The University of Science and Technology of China has developed a brain-like synaptic prototype device that is comparable to the energy efficiency of the american brain

IT Home News on February 9, according to the official website of the University of Science and Technology of China, The team of Professor Li Xiaoguang of the University of Science and Technology of China has made important progress in high-performance synaptic prototype devices. Based on the previous research, based on the design of ferroelectric domain morphology and flip kinetics, the team realized a brain-like synapse device with nonvolatile continuous regulation of electrical conductivity under sub-nanosecond electrical pulses in the ferroren tunnel junction, which can be used to build artificial neural network brain-like computing system, which was recently published in the journal Nat. Commun.）。

Brain-like artificial intelligence technology represented by neural networks is profoundly affecting human society. However, the hardware systems that currently run neural network computations are still based on traditional silicon-based operators and memory, and their energy efficiency is much lower than that of the human brain. The research and development of brain-like devices with neuromorphic simulation functions, such as the core device of neural network hardware systems - electronic synapses, is one of the important ways to further promote the development of artificial intelligence. In order to perform complex artificial intelligence tasks, neural network hardware systems put forward many demanding requirements for electronic synaptic devices, such as: the number of nonvolatile conductive states (used to simulate the continuous tunableness of brain synapses) is greater than 100, the nonlinearity is less than 1 (good linearity helps to accurately regulate conductivity), the switching ratio is greater than 100, the flip durability is greater than 109 times, and the cycle randomness is less than 3%. However, the reported synapse-like brain devices do not fully meet the above specification requirements.

Professor Li Xiaoguang's team prepared high-quality ferroelectric tunnel junctions, and through the design of PZT (piezoelectric ceramic driver) ultra-thin thickness and orientation, smaller ferroelectric domains and more continuous flipping kinetic behavior were obtained, and the richer ferroelectric multi-domain metastable is conducive to the controllable regulation of polymorphisms in brain-like synaptic devices. The device exhibits excellent overall performance: its 8-bit linear conductance regulation and high durability meet the core performance specification requirements of brain-like synapse devices. The neural network built based on the performance simulation of this device has a high image recognition rate, and even if pretzel noise or Gaussian noise is introduced into the picture, its accuracy in identifying the picture is still greater than 85%. In addition, the device features ultra-fast sub-nanosecond operating speeds and consumes energy down to the fly focal length. The researchers have calculated that the neural network computing system built by the ferroelectric tunnel structure may achieve excellent energy efficiency equivalent to that of the human brain, while the energy consumption of a single pulse of the human brain neuron synapses is about 10 fly focal points. The human brain synapse response speed is about a millisecond, and its response speed is also 6 orders of magnitude faster than the human brain synapse, which is comparable to the energy efficiency performance of the human brain synapse.

IT House understands that the above results show the important potential of ferroelectric tunnel junctions in building future high-performance brain-like ARTIFICIAL intelligence computing hardware systems.