The brain’s selective visual attention mechanism(SVAM)enables robust visual recognition in noisy environment through diverse neural action potential peaks acting as filters.Spiking neural networks(SNNs)mimic this par...The brain’s selective visual attention mechanism(SVAM)enables robust visual recognition in noisy environment through diverse neural action potential peaks acting as filters.Spiking neural networks(SNNs)mimic this paradigm but limited noise immunity and high write current density hinder brain-like efficiency.Hardware implementing SVAM necessitates spiking spintronic devices with noise-resistant and low operation current densities;such devices remain unreported.Here,we report an orbit-torque(OT)actuated ferromagnetic spiking synapse and neuron featuring a tunable peak action potential.These are more akin to the biological neurons with varying sensitivities to external sensory stimuli,thereby augmenting the perception aptitude of the system in complex surroundings.Capitalizing on the high-efficient OT,the ferromagnetic device demands a write current density of 5×10^(6) A/cm^(2),which is an order of magnitude lower than other spiking devices actuated by spin-orbit torque.Leveraging these neuromorphic devices,an all-spin SNN with low current density and tunable action potential peak has been fabricated,successfully mimicking the SVAM.In complex noise environment,the SNN achieves 92%on Cifar-10 and 95%on MNIST dataset,surpassing state-of-the-art spin-based SNNs by 5%.Our work provides a promising avenue for exploring the SVAM-inspired spiking neuromorphic devices,enhancing the bionic performance of the SNNs.展开更多
基金supported by the National Natural Science Foundation of China(12304160,12304161,62172155,U22A2027,62274183,and 62301595)the Research Foundation from National University of Defense Technology(ZK24-18,23-ZZCX-ZZGC-01-02,and 22-ZZCX-046-02)。
文摘The brain’s selective visual attention mechanism(SVAM)enables robust visual recognition in noisy environment through diverse neural action potential peaks acting as filters.Spiking neural networks(SNNs)mimic this paradigm but limited noise immunity and high write current density hinder brain-like efficiency.Hardware implementing SVAM necessitates spiking spintronic devices with noise-resistant and low operation current densities;such devices remain unreported.Here,we report an orbit-torque(OT)actuated ferromagnetic spiking synapse and neuron featuring a tunable peak action potential.These are more akin to the biological neurons with varying sensitivities to external sensory stimuli,thereby augmenting the perception aptitude of the system in complex surroundings.Capitalizing on the high-efficient OT,the ferromagnetic device demands a write current density of 5×10^(6) A/cm^(2),which is an order of magnitude lower than other spiking devices actuated by spin-orbit torque.Leveraging these neuromorphic devices,an all-spin SNN with low current density and tunable action potential peak has been fabricated,successfully mimicking the SVAM.In complex noise environment,the SNN achieves 92%on Cifar-10 and 95%on MNIST dataset,surpassing state-of-the-art spin-based SNNs by 5%.Our work provides a promising avenue for exploring the SVAM-inspired spiking neuromorphic devices,enhancing the bionic performance of the SNNs.
