AI Breakthrough: Synaptic Device Array Achieves Human-Like Visual Processing

Scientists have unveiled a compact 28×28 synaptic device array that mimics the human visual system, integrating sensing, memory, and processing for real-time image recognition. This innovation, built on MoS2 and gold nanoparticles, could revolutionize artificial vision and neuromorphic computing.

a Schematic diagram of the human visual perception system, including neural networks in the human lens, hemispherical retina, optic nerve, and visual cortex. b Optical image of the 28 × 28 device array. c 3D schematic diagram of the artificial synaptic device structure based on the MoS2 floating-gate device

In a development for artificial intelligence, researchers have unveiled a 28×28 synaptic device array that promises to revolutionize artificial visual systems. This innovative array, measuring a compact 0.7×0.7 cm², integrates the capabilities of sensing, memory, and processing to mimic the intricate functions of the human visual system. Utilizing wafer-scale monolayer molybdenum disulfide (MoS₂) and gold nanoparticles for enhanced electron capture, the array exhibits remarkable coordination between optical and electrical components. It is capable of both writing and erasing images. It has achieved a stunning 96.5% accuracy in digit recognition, marking a significant leap forward in the development of large-scale neuromorphic systems.

The human visual system processes complex visual data efficiently through an interconnected network that allows for parallel processing. However, current artificial vision systems face numerous challenges, including circuit complexity, high power consumption, and difficulties in miniaturization. These issues arise from the separation between signal devices and processing units, hindering the ability to process visual information in parallel. Despite previous attempts, simulating a complete, biologically inspired vision system with a single device has remained elusive, driving the need for more integrated, efficient solutions capable of real-time processing.

A study published in the journal Microsystems & Nanoengineering introduced a game-changing solution to these longstanding challenges. Led by a Beijing Institute of Technology team, the study presents a 28×28 synaptic device array fabricated using MoS₂ floating-gate field-effect transistors. This device replicates the neural networks of the human visual system and delivers exceptional optoelectronic synaptic performance, setting the stage for more efficient and integrated artificial visual systems.

a Implemented 28 × 28 synaptic arrays as the trainable memory. The image of the emblem of Beijing Institute of Technology was input into the memory array using 50 spikes of laser irradiation with P = 20 mW/cm², t = 0.2 s and ΔT = 0.2 s. b Applying different voltage spikes to write and erase the letter “B”, “I”, and “T”. c Schematic diagram of a simulated CNNs with an input layer that captures a 28 × 28-pixel image, a convolution layer with a 3 × 3 kernel, an average pooling layer followed by a fully connected layer and output layer. d The recognition accuracy of the visual signal stimuli for different training iterations under different light illumination

The Beijing Institute of Technology research team has successfully designed a 28×28 array where each device mimics the synaptic plasticity found in the human visual system. Using MoS₂ floating-gate transistors combined with gold nanoparticles as electron capture layers, they achieved stable and uniform optoelectronic performance, capable of simulating key synaptic behaviors like excitatory postsynaptic current (EPSC) and paired-pulse facilitation (PPF). The array demonstrated an on/off ratio of around 10^6 and an average mobility of 8 cm²V^-1s^-1. Notably, the array was able to store and process image data, such as the emblem of the Beijing Institute of Technology, showcasing its potential for optical data processing. Furthermore, the ability to adjust light intensity and fine-tune recognition accuracy provides a new method for optimizing the system's performance in varying lighting conditions.

Jing Zhao, the corresponding author of the study, emphasized the importance of these findings: "Our results offer a viable pathway toward large-scale integrated artificial visual neuromorphic systems. The performance of the MoS₂-based synaptic array represents a major step toward practical applications, from device-level simulations to system-wide integration."

The advances in artificial synaptic neural networks present numerous advantages, including high integration, stable uniformity, and powerful parallel processing capabilities. These attributes could transform the performance of computational systems. The network's ability to simultaneously process optoelectronic signals and adjust synaptic weights via light signals has already demonstrated impressive results in handwritten digit recognition, with an accuracy of 96.5%. This breakthrough opens up exciting possibilities for the future of deep learning and artificial vision, potentially ushering in smarter, more efficient systems in the near future.