R&D: Ferroelectric NAND for Efficient Hardware Bayesian Neural Networks

Nature Communications has published an article written by Minsuk Song, Department of Nanoscale Semiconductor Engineering, Hanyang University, Seoul, 04763, Republic of Korea, Ryun-Han Koo, Department of Electrical and Computer Engineering and Inter-university Semiconductor Research Center, Seoul National University, Seoul, 08826, Republic of Korea, Jangsaeng Kim, Department of Electronic Engineering, Sogang University, Seoul, 04107, Republic of Korea, and Department of System Semiconductor Engineering, Sogang University, Seoul, 04107, Republic of Korea, Chang-Hyeon Han, Jiyong Yim, Department of Electrical Engineering, Hanyang University, Seoul, 04763, Republic of Korea, Jonghyun Ko, Department of Electrical and Computer Engineering and Inter-university Semiconductor Research Center, Seoul National University, Seoul, 08826, Republic of Korea, Sijung Yoo, Duk-hyun Choe, Sangwook Kim, Thin Film Technical Unit, Device Research Center (DRC), Samsung Advanced Institute of Technology, Suwon, 16678, Republic of Korea, Wonjun Shin, Department of Semiconductor Convergence Engineering, Sungkyunkwan University, Suwon, 16419, Republic of Korea, and Daewoong Kwon, Department of Nanoscale Semiconductor Engineering, Hanyang University, Seoul, 04763, Republic of Korea.

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.“