Hybrid digital/analog processor pursues machine vision apps

Hybrid digital/analog processor pursues machine vision apps
BERKELEY, Calif. — Researchers in the United States and Hungary have developed a new kind of hybrid analog/digital computer that is being tailored for an increasing number of image-processing and machine-vision applications. Though still at the laboratory, prototype stage, the machine already has been used in real applications including the inspection of mammograms for particular kinds of cancer.

Called a cellular neural network (CNN), the processor is estimated to be three orders of magnitude faster than competing technologies for conventional operations. It can also perform functions that would be impossible with digital electronics. Researchers say they are hopeful the machine's immense power and easy programmability will encourage engineers to abandon their prejudices about analog computing.

"Sheer processing speed is the special ability of the CNN which underlies all other advantages," said Ken Crounse, a researcher at the Nonlinear Electronics Laboratory at the University of California at Berkeley, who has spent several years investigating the CNN's potential. "The area-to-flops and power-to-flops ratios look to be very good as a consequence." The device is particularly suited for image processing, he said.

Crounse continued: "The natural mapping from image pixels to CNN cells and the fast settling time of the analog dynamics could enable the CNN to perform certain types of image-processing operations on a whole image in nanosecond time frames: maybe three to six orders of magnitude faster than current digital techniques."

The CNN was invented by Leon Chua and his colleagues at UC Berkeley. The device consists of an array of analog processing circuits with a threshold that can be set externally. Besides taking input from a sensor connected to the outside world, each processor also takes information from each of its near neighbors.

Friendly neighbors

Thus, each processor has a "sphere of influence" that includes itself, the cells at its four sides and those at its four corners. Information going between the central processor and any of its neighbors is either amplified or diminished based on the weight of the interconnection between them. Each of these can also be set externally.

Because the network is homogeneous, its settings can be boiled down to just 19 numbers. These are the input and output weights between the central processor and its eight neighbors, the input and output weights between the central processor and itself, and the processor threshold. These numbers hold the key to the operation of the entire network.

To show what the CNN can do and demonstrate its superiority over conventional digital-signal processing (DSP) chips, professor Tams Roska and his group at the Computer and Automation Research Institute of the Hungarian Academy of Sciences built a digital CNN simulator. Among other things, the board can calculate how much digital accuracy a given program would need so the simulated result matches the accuracy of an actual CNN chip.

Using these numbers, Roska was able to show that, for a binary 128 x 128 image — close to the maximum size that can be achieved with today's technology — a CNN chip could calculate a Laplace transformation more than 1,000 times faster than a Pentium II. For a gray-scale image, the CNN would be more than 2,000 times faster.

The largest gray-scale CNN chip built so far has 64 processors and a 128 x128 chip is currently on the drawing board. The ultimate goal would be to produce arrays of cells large enough to directly process incoming pixel data from large, integrated sensor arrays such as CCD cameras. Researchers expect the size of the CNN cells to drop, and array sizes to grow, as development continues.

The new technology can take advantage of many components already used in digital and analog electronics. But one novel device, known alternately as a voltage-controlled current source or operational transconductance amplifier, provides the cell's analog nonlinearity.

While they wait for bigger arrays to become available, the researchers have designed a CNN "Universal Machine" to deal with large images by processing them in blocks. They found little difficulty in keeping track of the boundary conditions, and results have generally been good.

One possible obstacle to the technology's success, however, is that it requires a completely different way of thinking on the part of engineers. According to Crounse, "Transform-domain processing is very difficult on the CNN hardware, but by using the proper design tools, equivalent functions can be performed on the CNN in the spatio-temporal domain. Just like digital architectures, the CNN has its limitations, but they are different limitations. It's just that we're used to designing around the limitations of digital."

Crounse gave another example: "Digital designers are used to thinking in terms of bits of accuracy and quantization noise, but analog noise is quite different, and biological systems have found ways of dealing with it robustly."

Another barrier may be economic. "The CNN would almost certainly provide a benefit for numerous applications," Crounse said. "However, no one of them has yet to justify the cost of full-blown development. The key will be to find an incremental, though rapid, development path."

The engineering community has invested heavily in the digital technological path over the past 20 to 30 years, he said, so much so "that analog design and manufacturing has a lot of catching up to do to be able to exploit even today's technological capabilities."

On the positive side, he sees the coming availability of relatively small CNN chips as a major step forward. "I'm expecting an explosion of new algorithm development: methods that were previously undreamed of because they would have taken hours or days to run on conventional processors but could be done in real-time on the CNN." One possibility: new video-compression algorithms that have a complex visual-system model in a feedback loop. This would allow iterative evaluation of the human perception of compression quality.

Despite the success of the CNN in image processing, the network is much more than just a fast DSP: It offers a means of simulating nature. In his recent book, CNN: A Paradigm for Complexity, Berkeley's Chua shows that CNN functions can produce all sorts of natural patterns, from phosphenes — the patterns you see when you close your eyes — to the patterns on animal fur or fish scales, to the growth of a dendritic tree that looks like nerve nets in the brain. They can also solve difficult perceptual problems, such as finding their way through a maze or labyrinth, a classic problem that the single-layer Perceptron neural network couldn't handle. On a more bizarre level, CNNs have even been able to imitate common optical illusions like the "tunnel of light" reported by survivors of near-death experiences.

The CNN's success in these biological problems comes from the system's ability to perform the analog equivalent of reaction-diffusion equations: equations that are extremely computer intensive. Chua makes no claim that the CNN mechanism has any direct similarity with the biological systems it can simulate, but vision researchers, including some at UC Berkeley, are already starting to use the device to test theories about the workings of the eye.