Physical Implementation of Reinforcement Learning with a Dual-Input Synaptic Transistor
발표자
유지현 (연세대학교)
연구책임자
조정호 (연세대학교)
공동저자
유지현 (연세대학교), 조정호 (연세대학교), 노동규 (연세대학교)
초록
내용
The growing computational cost of modern artificial intelligence (AI) workloads motivates hardware approaches that reduce energy consumption and circuit overhead. We report a dual-input synaptic transistor (DIST) that realizes in-device analog current summation by combining independent voltage and light driven conductance updates. A photoactive Au(I) complex (Au(DippPZI)(DPA)) undergoes dipole inversion under illumination, while -OH trap sites in crosslinked PVP enable gate voltage modulation of an IGZO channel. Optical and electrical pulses add linearly with little interference, enabling direct mapping to value-advantage integration in a dueling deep Q-network. In a Python racing task, a DIST-based agent learns stably and matches software-level performance, highlighting a scalable route toward device-level acceleration of reinforcement learning inference and training.