Most people are still trying to get to grips with the digital revolution. But a small group of engineers thinks the whole thing is, if not exactly a mistake, at least not the be-all and end-all of electronics. These engineers reckon that digital's alternative-analogue technology-is being needlessly ignored. In fields such as optical recognition, for example, digital technology tends to make pretty heavy weather of things. Now, with the help of a newly designed analogue neural net, they hope to put that right.
Analogue systems work with continuous signals that are processed according to the physical layout of a machine, rather than by ordering a series of ones and zeros according to the dictates of a logic-based program (which is the digital way). That generally gives analogue machines an advantage in terms of size, efficiency and raw computing power. The reason that people have come to prefer digital computers is that it is difficult to harness that power-for the two things that analogue computers do not have are flexibilityand simplicity. If armed with the right software (and assuming its processor is powerful enough) a digital computer can solve any problem that can be expressed as a logical algorithm. But, because analogue computers have their problem-solving abilities built directly into their electronics, if you want to perform a different sort of calculation, you have to design and build a new circuit to do it.
A collaboration between researchers in America and Hungary has, however, come up with an answer to this problem: the cellular non-linear network (CNN). The CNN chip was invented by Leon Chua at the University of California, Berkeley. It consists of an array of identical analogue processors, each connected to its eight nearest neighbours in a way that can affect those neighbours' behaviour. This arrangement makes the chip particularly suitable for image-processing, because each analogue element of the array can be made to correspond to a single pixel (one of the individual dots of which an image is composed). The pixels in an image can thus be processed simultaneously.
Since most image-processing operations, such as sharpening up a fuzzy document, or altering the colours in a photograph, involve looking at the value of a particular pixel and then altering it depending on the value of adjacent pixels, this is a real advantage. Digital systems, which process pixels one at a time, have to translate the two-dimensional image into a one-dimensional stream of data. Most neighbouring pixels are no longer neighbours in such a data stream, so comparing them is a long-winded business, since the processor has to hunt down the relevant information before it starts. In order for the CNN technology to be really useful, however, it needs to be programmable, unlike traditional analogue machines. This problem was solved by Dr Chua's collaborator, Tamas Roska, and his group at the Computer and Automation Research Institute of the Hungarian Academy of Sciences in Budapest. Dr Roska has invented something he calls the CNN Universal Machine. This is a system that allows a CNN chip to be programmed, somewhat in the way of a digital chip, to perform many image-processing tasks. The "program" consists of sequences of a mere 19 numbers. Sixteen of these modify the strengths of the inward and outward connections between each element in the array and each of its eight neighbours. Two modify the elements' connections with themselves. And one is a "threshold" value. If the combined values of the inputs to a given element (as modified by the appropriate numbers) exceed this threshold, that element's behaviour changes radically. This is the point at which the element makes its alteration to the image (or not, as the case may be).
There is just one problem. Although they know that by playing with these 19 numbers, they can make the Universal Machine do lots of things, Dr Chua, Dr Roska and their colleagues cannot easily predict which sets of numbers will result in what behaviour. Trial and error has turned out to be the best way to find out. But among the tasks the machine can perform in a single step are edge detection, inversion, the identification of lines and shapes of different orientations, and the filling-in of concave shapes-all things that are useful in image-processing, and not all of which can be managed by digital technology. The ultimate prize for the researchers is machine vision. Their device can already manage this to a limited extent. It is, for instance, being tested in Hungary for detecting signs of cancer in mammograms.
Proper machine vision, though, will require a lot of pixels. So far, the largest CNN chip that the researchers have managed to build is a 64 x 64 array, though they are working on a 128 x 128 version that tests the limits of the current fabrication technology. The ideal, however, would be a chip with a 1,000 x 1,000 grid of processors. That would be enough to handle every pixel coming in from the average video camera in a single step. The camera would then not only take pictures; it would know what it was looking at.
THE ECONOMIST, March 6th