Chinese scientists unlock potential of memristor semiconductor building block that could boost artificial intelligence, self-driving cars and more
- World’s first fully system-integrated memristor chip could make artificial intelligence smarter and up to 75 times more efficient, researchers say
- Advances could lead to AI that is capable of more human-like learning, with implications for how smart devices and autonomous driving work
The world’s first fully system-integrated memristor chip has been unveiled by a team of Chinese scientists who believe it could not only make artificial intelligence smarter, but also more time and energy efficient.
“Learning is highly important,” for edge intelligence devices, the research team from Tsinghua University said in their study released in the journal Science on September 15, referencing devices that process data internally with technology like AI.
02:39
Korean researchers unveil world’s first humanoid robot pilot
China’s new memristor chip is a significant step forward in developing such technology, according to the researchers.
“Memristor-based computing technology has recently received considerable attention because of its potential to overcome the so-called ‘von Neumann bottleneck’ of conventional computing architecture,” Yury Suleymanov, an associate editor at Science, said in his editor’s summary of the paper, referring to the computational limitation set by the separation of memory and processing.
A resistor is an element of a circuit that is able to limit the flow of energy by providing resistance to electrons flowing through.
China downplays reports of Taiwanese firms helping Huawei’s semiconductor ambitions
A memristor – a contraction for memory resistor – is able to remember the most recent value of current that was passed through it when it was turned on, meaning future resistance can depend on prior history, according to Nature Electronics.
As such, it is capable of improvement-based learning, or maintaining pre-acquired knowledge when something new is learned.
This differs from transfer learning, which focuses on moving to a new set of data and can sacrifice the accuracy of prior data, according to the paper.
To train artificial neural networks, which mimic how human neurons pass on data in the brain, conventional hardware requires a great deal of energy and time to move data between the computing and memory units.
The study is an important step towards future chips with high energy efficiency and extensive learning capabilities
Memristor-based computing is able to reduce the energy required for a task by allowing the learning to occur on-chip with no external memory source.
Several studies have investigated memristors, but they still used additional external processors, the team said in their paper.
The researchers produced a chip capable of conducting complete on-chip improvement learning, and proposed a learning architecture for it.
They said they were able to demonstrate on-chip learning with multiple tasks, including image classification, motion control, and audio recognition.
How artificial intelligence will redefine the future of work
In one demonstration, the motion control of a model car designed to pursue a laser light was better at finding the light in a dark environment and would lose track of it in a brighter environment.
But after implementing learning, it was able to find the light equally well in both environments.
The accuracy in the bright environment greatly increased, while accuracy in the dark environment did not decrease with the addition of the new ability.
05:03
How does China’s AI stack up against ChatGPT?
With further research on the learning architecture, the team said it believed it could enable on-chip learning that is 75 times more energy efficient than current machines designed for AI processing.
Still, challenges remain in the research and development of such chips – in particular, engineering the chips for large-scale integration.
It could take time for the technology to make it out of the laboratory, Wu Huaqiang and Gao Bin, professors at Tsinghua and leads of the research team, admitted to Chinese-language news site Science Times.
“The study is an important step towards future chips with high energy efficiency and extensive learning capabilities,” the researchers said, adding they hoped their findings “will accelerate the development of future smart edge devices”.
