计算机科学
尖峰神经网络
编码器
Spike(软件开发)
人工神经网络
算法
泊松分布
人工智能
模式识别(心理学)
提炼听神经的脉冲
电子工程
编码(内存)
自编码
随机过程
神经编码
作者
Jiaying Chen,Lirou Su,Yongsheng Huang,Beibei Huang,Junfeng Zhang,Mingkun Xu,Yishu Zhang,Shuai Zhong
标识
DOI:10.1109/led.2026.3750454
摘要
Poisson rate coding converts analog inputs into stochastic spikes for low-power spiking neural networks (SNNs). However, a compact genuine Poisson encoder is yet to be reported. In this letter, we present a Poisson spike encoder implemented on a volatile Pt/CeO2/V threshold-switching memristor with high switching uniformity. Intrinsic thermal randomness of the device is exploited to generate spike trains following Poisson-like statistics, which is validated by exponential interspike interval distribution, Fano factor near unity, and near-zero autocorrelation function. Benchmarked on a FashionMNIST spiking multilayer perceptron, our memristive Poisson encoder matches the clean classification accuracy of non-Poisson stochastic coding under equal firing rates, and exhibits prominent robustness against spike dropout, insertion, and flip noise. This work delivers a validated, noise-resilient memristive Poisson spike hardware primitive to facilitate scalable energy-efficient on-chip SNN implementations.
科研通智能强力驱动
Strongly Powered by AbleSci AI