Dual‐Mode Optoelectronic Neuromorphic Memory for Complex Edge Detection and Recognition

Abstract In‐sensor neuromorphic computing possesses great potential in in‐sensor edge computing for massive image data processing, but today optoelectronic devices cannot meet the requirements on multifunction and high‐precision computing. Here MoS ₂ heterojunction‐based optoelectronic memory device is proposed that can integrate two modes, the dynamic (short‐term memory, STM) and non‐dynamic (long‐term memory, LTM) into one unit to efficiently execute image processing. In the STM mode driven by negative bias, the memory device exhibits huge memory capacity, which enables the device to possess 128 photoconductance states that can supply 7‐bit spatiotemporal feature encoding of reservoir computing. The LTM mode that the heterojunction is positive bias, the memory device with multiple stable photoconductance states can supply physically parallel and one‐step hardware convolution acceleration. Under the photoconductance modulation mechanism, the energy consumption for a single convolutional kernel operation is ≈1.6 fJ. This result demonstrates that the type of optoelectronic memory can achieve energy efficiency advantages as well as enabling accelerated convolutional computing, yielding an accuracy 100% for 26‐letter image classification. This work lays a significant foundation on emerging image sensors.