The emerging technology of neural networks, which attempts to imitate in silicon and software the way that living brains work, is already showing signs of practical application. Microbytes reports that one of the pioneers of the new science, Hecht-Nielsen Neurocomputer Inc of San Diego, California, has come up with a co-processor and software for MS-DOS micros that is designed to simplify the task of teaching a computer to recognise patterns and shapes. Pattern recognition is a vital function of a vision robot, which for simple tasks like placing components onto a circuit board must be able to recognise the right component for each position whatever way round it happens to be in the bin, and to be able to see precisely the position and orientation in which it must be placed. But Robert Hecht-Nielsen, president of the eponymous company, reckons that his AR/NET software and Anza neurocomputing co-processor for MS-DOS micros, can effectively apply the technology of adaptive resonance neural networks to enable unsupervised pattern learning by machine. Aimed at tasks such as image recognition for security purposes, the AR/NET software is claimed to be the first fully functional product of this type. AR/NET, in conjunction with the neurocomputing board, uses a massively parallel structure to recognise a pattern it has learned immediately. Regardless of the number of patterns the net has learned, it can respond to any one of them, says Hecht-Nielsen. It doesn’t slow down as it learns more patterns, either, he adds. The AR/NET software can be set to signal the user when it encounters a pattern it does not recognise. Users can change the amount of error the program will tolerate in matching a pattern to a learned category. No idea of price was given for the product.