Cortex chip goes both waysBy Kimberly Patch, Technology Research News
In electronics, digital and analog signals are very different languages that generally don't mix. Digital circuits speak in precise ones and zeros while analog circuits speak in waves.
The brain's neurons, however, are bilingual.
When researchers put together a simple silicon chip that mimiced the brain's communication paths in a slightly different way than traditional neural nets they found something unexpected: it spoke both languages.
Digital circuits use feedback to turn a switch on or off, while analog systems use feedback to amplify a signal. Traditionally, a circuit exhibits either digital or analog behavior. For example, if an amplifier has positive feedback, or a gain, of five, it will amplify an input of one to an output of five. But if the signal is too strong it will overload the system and produce a digital off.
The cortex inspired chip, however, employed both types of signals at the same time, said researcher Rahul Sarpeshkar. "Positive feedback created explosive instability that completely shut off some neurons, but the neurons that were not shut off actually were stable and amplified their analog inputs," said Sarpeshkar, an assistant professor in the electrical engineering and computer science department at MIT and a scientific consultant for Bell Labs.
There are a lot of potential applications for hybrid chips simply because information from sensory devices like cameras, microphones and electronic noses is voluminous and analog: "The light that you see comes in many, many shades of gray and many, many shades of intensity. The sounds that you hear come in 12 orders of magnitude of intensities, said Sarpeshkar. Today's computers convert this sensory input to digital, then process the whole mountain of data.
Analog wave signals can, like the human brain, approximate well, but digital circuits must go through many steps of logic to do something as simple as adding one and one. It is only because digital computers run millions of steps a second that they seem fast. "Moderate precision addition can be done in analog by just hooking two signals together on a wire. It would take 3,000 transistors in digital technology to set up a circuit that would accomplish the same feat," said Sarpeshkar.
So wholesale digital processing is "computing with needless precision on gobs and gobs of data that you don't need," he said.
Analog signals, however, are not precise and are prone to drift, making complicated analog systems an impossibility. Digital signals are very precise and remain so even when the problem gets very complicated. Analog recordings, for instance, degrade with each copy and may eventually drift far from the original, while digital copies remain stable throughout many generations.
A hybrid approach could potentially use analog systems to quickly and efficiently pare down large amounts of analog data to a much smaller pile, then and use its digital abilities to crunch just that data.
Some of today's devices take advantage of both strengths by using digital and analog circuits. But they must use converters to translate information from analog to digital, which takes a lot of power. In a world where devices are getting more portable, a hybrid chip that does not need to convert anything has a lot of potential, Sarpeshkar said.
The researchers' experimental chip contained 16 hardware neurons. While traditional neural networks are fully connected, in this model the neurons were connected only to their four nearest neighbors and to a global inhibitory neuron -- a pattern closely modeled after the human cortex, Sarpeshkar said.
The connections among the neurons were excitatory, meaning when one neuron received output it stimulated those it was connected to, which created a positive feedback loop. Each neuron also stimulated the global inhibitory neuron, which then served out negative feedback, meaning when it was stimulated it calmed all the other neurons down. "This is the pattern of connectivity that is seen in [the] cortex," said Sarpeshkar.
"It certainly is an interesting circuit and a clever circuit, but the kind of generalities you can draw about the influence this will have on electronics in the future is a tough call," said Bernard Widrow, a professor of electrical engineering at Stanford University.
Sarpeshkar is taking the work further with project on pulsed hybrid circuits that use pulses and the space between pulses to represent digital and analog information, respectively. "The time between pulses is analog but the pulse is an all or nothing event. [Pulses] are uniquely suited to do hybrid computation because of that," he said.
The three-year project will attempt to build reprogrammable hybrid computers that can do digital arithmetic and store information digitally, but also have analog feedback loops that learn and adapt, he said.
It will be at least three to five years before practical applications like sensors that can take advantage of hybrid circuits may appear, said Sarpeshkar. Widespread applications of hybrid chips are probably 10 years away, he said.
Bell Labs funded the cortex inspired silicon circuit project, which is detailed in a paper that appeared in the June 22 issue of Nature magazine. Sarpeshkar's colleagues on the paper were Richard H. R. Hahnloser, Misha A. Mahowald and Rodney J. Douglas, all from the Institute of Neuroinformatics in Zürich Switzerland and Sebastian Seung from MIT and Bell Labs.
Sarpeshkar's ongoing work on hybrid circuits is funded by the Office of Naval Research.
Timeline: >3 years, >10 years
Funding: Corporate, Government
TRN Categories: Neural Networks; Integrated Circuits
Story Type: News
Related Elements: Technical Paper in Nature, June 22, 2000, p. 947
June 28/July 5, 2000
Cortex chip goes both ways
Sampling ability broadens quantum computing
Nano-scale plotter goes parallel
NASA grasps intricacies of human hand
Algorithm evolves more efficient engine
Research News Roundup
Research Watch blog
View from the High Ground Q&A
How It Works
News | Blog | Books
Buy an ad link
Ad links: Clear History
Buy an ad link
© Copyright Technology Research News, LLC 2000-2006. All rights reserved.