Artificial synapses copy brain dynamics

By Kimberly Patch, Technology Research News

The human brain is a stunningly vast and complicated network. It is a massively parallel web of about 100 billion neurons that each harbor about 10,000 connections to other neurons. The functions the biological brain carries out are even more impressive -- a network of neurons can easily outperform any electronic network of the same size.

In an attempt to uncover at least some of the brain's computing efficiencies, a group of researchers from Graz University of Technology in Austria and Cold Spring Harbor Laboratory have built an artificial neural network architecture that mimics the way neurons time their signals to each other.

In going about its routine business, the brain regularly solves difficult computational problems that remain beyond the reach of the most powerful computers, said Thomas Natschlager, a research and teaching assistant at the Institute for Theoretical Computer Science at Graz University of Technology. The processes the brain uses for doing this, however, have remained elusive.

One thing the brain can do that is very difficult for computers is react appropriately to a signal that varies over time, like spoken language. This is essentially a complicated, or nonlinear filtering function. A filter is any device that produces output in response to an input signal over time, and the output of a simple linear filter varies directly with the signal it's reacting to.

Nonlinear filters, which are more complicated mathematically and thus more difficult to simulate, are also prevalent in the natural world. The brain is essentially a big nonlinear filter that perceives input like auditory and visual signals and produces output like arm movements or speech, Natschlager said.

Recent research into the dynamics of neurons provided the clue the researchers needed to work out a neural network architecture that performs more like the brain. The research shows that neurons gain information from the timing of electrical signals in ways that electronics-based computers do not.

It is well known that synapses incorporate learning by strengthening and weakening connections over a period that may encompass several days. In the case of long-term learning, synapses change in order to store knowledge, and the changes remain stable for years.

It is also becoming apparent that neurons can also change temporarily for a few seconds or milliseconds in order to perform a task of the moment. This ability to reconfigure the network on-the-fly may help neurons increase their computing power.

"Usually it is assumed that a synapse responds to each incoming input, [or] spike, with an electrical impulse, [or] postsynaptic potential with the same amplitude, but this is not true," said Natschlager.

Experimental evidence has shown that the amplitude of the responding electrical impulse depends on the exact timing of previous inputs. Essentially, the response of a neuron to a neighboring neuron's signal depends on the history of multiple signals, or spike trains between them. The synapses between neurons "are not static devices, but respond differently to each input," said Natschlager.

The researchers model incorporates this biological trick, allowing synaptic strength between neurons to continually change. "The architecture we propose provides a framework for studying how neural circuits compute in real-time," said Natschlager.

The researchers' work added evidence that synaptic dynamics are a key to the brain's efficiency by showing that a relatively simple artificial neural network that incorporates dynamic synapses can do some of the difficult things a brain does.

The dynamic model of timing between neurons gave the network another parameter to work with, which increased its real-time processing capabilities. The model allows "rather simple networks to compute an arbitrary nonlinear filter. Such networks can approximate a surprisingly large class of nonlinear filters," according to the researchers.

The research is a contribution to understanding how the biological model of temporarily fluctuating synapse strengths can speed networks, said Michael Arbib, a professor of computer science and neuroscience at the University of Southern California. "It's... the latest chapter in the idea that if you take seriously the rapid changes of the synapses you can get a lot more computing power in your network," he said.

The sophistication of the brain's network hides a lot of complications from us, which makes what we do seem easy, Arbib added. For instance, "the brain doesn't work by... keeping only [like] information together. It has lots of specialized subsystems. The question is how do those different subsystems know which of the information they have goes with specific information that another system is processing. That's called the binding problem -- how do you bind together these different pieces of information," he said.

The Graz and Cold Spring Harbor researchers are working on a new framework that will allow them to further investigate the types of communications that can be carried out on spike trains, said Natschlager. "The ultimate goal is to understand how... neurons and synapses work together to allow the brain to process time-varying signals so efficiently."

The researchers are also applying their work practically in collaboration with speech recognition researchers. "We think that within a time frame of five years we will know how well dynamic synapses are suited for real world applications," Natschlager said.

Natschlager's research colleagues were Wolfgang Maas of Graz University of Technology in Austria and Anthony Zador of Cold Spring Harbor Laboratory. They published the research in the February, 2001 issue of the journal Network: Computation and Neural Systems. The research was funded by Neurocolt2 and the Austrian Science Fund (FWF).

Timeline:   > 5 years
Funding:   Government, Institutional
TRN Categories:   Neural Networks
Story Type:   News
Related Elements:  Technical paper, "Efficient Temporal Processing with Biologically Realistic Dynamic Synapses," Network: Computation and Neural Systems, February, 2001. The paper is posted at .


June 13/20, 2001

Page One

Stressed silicon goes faster

Artificial synapses copy brain dynamics

DNA device detects light signals

Lightwaves channel atoms to make chips

Process promises better LCD production


Research News Roundup
Research Watch blog

View from the High Ground Q&A
How It Works

RSS Feeds:
News  | Blog  | Books 

Ad links:
Buy an ad link


Ad links: Clear History

Buy an ad link

Home     Archive     Resources    Feeds     Offline Publications     Glossary
TRN Finder     Research Dir.    Events Dir.      Researchers     Bookshelf
   Contribute      Under Development     T-shirts etc.     Classifieds
Forum    Comments    Feedback     About TRN

© Copyright Technology Research News, LLC 2000-2006. All rights reserved.