Deep spike heterostructure photonic neuro-transistors for effective neuromorphic computation and low energy consumption

초록

Photonic neuro-transistors that utilize the persistent photoconductivity behavior by heterojunction structures has been previously used for light-driven synaptic performance. By varying the anion-to-cation ratio of the light-absorbing layer and the semiconductor, the photonic transistors were able to precisely mediate the degree of energy band-bending at the heterointerface, leading to efficient accumulation of photo-generated charge carriers and the emulation of biological synaptic functions. The photonic neuro-transistor with the optimized structure achieved a high ratio of effective conductance-level states for both long-term potentiation and long-term depression, along with linear conductance change and less energy consumption compared to previously reported optoelectronic neuromorphic devices. Deep spike synaptic transistor with deep level potential well enables linear conductance change with low non-linearity values (NL) of 1.1 during long-term potentiation (LTP) behaviors along with low energy consumption (45.04 pJ). We also demonstrate the feasibility of large-area optoelectronic neuromorphic arrays and explore training and inference tasks simulation using Modified National Institute of Standards and Technology (MNIST) data set, achieving high recognition accuracy of 85.96% . This study shows potential for the development of energy-efficient neuromorphic computing systems for artificial intelligence applications. © 2023 SPIE.

키워드