Intelligent Systems Report € October € 1996 € Volume 13€ No. 10

WCNN '96

Neural nets are up, trade shows are down

---

By David Blanchard
Editor

Neural networks have never been more popular than they are today. They are being used by physicians to detect cervical cancer, by police forces to identify and prosecute felons, by banks to predict credit card fraud, by mortgage companies to assess real estate values, by Internet providers to perform intelligent concept searches, by printed circuit designers to improve the quality of their board layouts, by industrial plants to control the level of NOx emissions, and even by airplane manufacturers to build safer planes. The floodgates have been opened wide for neural network-based applications, with no end in sight.

It's ironic, then, that neural network conferences have never been less popular than they are now, at least to judge by attendance figures. For instance, the World Congress on Neural Networks (WCNN '96), held last month in San Diego, Calif., drew a mere 350 attendees, mostly researchers rather than application developers. (For the record, the IEEE's ICNN '96 show, held earlier this year in Washington, D.C., had roughly the same low attendance).

The irony of the situation takes on poignant overtones when you consider that the wound is self-inflicted. Once upon a time there used to be a large conference and trade show called the International Joint Conference on Neural Networks, co-sponsored by the two largest neural network trade associations -- the IEEE's Neural Network Council and the International Neural Network Society (INNS). The events were well attended, gathered together a diverse international collection of researchers and developers, dealt with real-world business applications and entrepreneurial pursuits, and hosted most if not all of the major tool vendors. However, the joint conferences came to an end a couple of years ago when the two associations decided -- in their infinitesimal wisdom -- to go their separate ways and host separate conferences. These conferences have become "preaching to the choir" affairs, predominantly drawing only the main players in the neural network field rather than attracting any new blood. As a result, the IEEE and the INNS are fighting for the same slice of an ever-shrinking pie.

Which is not to say that WCNN '96, sponsored by the INNS, was bereft of any interest. After all, while attendance was light, the quality of the presenters was quite high, representing as it did the state of the art of neural network research.


Neuro-medical applications
For instance, following in the wake of the highly successful Papnet pap smear reading system (see ISR, September 1996), researchers are seeking to exploit other medical areas that could benefit from neural network technology. Detecting breast cancer and classifying mammograms looks to be the next major area of concentration, as several research groups -- Duke University, Rutgers University, University of Chicago, Southern Methodist University, Baylor University, and Paris-based INSERM -- discussed their latest efforts in developing intelligent systems to identify cancerous cells.

In fact, the trend of intelligent diagnostics has yielded yet another acronym -- CADx, for computer-aided diagnosis. The FDA's Center for Devices and Radiological Health (Rockville, Md.) has seen a measurable increase in the number of products submitted for premarket review which contain neural networks. According to Dr. Harry Burke of New York Medical College (Valhalla, N.Y.), "CADx devices do not replace physicians; rather, they aid the physician in making the most accurate diagnoses. The diagnostic decision and responsibility continues to reside with the physician."

Burke thinks it is highly unlikely that any CADx program will generate hundreds of millions to billions of dollars in sales, because "sales will be to physicians, physicians will not buy many copies of the software many times, and it is relatively easy to create me-too products in the CADx domain." Burke's point is a good one, but given the national advertising campaign undertaken by Neuromedical Systems (Suffern, N.Y.) to market its Papnet system, the potential for a neural network CADx system has yet to be really tapped.

Other neural network-based CADx projects in the works include Duke University's system to diagnose pulmonary embolisms, an ophthalmic system for image segmentation developed at the University of New Mexico, and a neural system to diagnose coronary artery disease under development at SUNY at Stony Brook.


Awakening to a neural world
WCNN '96 made a curious choice in its keynote speaker: Dr. Oliver Sacks, best known for his book "Awakenings" and the movie based on it, starring Robin Williams as Sacks. What made Sacks such an unusual speaker was that his life work has been in clinical neurology, rather than artificial neural networks. In fact, he never even used the words "neural networks" in his presentation, focusing instead on his insights into the nature of consciousness. As a noted author and clinician, Sacks spoke before a packed house at WCNN; however, his remarks were more of an anecdotal variety than they were of any fundamental use to neural network researchers and developers.

Another figure of international repute (though no movie has yet been made of his life) was also on hand at WCNN '96 -- Lotfi Zadeh, the "father" of fuzzy logic. Zadeh's work lately has centered on what he refers to as "computing with words," a less awkward way of referring to the "fuzziness" factor. Computing with words provides approximate methods for solving nonlinear problems, and leads ultimately to more intelligent machines. In fact, Zadeh believes that machines have intelligence quotients that can be gauged, much as a human being's IQ can be measured.


Vendor no-shows
The WCNN '96 exhibition -- if you want to call it that -- was virtually invisible. Other than the usual collection of book publishers, only two neural software vendors showed their wares at the show: NeuroDimension, vendor of the Neuro-Solutions system, and Logical Designs, vendor of the Think system. It was a far cry from the IJCNN glory days of the late 1980s/early 1990s, when more than 10 times that many vendors displayed their neural wares. The major neural network vendors continue to be extremely active, of course -- just a glance at the "News Briefs" and "Intelligent Applications" departments in this newsletter every month bear witness to that -- but clearly the word has gotten out that the INNS and IEEE neural network shows are no longer a prime attraction for corporate decision-makers with deep pockets.

In the final analysis, a lot of important and trail-blazing work is underway in the neural network field, some of it of immediate application, such as medical systems comparable to Papnet or electronic noses such as the one developed by Pacific Northwest National Laboratory (see sidebar). However, as cutting edge technologies mature, the interest in the research area tends to drop way off, reducing what once were impressive trade shows and conferences to mere shadows of their former selves. It's ultimately much more important that the industry survive than the trade shows, and while frankly WCNN '96 wasn't much of a show, neural networks themselves are hotter than ever.