Reward-modulated spike-timing-dependent plasticity(R-STDP)is a promising biomimetic learning rule in neuromorphic intelligent systems for implementing tasks in variable environments.Nevertheless,realizing R-STDP in a ...Reward-modulated spike-timing-dependent plasticity(R-STDP)is a promising biomimetic learning rule in neuromorphic intelligent systems for implementing tasks in variable environments.Nevertheless,realizing R-STDP in a single synaptic device for building compact and energy-efficient neuromorphic systems remains challenging.Here,we report a two-dimensional ferroelectric memtransistor to emulate the RSTDP learning rule by effectively reconfiguring the STDP and anti-STDP.The thermionic emission and tunneling behavior of charges at the ferroelectric interface can be regulated via vertical electric field in a multi-terminal manner,allowing for controllable polarization reversal of synaptic plasticity and transition between STDP and anti-STDP.This enables faithful realization of the R-STDP feature in a single device with energy consumption of~1.3 nJ(the lowest known to date),approximately 10^(6) times lower than that of its complementary metal-oxide-semiconductor(CMos)counterpart.By leveraging the synaptic characteristics in the hardware device,we construct spiking neural networks(SNNs)trained with R-STDP to perform robotic recognition and tracking tasks.The SNN achieves 95.1% accuracy on the MNIST dataset using only 8000 parameters,and faster convergence speed requiring only one data batch with 100% inference in the few-shot learning task.Moreover,a robotic arm motion control system configured with R-STDP exhibits 85.5% success rate in tracking both the static and moving targets,illustrating its outstanding adaptability to the dynamic environments.This work provides a potential hardware building block to support compact neuromorphic systems for the application of interactive artificialintelligenceagents.展开更多
Memtransistors in which the source-drain channel conductance can be nonvolatilely manipulated through the gate signals have emerged as promising components for implementing neuromorphic computing.On the other side,it ...Memtransistors in which the source-drain channel conductance can be nonvolatilely manipulated through the gate signals have emerged as promising components for implementing neuromorphic computing.On the other side,it is known that the complementary metal-oxide-semiconductor(CMOS)field effect transistors have played the fundamental role in the modern integrated circuit technology.Therefore,will complementary memtransistors(CMT)also play such a role in the future neuromorphic circuits and chips?In this review,various types of materials and physical mechanisms for constructing CMT(how)are inspected with their merits and need-to-address challenges discussed.Then the unique properties(what)and poten-tial applications of CMT in different learning algorithms/scenarios of spiking neural networks(why)are reviewed,including super-vised rule,reinforcement one,dynamic vision with in-sensor computing,etc.Through exploiting the complementary structure-related novel functions,significant reduction of hardware consuming,enhancement of energy/efficiency ratio and other advan-tages have been gained,illustrating the alluring prospect of design technology co-optimization(DTCO)of CMT towards neuro-morphic computing.展开更多
基金supported by the National Key Research and Development Program of China(2023YFA1407800 and 2024YFA1410700)the National Natural Science Foundation of China(62374038,62104041,62204051,and 62374040)+1 种基金the Natural Science Foundation of Shanghai(22ZR1405700)the Shanghai Rising-Star Program(22QA1401000 and 24QA2700400).
文摘Reward-modulated spike-timing-dependent plasticity(R-STDP)is a promising biomimetic learning rule in neuromorphic intelligent systems for implementing tasks in variable environments.Nevertheless,realizing R-STDP in a single synaptic device for building compact and energy-efficient neuromorphic systems remains challenging.Here,we report a two-dimensional ferroelectric memtransistor to emulate the RSTDP learning rule by effectively reconfiguring the STDP and anti-STDP.The thermionic emission and tunneling behavior of charges at the ferroelectric interface can be regulated via vertical electric field in a multi-terminal manner,allowing for controllable polarization reversal of synaptic plasticity and transition between STDP and anti-STDP.This enables faithful realization of the R-STDP feature in a single device with energy consumption of~1.3 nJ(the lowest known to date),approximately 10^(6) times lower than that of its complementary metal-oxide-semiconductor(CMos)counterpart.By leveraging the synaptic characteristics in the hardware device,we construct spiking neural networks(SNNs)trained with R-STDP to perform robotic recognition and tracking tasks.The SNN achieves 95.1% accuracy on the MNIST dataset using only 8000 parameters,and faster convergence speed requiring only one data batch with 100% inference in the few-shot learning task.Moreover,a robotic arm motion control system configured with R-STDP exhibits 85.5% success rate in tracking both the static and moving targets,illustrating its outstanding adaptability to the dynamic environments.This work provides a potential hardware building block to support compact neuromorphic systems for the application of interactive artificialintelligenceagents.
基金supported by the National Key Research and Development Program of China(No.2023YFB4502200)Natural Science Foundation of China(Nos.92164204 and 62374063)the Science and Technology Major Project of Hubei Province(No.2022AEA001).
文摘Memtransistors in which the source-drain channel conductance can be nonvolatilely manipulated through the gate signals have emerged as promising components for implementing neuromorphic computing.On the other side,it is known that the complementary metal-oxide-semiconductor(CMOS)field effect transistors have played the fundamental role in the modern integrated circuit technology.Therefore,will complementary memtransistors(CMT)also play such a role in the future neuromorphic circuits and chips?In this review,various types of materials and physical mechanisms for constructing CMT(how)are inspected with their merits and need-to-address challenges discussed.Then the unique properties(what)and poten-tial applications of CMT in different learning algorithms/scenarios of spiking neural networks(why)are reviewed,including super-vised rule,reinforcement one,dynamic vision with in-sensor computing,etc.Through exploiting the complementary structure-related novel functions,significant reduction of hardware consuming,enhancement of energy/efficiency ratio and other advan-tages have been gained,illustrating the alluring prospect of design technology co-optimization(DTCO)of CMT towards neuro-morphic computing.
