The biggest fear of those who champion neural networks is guilt-by-comparison with the artificial intelligence camp. They're not alone in this. Object-oriented advocates, as well as most other popular technologies that make the front pages of pseudo-technology news tabloids, don't want to be snake-bit by the same type of hype that poisoned AI development. The frontal assaults of AI and expert systems, fueled by big money and bigger promises, have been nonexistent for neural nets and, although those neural net developers trying to eke out a living might disagree, the lack of venture capital has probably been a blessing. Less hype buys more time, at least as long as enough money comes in to keep the lights turned on.
Nevertheless, there continues to be a lot of interest and development in neural nets. A survey recently published by Future Technology Surveys of Madison, Georgia listed over 200 companies and organizations currently producing neural-related products or undertaking serious neural net research. And it just isn't the little guys doing all this research, either.
Among the big outfits testing the neural net waters is Intel. Early last summer, a ten-member Intel engineering team, under the direction of Mark Holler, rolled out an Electronic Trainable Artificial Neural Network (ETANN) chip that is capable of up to 2 billion multiplies and accumulates per second. To put the chip to work, Holler and his crew have built a prototype ETANN-based board that plugs into the PC AT bus; Mark Lawrence and the folks at California Scientific (developers of BrainMaker, a neural net simulation package for PCs) are developing the software tools that let you use the system. There's also a rumored Intel research project that will put a version of the ETANN board into an i860-based system that can achieve 33 billion connections per second.
Intel isn't the only big IC manufacturer poking around in neural nets. A few months ago, Sharp introduced a neural-network image-processor chipset that simulates human vision and, the company claims, supports PC applications at speeds up to 700 MIPS.
These examples illustrate another trend in the neural net world -- a transition from software to hardware. Within ten years, or so say the experts, more neural nets will be implemented in hardware than software. Until then, engineers will begin to overcome many challenges, including the implementation of back propagation in hardware and the parallelization of the entire scheme.
So where does this leave software developers? For one thing, a whole new class of development tools is in the offing, designed for specific neural net hardware implementations. Another type of tool will be like that described by Andy Czuchry in this issue, whereby designers can match the right neural model with the task at hand. Nor will the simulators go away; they may be used to simulate the right net with appropriate learning, then generate source code to be frozen in silicon.
In her keynote address at Miller-Freeman's SD'90, Smalltalk pioneer and ParcPlace System's president Adele Goldberg expressed a concern similar to that I wrote about in this space last month -- the spread of litigation and its effect on the software industry.
Although her talk concerned a wide variety of legal issues -- from intellectual property to the emerging problem of who owns the design and implementation of objects, as in object-oriented programming -- she spent a fair amount of time on copyrights and patents. "Lawyers will always tell you two things," she said, "try to patent or copyright whatever you do." She went on to describe a speech she gave to a group of lawyers, where she was asked how to convince software developers to protect their works. "My answer was simple," she said. "Tell them to protect their work so that they have the choice later on to give it away." Not doing so, she explained, opens the door for someone to come along and take it away. But Goldberg wasn't engaged in lawyer-bashing, no matter how easy that is. What she was presenting was a persuasive argument for open standards and open licensing. She pointed out that among the problems litigation forces upon us are the waste of time and money, the fear of alliances, the inability of entrepreneurs who lack clear patent or copyright protection to attract investors, and the expense of starting up new businesses.
One of the main points of her talk was simply to "reassert an often unstated goal of our industry -- to share ideas and to challenge one another with our innovative expressions of those ideas." Nicely put.
And no, the favored horse running in the first race at Bay Meadows racetrack the other night wasn't our official mascot, even though the nag's name was "Dr. Dobbs." Although, he lost by a nose in a photo finish, the good Doctor is surely chomping at the bit to get into the next race. We'll keep you posted on his progress this season.