Screen-Gated Organic Transistors Enabling Tunable Sigmoid and Gaussian Activation Functions
발표자
한영민 (한양대학교)
연구책임자
유호천 (한양대학교)
공동저자
한영민 (한양대학교), 유호천 (한양대학교)
초록
내용
Tunable analog activation functions are critical for energy-efficient AI hardware. Here, we demonstrate sigmoid-like activation transistors (SA-transistor) and Gaussian-like activation transistors (GA-transistor) based on a screen-gate architecture, enabling continuous and precise control of slope, amplitude, and width at the device level. The SA-transistor improves lung MRI classification accuracy from 77% to 84%, while the GA-transistor enhances time-series forecasting performance (R²: 0.82 → 0.93). Furthermore, a hardware multilayer perceptron (MLP) incorporating these devices achieves 96.7% accuracy on the IRIS dataset, validating their effectiveness for low-power neuromorphic inference with reduced circuit complexity.