Most of the current action in deploying brain-emulating neural networks focuses on a handful of sensor programs sponsored by the Defense Advanced Research Projects Agency (Darpa), which funded much of the early work on neural networks. Jasper Lupo, the Pentagon's chief scientist and a pioneering researcher in the field, said Sunday (July 15) at the International Neural Networks Society conference here that the agency is seeking industry and academic proposals for new approaches to developing large sensor arrays based on the technology.

One solicitation seeks new mathematical ideas for large sensor arrays and related communications links. The second focuses on "multimodal, multidimensional sensor methods." Both are geared to developing what Lupo called "a global nervous system" made up of networked sensor arrays covering everything from handheld forward-looking IR to chemical sensors.

"Every object on the battlefield can be a sensor platform, even a human being," Lupo said. Besides wearable sensors, he said Pentagon planners could install chemical and other detectors on a fence post or a traffic light.

Darpa's Smart Sensor Web program seeks to use neural network technology to deploy large arrays of small sensors, each costing no more than $300 a piece. It is also seeking to use microelectromechanical systems for a "smart dust" program that would release tiny sensors into the wind.

The approach is fundamentally different from the DOD's traditional approach of packing platforms with huge, costly, standalone sensors, said Lupo, who oversees U.S. military sensor programs.

While most of the new sensor development is in its infancy, Lupo said neural network technology is already being fielded in a handful of first general weapon systems. In the mid-1990s, for instance, the Office of Naval Research developed a model of the rabbit's hippocampus used to listen to helicopter rotors. The system proved better than mechanics in detecting potential catastrophic failures in military helicopters and was deployed by the U.S. military.

Another sensor in development by the military since the 1980s is automatic target recognition (ATR) systems. "ATR is being fielded in many systems in large measure due to neural networks," Lupo said. ATR combines different passive and active sensors to help pilots and other forces spot targets concealed by ground clutter and signal noise. It then helps direct weapons to their targets.

Such systems are part of a "third wave of computing" that could use neural networks to build sensors and other machines capable of the "unsupervised learning" exhibited by the human brain. Researchers here said they are using a range of new algorithms designed to improve machine vision and other systems.

Harold Szu, director of the Digital Media RF Laboratory at George Washington University, based here, reported at the conference on a remote sensing technique that uses neural net algorithms to improve resolution on Landsat satellite images. The technique used the Bell-Sejnowski-Amari-Oja algorithm for independent component analysis post-processing.

"We solved the problem pixel by pixel in real time, independent of the image size, hyper-spectral remote sensing or brain imaging," Szu reported.

As neural network technology development proceeds apace, it is also running into an old problem in the research community: The reluctance to share hard-earned data with fellow researchers. Stephen Koslow, director of the Office of Neuroinformatics at National Institutes of Mental Health (Bethesda, Md.), said data sharing is lagging just as the application of neural network technology in neuroscience is soaring. Information sharing is critical, Koslow said, as the problems being addressed grow in complexity and research funds grow more scarce.

One hot research area focuses on how to create functional connections between microcircuits and the human nervous system. Koslow said the National Institute of Biomedical Imaging and Engineering was formed in the late 1990s to tackle that problem.