Scholarlink

<p>Bayesian neural networks (BNNs) enable trustworthy edge intelligence by quantifying predictive uncertainty. However, hardware BNN implementations face a bottleneck: digital approaches suffer from high latency, while memristors are limited by the intrinsic coupling between their conductance state (mean) and stochastic noise (variance). Here we report a coupled dual-channel memristor (CDCM) based on an ion gel/ZnO heterostructure that breaks this fundamental trade-off. Utilizing vertical ion-gating to establish two tunable memristive channels, our device defines synaptic weight as the differential conductance between the two channels. This architecture enables the hardware-native orthogonal control over the synaptic weight mean (<i>μ</i>) and standard deviation (<i>σ</i>), allowing for the precise synthesis of decoupled Gaussian weights. We validate our approach with a hardware-calibrated BNN for multimodal human activity recognition, achieving 79.08% accuracy while reliably detecting unseen activities as out-of-distribution anomalies. This work provides a scalable, physics-driven paradigm for energy-efficient, inherently trustworthy probabilistic computing.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Coupled dual-channel memristors for hardware-native trustworthy Bayesian intelligence

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月26日

查看全文
  • Guowei Liu,
  • Jiaqi Ding,
  • Ziyu Wan,
  • Jiakun Xue,
  • Yu Zhang,
  • Zhenqi Gong,
  • Chengyuan Peng,
  • Wenpei Shi,
  • Jiale Su,
  • Jiangnan Xia,
  • Huipeng Chen,
  • Lang Jiang,
  • Lei Liao,
  • Yuanyuan Hu

摘要

Bayesian neural networks (BNNs) enable trustworthy edge intelligence by quantifying predictive uncertainty. However, hardware BNN implementations face a bottleneck: digital approaches suffer from high latency, while memristors are limited by the intrinsic coupling between their conductance state (mean) and stochastic noise (variance). Here we report a coupled dual-channel memristor (CDCM) based on an ion gel/ZnO heterostructure that breaks this fundamental trade-off. Utilizing vertical ion-gating to establish two tunable memristive channels, our device defines synaptic weight as the differential conductance between the two channels. This architecture enables the hardware-native orthogonal control over the synaptic weight mean (μ) and standard deviation (σ), allowing for the precise synthesis of decoupled Gaussian weights. We validate our approach with a hardware-calibrated BNN for multimodal human activity recognition, achieving 79.08% accuracy while reliably detecting unseen activities as out-of-distribution anomalies. This work provides a scalable, physics-driven paradigm for energy-efficient, inherently trustworthy probabilistic computing.