Ferroelectric NAND for efficient hardware bayesian neural networks.

Area of Science:

• Artificial Intelligence
• Neuromorphic Computing
• Materials Science

Background:

• Conventional neural networks lack uncertainty quantification, limiting reliability with real-world data.
• Bayesian neural networks offer improved robustness by modeling weights as probability distributions.
• Hardware implementations of Bayesian neural networks face challenges in controlling weight distributions.

Purpose of the Study:

• To propose an efficient and scalable 3D ferroelectric NAND-based Bayesian neural network system.
• To enable precise probabilistic weight control for enhanced AI reliability.
• To demonstrate the system's effectiveness in uncertainty estimation and robustness for medical image analysis.

Main Methods:

• Utilized a 3D ferroelectric NAND architecture for probabilistic weight control.
• Employed incremental step pulse programming technology for efficient weight distribution tuning.
• Leveraged page-level programming and device-to-device variations for Gaussian weight distributions.

Main Results:

• Achieved efficient and scalable probabilistic weight control using ferroelectric NAND.
• Demonstrated precise control over weight distributions by modulating programming voltage.
• Successfully implemented uncertainty estimation and enhanced robustness for medical images.

Conclusions:

• The proposed 3D ferroelectric NAND-based Bayesian neural network system offers an efficient solution for uncertainty quantification.
• The system enhances AI robustness and energy efficiency, particularly for applications like medical diagnostics.
• This approach overcomes hardware implementation challenges for Bayesian neural networks.