Ferroelectric NAND for efficient hardware bayesian neural networks

Abstract

The rapid advancement of artificial intelligence has enabled breakthroughs in diverse fields, including autonomous systems and medical diagnostics. However, conventional deterministic neural networks struggle to capture uncertainty, limiting their reliability when handling real-world data, which are often noisy, imbalanced, or scarce. Bayesian neural networks address this limitation by representing weights as probabilistic distributions, allowing for natural uncertainty quantification and improved robustness. Despite their advantages, hardware-based implementations face significant challenges due to the difficulty of independently tuning both the mean and variance of weight distributions. Herein, we propose a 3D ferroelectric NAND-based Bayesian neural network system that leverages incremental step pulse programming technology to achieve efficient and scalable probabilistic weight control. The page-level programming capabilities and intrinsic device-to-device variations enable gaussian weight distributions in a single programming step, without structural modifications. By modulating the incremental step pulse programming voltage step, we achieve precise weight distribution control. The proposed system demonstrates successful uncertainty estimation, enhanced energy efficiency, and robustness to external noise for medical images.