> Date: Mon, 31 May 1999 18:51:18 +0000 > From: Boanne Lorraine <[EMAIL PROTECTED]> > }- The PC-MSDOS-CP/M approach is definitely a step in the right > }- direction, but we won't see the problems resolved until the > }- true ANALOG neural-net computer ( *NOT* a digital simulation > }- of one ) becomes reality. > I have always heard that "analog computers are impossible" and > that they would be HUGE... Could you explain what the difference > between an analog and a digital computer is, and why they would > be impossible? Certainly not impossible; one form of analog computer ( which Mike Millen has explained most capably here ) has been around a long time. In use for nearly a century in Naval ships for computing the ballistics of the big guns, even earlier examples exist. The French mathemetician, Blaise Pascal ( namesake for the Swiss-designed Pascal programming language ) built a "numerical engine" out of gears and clockworks about 300 years ago; it was a true analog computer, too. However, the analog nueral net computer is a different animal altogether. These model the brain activity of living organisms, and unlike your standard digital computer, LEARNS what to do, rather than being programmed to do it. The concept is deceptively simple - and always includes an element of FEEDBACK which reinforces the correct answer, and rejects the wrong one. Like people, it can make valid judgements about situations it has never "seen" before; the level of confidence about the answer is dependent upon its previous learning experiences. ( Just like children/adults experiencing new situations that are similar to, but not exactly the same as others which they remember.) There are plenty of "digital simulations" of nueral-net activity available; they "learn", too - but not with anywhere near the efficiency of an analog system. In fact, advanced digital simulations have been the key to discovering the capabilities that are possible in such an analog computer. ( The big holdup is development of op amps inside integrated circuits with the economy of scale - and huge numbers - we see today in RAM chips...) About ten years ago I attended an AI ( "artificial intelligence" ) seminar where an IBM researcher ( feverishly, but quietly developing these things for the military ) detailed an early neural net experiment done for the Navy. It seems it takes about seven years experience before a navy sonarman can truly be considered an "expert", able to tell the difference between benign and dangerous "targets" with a great degree of confidence. ( Imagine a whale and a very quiet submarine...) An IBM experimental nueral-net system was "taught" the techniques by ob- serving an expert sonarman's responses for EIGHT HOURS. It then went "online", identifying sonar targets by what it had learned. It had learned to correctly identify these things with about 85% accuracy - in 8 hours. It took human subjects approximately three years of training to accomplish a similar level of accuracy !! - John T. -- Arachne V1.5a;alpha, NON-COMMERCIAL copy, http://home.arachne.cz/ To unsubscribe from SURVPC send a message to [EMAIL PROTECTED] with unsubscribe SURVPC in the body of the message. Also, trim this footer from any quoted replies.
