed to transform a flow of serial input data from a one-dimensional array into a parallel-serial data output by group of bits. Basic structures have been developed. They represent calculation algorithms in terms of hardware and serve as the foundation for synthesizing of recursive devices for vertical-group calculation of basic multi-operand neurooperations with specified parameters. The method for synthesis of recursive devices for calculation of basic multi-operand neurooperations with vertical-group data processing has been improved. Through the use of mechanisms for synchronizing calculation time with data arrival time, this method provides the selection of structure which performs real-time data processing and with high hardware utilization efficiency. It has been demonstrated that the use of the improved vertical-group methods, designed basic structures of devices for finding maximum and minimum numbers in one-dimensional arrays, calculation of the sum of squared differences and scalar product, as well as the improved synthesis method, enables real-time mode and the implementation of devices for calculation of basic multi-operand neurooperations with vertical-group data processing with high hardware utilization efficiency.