Abstracts

Programmirung

The purpose of development of a universal hardware-software medium for a realization neural networks is pursued. As basic the principles are used:
Self-management by the process of evaluations;
Parallelism of evaluations;
Programmability of network topology;
Combination of an asynchronous character of data transfer between neurons with a synchronization of level-by-level work of a network;
Solidity of a hardware-software medium;
Multifunctional use of memory.

The first principle is control by data: the operation is fulfilled at readiness of input data. The principle allows to depart from sequential command control and potentially supports a realization of maximum parallelism peculiar to the neural tasks. For the second principle the presence of rather large number of the simple processor elements for realization of functional transformations connected to toting of weighed values of entering variables and a comparison of an obtained sum with a value of a threshold is necessary. The programmability of neural network topology can be supplied with the explicit instruction ?radiants, receivers? of data in a format of command packages, that connects the neurons of a network, as in direct direction (operating duty of a network ), and in the opposite direction (condition of tutoring). The self-management eliminates predetermining about handling command packages by the processor elements and moment of data transfer between neurons of a network , expedient fixing of transient states of a network especially in learning process is represented. The solidity of devices embodies natural rushing to security of a reliability and constructive wholeness of hardware. The multifunctional use of memory ensures data processing on a place of storage and minimizes information interchange with an exterior medium. The hardware-software medium is represented as a regular structure on basis of multifunctional memory and simplified up to a level of a neuron of the processor elements. The memory is necessary for storage of network topology, functional parameters of a network, values of information signals spreaded on a network during work, and also codes of operations of the processor elements. Transmission of input data in the processor elements is addressless, and outcomes of handling on inputs of neurons is address.

Note. Abstracts are published in author's edition

Comments

[Home] [Conference]

©2001, Siberian Branch of Russian Academy of Science, Novosibirsk
©2001, United Institute of Computer Science SB RAS, Novosibirsk
©2001, Institute of Computational Techologies SB RAS, Novosibirsk
©2001, A.P. Ershov Institute of Informatics Systems SB RAS, Novosibirsk
©2001, Institute of Mathematics SB RAS, Novosibirsk
©2001, Institute of Cytology and Genetics SB RAS, Novosibirsk
©2001, Institute of Computational Mathematics and Mathematical Geophysics SB RAS, Novosibirsk
©2001, Novosibirsk State University
Last modified 06-Jul-2012 (11:45:21)