Probabilistic Weight Elements in Bayesian Physical Neural Networks

Shih-Chuan Hung^1*, Sheng-Chung Chen¹, Sheng-Ting Huang¹, Der-Hsien Lien¹

¹Institute of Electronics, National Yang Ming Chiao Tung University, Hsinchu, Taiwan

* Presenter:Shih-Chuan Hung, email:schung.ee13@nycu.edu.tw

Bayesian probabilistic weights offer a promising route toward efficient storage and computation in neural networks. In this work, we propose compact analog physical elements that enable Bayesian weights by leveraging the current characteristics of integrated p- and n-type transistors within the framework of physical neural networks (PNNs). We demonstrate that the parameters of a weight distribution, including mean and variance, can be directly mapped onto the physical properties of integrated devices through an in-situ tuning mechanism. By modulating the applied voltage, the shape and position of the Bayesian distribution can be precisely controlled, functioning as a tunable probabilistic weight element. This physical-level modulation proposes a simple logic element into a dense, programmable probabilistic weight cell. By encoding distributional parameters directly into device physics, it offers a pathway toward low-power, area-efficient hardware for Bayesian probabilistic inference.

Keywords: Bayesian Neural Networks, Physical Neural Networks, Transistor Characteristics