Design enables large neural nets

October 8/15, 2003

Electricity and light each have strengths and weaknesses as communications media. This means, inevitably, that the two must be used together.

Researchers from Carlos III de Madrid University in Spain and the Massachusetts Institute of Technology have devised a neural network architecture that uses a different mix of optics and electronics than previous schemes in order to accommodate large numbers of neurons. The architecture leverages the computational strength of electronics and the fast communications abilities of light. It could be useful in systems that require optical input and neural net computation -- like those used for robotic vision.

Neural networks, like the human brain, have many interconnected elements. Neural networks learn by assigning weights to the connections between neurons and changing the weights based on use. This allows a specific set of inputs to be associated with a pattern of weighted neural connections.

The researchers' system carries out the neural weights assignment electronically, which cuts down on the number of difficult-to-align optical connections compared to previous optoelectronic neural networks, according to the researchers.

The design makes it possible to scale the system to a very high number of elements, or neurons, and the systemís optical interconnects allow for fast communications among neurons.

The device can be used in practical applications in two to five years, according to the researchers. The work appeared in the September, 2003 issue of Optical Engineering.

Page One

E-paper closes in on video

Magnetic memory makes logic

Old idea retooled for security

Crystal slows and speeds light

News briefs:
Process orders nanowire arrays
CD writer generates holograms
Nanotubes boost storage
Nanotubes harvest electrons
Bacteria make more electricity
Design enables large neural nets

Research Watch blog

View from the High Ground Q&A
How It Works

RSS Feeds:
News  | Blog

Ad links:
Buy an ad link


Ad links: Clear History

Buy an ad link

Home     Archive     Resources    Feeds     Glossary
TRN Finder     Research Dir.    Events Dir.      Researchers     Bookshelf
   Contribute      Under Development     T-shirts etc.     Classifieds

© Copyright Technology Research News, LLC 2000-2010. All rights reserved.