Active Hyperparameters Training in Adaptive Physical Neural Networks

Sheng-Chung Chen^1*, Shih-Chuan Hung², Sheng-Ting Huang², Der-Hsien Lien²

¹Institute of Pioneer Semiconductor Innovation, National Yang Ming Chiao Tung University, Hsinchu, Taiwan
²Institute of Electronic, National Yang Ming Chiao Tung University, Hsinchu, Taiwan

* Presenter:Sheng-Chung Chen, email:sheng8705@gmail.com

We demonstrate an electronics-based Physical Neural Network (PNN) that harnesses measured nonlinear I–V characteristics of semiconductor devices as physical activation functions for analog inference and learning. The system operates under both Backpropagation (PAT) and Forward-Forward (FF) training paradigms, with Optuna-assisted hyperparameter optimization and adaptive training thresholds to minimize redundant hardware-in-the-loop iterations. Experimentally, the device-level nonlinearity—originating from transistor transfer curves and voltage-dependent conductance—is embedded into the computational graph, enabling in-situ activation without digital emulation. Compact architectures incorporating pooling and slim fully connected layers reduce parameter overhead while preserving recognition accuracy, achieving >90% on MNIST and ≈80% on Fashion-MNIST. This synergy between experimental device physics, algorithmic optimization, and efficient hardware interfacing establishes a practical framework for energy-aware neuromorphic computing, and points toward scalable extensions across diverse device platforms and locally trained analog networks.

Keywords: physical neural networks, hyperparameter, deep learning