Analog Computing
Dateline: May 31, 1998
Digital computing today lends itself to a wide variety of applications. Analog computing does not, but it does lend itself to some very specific applications. Early digital computers took hours to compute a mathematical integration or differential, for example, whereas a simple analog circuit could do the job almost instantly, and in the 1940s and 50s analog computers were the instruments of choice for flight simulators and weapon guidance systems—both of which applications make heavy use of integrals and differentials. (Engineers may be interested in a paper by James Harman at the University of Minnesota.)
But analog computers lack precision, and you need precision if you want an ICBM to travel thousands of miles and drop its warhead on a dime. The ideal interim computing platform, then, for such applications was a hybrid combination of analog and digital computers. Eventually, however, digital computers became fast enough to equal and often outperform their analog cousins in most applications, and they had the added flexibility of being able to store their programs and their data.
Today, only a handful of companies produce analog circuits on a large scale, mostly for battery voltage regulators in laptops and cell phones, and for analog circuits in communications equipment. But analog circuits have found a recent niche in the AI world as neural net chips for pattern recognition. According to Prof. P. Masa at the University of Twente, who has been involved in the design of analog VLSI chips, they may be cheaper and faster than digital chips for such problems as real-time pattern classification in high-energy physics. Dr. Masa wrote that:
Analog VLSI offers compact, high speed but moderate precision analog computing. With a standard CMOS process up to tens of thousands of synapses can be integrated on a single chip resulting [in] up to ten-thousandfold parallelism in the computing. This parallelism can not be achieved with digital technique, because the inner product computing requires too large chip area.
But is the "moderate" precision of analog computing a serious problem? Dr. Masa says not:
Artificial neural systems—such as their biological counterparts—may exhibit significant robustness against the limited accuracy of components. They achieve "perfection" by collective computing on the system level, rather than on the device level in case of the digital (Boolean) approach.
Dr. Masa and his colleagues demonstrated the potential for such VLSI analog chips by making one, which they called the NeuroClassifier, a "general purpose neural network pattern classifier chip" which he claimed was "superior in computing speed not only to the state-of-the-art CPUs but also to the state-of-the-art neural network ASICs (Application Specific Integrated Circuits)."
Alan H. Kramer of the Neural Network Design Group at SGS-Thomson Microelectronics has also explored the advantages and limitations of analog computation and neural network architectures. As of October 1996 (IEEE Micro, Vol. 16, No. 5, October 1996) SGS-Thompson was investigating three analog VLSI chips to work on problems in image processing.
Some people argue (e.g., neuroscience writer John McCrone, in a personal communication) that analog circuitry will be necessary for a true artificial intelligence—the Machina sapiens I keep talking about. This argument could be right, but I'm sticking by my belief that since everything in the universe is ultimately quantum, and therefore composed of discrete—not analog—elements, then there is no reason in principle why discrete binary computing (or, at least, the multiple discrete states of quantum computing) cannot be made to emulate anything, including intelligence, mind, and consciousness.
My backup position is that even if analog computing does turn out to be fundamentally necessary for true AI: (1) we'll know it soon enough, given the rate at which the cognitive sciences are finding answers to how and why mind works; (2) analog computing is here already in the form of not only of the analog VSLI neural net chips just discussed, but also in the form of analog robots (the subject of next week's article); and (3) the appropriate technology will be developed well within my 30-year timeframe for the emergence of Machina sapiens.
Until next week,
NEXT WEEK: Analog Robots.
Help Wanted: Got questions or comments on this article or on any other AI-related subject under the sun? Post it in the AIBB!