Despite what might seem a fairly workaday brief compared with that of its elder brother AT&T Bell Laboratories, Bell Communications Research Inc, the research consortium owned by the seven regional Bell Operating Companies spun off from AT&T Co – plus two former AT&T affiliates, Cincinnati Bell Inc and Southern New England Telecommunications Inc – and mandated to apply itself to local-telephone-related applications, regularly comes up with some pretty exotic technology. The latest experimental baby to be born at the Livingston, New Jersey laboratory is a neural network computer that is designed to learn like people do, by example. The computer can learn and process patterns in more than 100,000 individual signals a second, about 10,000 times faster than a leading edge workstation. Bellcore’s computer uses an experimental chip the consortium first demonstrated in 1988. The lab says that the experimental system could herald a new generation of neural learning computers that will be able to solve some classes of complex problems faster and more efficiently than conventional machines, but without requiring tedious programming – identify spoken words, read handwriting, identify fingerprints and recognise a smell. The new version of the chip has 32 neurons and 496 synapses and is cascadable. Why should Bellcore be playing with neural computers? It plans to add additional chips to the prototype to create an experimental neural learning computer aimed at solving problems commonly faced in telecommunications network management and operations systems, including least-cost and least-congested routing of phone calls, assigning frequencies to radio equipment, compressing phone company business data for storage and transmission and recognising speech.
