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

Area of Science:

• Materials Science
• Computer Engineering
• Artificial Intelligence

Background:

• Bayesian neural networks (BNNs) are crucial for trustworthy edge intelligence due to their uncertainty quantification.
• Existing hardware implementations face challenges: digital methods have high latency, and memristors struggle with coupled mean and variance control.

Purpose of the Study:

• To overcome the trade-off between synaptic weight mean and variance in memristive devices for BNN hardware.
• To enable hardware-native orthogonal control of synaptic weight statistics.

Main Methods:

• Development of a coupled dual-channel memristor (CDCM) using an ion gel/ZnO heterostructure.
• Utilizing vertical ion-gating to create two tunable memristive channels.
• Defining synaptic weight as differential conductance for independent control of mean (μ) and standard deviation (σ).

Main Results:

• Demonstrated orthogonal control over synaptic weight mean and standard deviation, enabling decoupled Gaussian weight synthesis.
• Achieved 79.08% accuracy in a hardware-calibrated BNN for multimodal human activity recognition.
• Successfully detected out-of-distribution anomalies, indicating reliable uncertainty estimation.

Conclusions:

• The CDCM device offers a scalable, physics-driven solution for energy-efficient probabilistic computing.
• This approach breaks the fundamental trade-off in memristive BNNs, paving the way for trustworthy edge AI.
• The technology enables precise synthesis of Gaussian weights for enhanced BNN performance and reliability.