Brain-inspired synaptic transistors for in-situ spiking reinforcement learning with eligibility trace.

Abstract

Brain-inspired reinforcement learning is pivotal for artificial general intelligence, yet current artificial neural network-based hardware lacks critical biological mechanisms like third-terminal modulated eligibility traces and dynamic reward signaling. Emerging materials address these challenges by efficiently mimicking complex reinforcement learning dynamics. Here, we demonstrate a brain-inspired spiking neural network-based reinforcement learning computing architecture using α-In₂Se₃ ferroelectric semiconductor field-effect transistor. By leveraging the intrinsic in-plane and out-of-plane polarization coupling of α-In₂Se₃, the multi-terminal conductance modulation in the device enables reward signal modulation of reinforcement learning. The ferroelectric relaxation is utilized to implement biological eligibility trace decay, thereby enhancing the algorithm's processing capability. autonomous driving tasks are then demonstrated with an RL neural network constructed by the α-In₂Se₃ transistor array, where in-situ reward-based weight updates and eligibility trace decay are performed without any external memory or computing units. Our solution enables a fully functional, energy-efficient, and low-overhead spiking-based reinforcement learning architecture.