Neurons store and transmit information in the brain. Source: CNRI / SPL
A neuron-based superconducting computing chip that processes information more efficiently and quickly than the human brain.New achievements recently published in Progress in Science may lead scientists to develop state-of-the-art computing devices to design mimicry biological systems A major benchmark.Although there are still many obstacles before it is commercially available, the study opens the door to more natural machine learning software.
Nowadays, artificial intelligence software is increasingly starting to mimic the human brain, and algorithms such as Google's automatic image classification and language learning programs can also perform complex tasks using artificial neural networks, but because conventional computer software can not be designed Running a brain-like algorithm, these machines require more computational power than the human brain does.
'There will surely be a better way to do this because nature will find a better solution,' said study co-author Michael Schneider, a physicist at the National Institute of Standards and Technology (NIST).
NIST is one of several teams that want to develop neuromorphic hardware that mimics the human brain and hopefully the neuromorphic hardware will run brain-based software more efficiently. In conventional electronic systems, transistors are often spaced at regular intervals The exact amount of information is processed (binary digits 0 or 1), but neuromorphic hardware can accumulate a small amount of information from multiple sources and change that information to produce a different type of signal and, if needed, emit a current , Just as neurons discharge, so this neuromorphic hardware requires less energy to run.
However, these devices have so far been ineffective, especially when transistors need to transmit information across gaps or synapses, so the Schneider team used niobium superconductors to create neuron-like electrodes that can conduct electricity without resistance , Researchers use thousands of magnetic manganese nanoclusters to fill gaps in superconductors.
By changing the number of magnetic fields in the synapse, these nanoclusters can be aligned in different directions, allowing the system to encode information both in the electrical and magnetic directions, giving the system more power than other neuromorphic systems Of computing power, while not taking up additional physical space.
These synapses discharge 1 billion times per second, several orders of magnitude faster than human neurons, and consume only about ten thousandth of the biological synapse energy in the computer simulation process, passed to the next Synthesizing neurons can be used to check input information from up to nine sources before one electrode, but thousands of synapses are needed before the system based on this technology is used for complex calculations, and Schneider said it can be expanded to that level For further study and analysis.
Another problem is that the synapse can only operate at near absolute zero temperatures and need to be cooled with liquid nitrogen.While Steven Furber, a computer engineer at the University of Manchester in the UK, points out that this may make the chip impractical in small devices Although large data centers may be able to service it, Schneider said cooling the device may require less energy than operating a conventional electronic system with a significant amount of computing power.
Carver Mead, an electrical engineer at the California Institute of Technology praised the study as a new way of calculating neuromorphic shapes. "There is a lot of speculation in the field right now, and we're happy to see that fine work can be done objectively "He said, but it may take a long time before the chip is actually used in the computing world, and there are currently fierce competition and challenges from many other neuromorphic computing devices.
Furber also emphasizes the wide range of practical applications of this new device. "This device technology is also very interesting, but we are not fully aware of the key features of these biological synapses today and do not know how to use them more efficiently. For example, he said, there are still many issues that need to be addressed, such as how these synapses reshape themselves as they form? This makes it very difficult for researchers to rebuild the process in memory chips.
Despite this, Furber said it will take 10 years or more for a new computing device to enter the market, and even though neuroscientists have a hard time understanding the human brain, it is imperative that they develop as many different techniques as possible.