Silicon Investor (SI) -- The First Internet Community

STOCKTALK

We've detected that you're using an ad content blocking browser plug-in or feature. Ads provide a critical source of revenue to the continued operation of Silicon Investor. We ask that you disable ad blocking while on Silicon Investor in the best interests of our community. If you are not using an ad blocker but are still receiving this message, make sure your browser's tracking protection is set to the 'standard' level.

Strategies & Market Trends : Neural Nets - A tool for the 90's -- Ignore unavailable to you. Want to Upgrade?

To: Optim who wrote (665)	9/27/1999 12:07:00 PM
From: Larry Livingston	Respond to of 871

Thanks Optim, I think I'll go ahead and try them out. L.L.

To: Optim who wrote (665)	10/2/1999 12:16:00 PM
From: Larry Livingston	Read Replies (1) \| Respond to of 871

SLIGHTLY OFF TOPIC but of interest nonetheless, I think. Here are a few excerpts of an interesting article on a new neural net technology that has been developed for speech recognition. Perhaps there are some of us who might be familiar with the technology and who could provide more information. Maybe it could useful in developing better nets for financial prediction. Superhuman Speech Machine Leander Kahney "Two biomedical engineers claim to have created the first "superhuman" speech recognition system following a breakthrough in neural network technology. " "If the claims pan out, the system may herald a new era of neural network computing, which has languished for a decade after failing to deliver on the promise of software that mimics human intelligence. " "Developed by biomedical engineering professor Theodore Berger and Jim-Shih Liaw, director of the university's Laboratory for Neural Dynamics, the Berger-Liaw Neural Network Speaker Independent Speech Recognition System (SRS) for the first time displayed better-than-human performance in word recognition tests. " "According to the researchers, the SRS dramatically outperformed an off-the-shelf word recognition system as well as human listeners, showing an uncanny ability to pick out keywords almost completely drowned out by white noise. Speaking by phone from his home, Berger said the new technology may lead to smaller, faster neural nets that outperform their predecessors to the extreme." "Neural nets simulate biological nervous systems, consisting of mathematical models of neurons and the connections between them. Unlike conventional software, they aren't programmed, but trained to associate patterns of input with outputs. Like biological systems, neural nets are highly adaptive, cope well with incomplete information and noise, and are suited to complex pattern recognition tasks like identifying words. However, after a flurry of activity in the 80s, neural nets fell out of fashion after failing to create truly artificially intelligent software, as boosters had predicted. Incredibly, Berger said the SRS's superhuman performance is the result of implementing a fairly obvious feature that had previously been overlooked. Berger, a neurobiologist by training, said he and Liaw realized that a crucial characteristic of biological nervous systems -- the ability of neurons to change their behavior according to the timing of input signals -- had never before been implemented in neural nets. "It's hard to believe, I know," he said. "But there it is." In Berger and Liaw's model, the timing of inputs are crucial. A neuron may fire if two input signals come very close together, but will not fire if the same two signals are just slightly further apart. Thanks to this, the SRS comprises just 11 neurons and 30 links. By contrast, previous attempts to build word-recognizing neural nets required hundreds of nodes and thousands of connections, Berger said. "It's just wonderful what you can do with a small number of neurons," Berger said. "It's just amazing. We've haven't even started to make them complicated. It's going to be really interesting." Although the system was trained on only 12 different words and tested with eight different speakers, Berger said he was confident it could learn a much larger vocabulary. Two major advantages over today's commercial systems, Berger said, is that the SRS is speaker-independent -- it can recognize a word regardless of who says it -- and it never gives false-positive readings. If the system doesn't recognize a word, it doesn't hazard a guess. " The full article is from Wired Magazine wired.com