Electroluminescent perovskite QD–based neural networks for energy-efficient and accelerate multitasking learning

Abstract

The ability of multitasking (MT) learning in neuro-inspired artificial intelligence (AI) systems offers promise for energy-efficient deployment in robotics, health care, and autonomous vehicles. Here, an MT learning framework is established using a dual-output electroluminescent synaptic device array based on a mixed-dimensional stacked configuration with Cs1−xFAxPbBr3 (0.00 ≤ x ≤ 0.15) quantum dots. The device concurrently processes postsynaptic current (PSC) and postsynaptic electroluminescence (PSEL) signals, demonstrating stable and adjustable long-term plasticity with ~1000 individual states, along with spike rate-dependent plasticity and paired-pulse facilitation. By synthesizing the update behavior of both PSC and PSEL pathways, the MT framework simultaneously executes classification-regression and classification-image reconstruction tasks. This approach achieves computational speed improvements of up to 47.09 and 29.17% while reducing energy consumption by up to 8.2- and 32.4-fold compared to a combined single-tasking framework and graphics processing unit–based hardware accelerators, respectively. This innovative method emphasizes the potential of dual-output electroluminescent artificial synapse for MT learning applications.