Self-Powered Organic Optoelectronic Synapses with Binarized Weights for Noise-Suppressed Visual Perception and High-Robustness Inference Self-Powered Organic Optoelectronic Synapses with Binarized Weights for Noise-Suppressed Visual Perception and High-Robustness Inference AITranslate

Basic Information:

DOI：https://doi.org/10.1021/acsaelm.3c00698

Chinese Library Classification Number:

Citation Information:

Neuromorphic optoelectrical synapses have shown great potential in edge artificial intelligence (AI) for energy-efficient sensory computing. However, environmental noise and device nonidealities pose immense challenges to the inference accuracy and robustness of neuromorphic devices and networks. Here, inspired by the hierarchical biological vision system, self-powered organic optoelectronic synaptic devices, which benefit from a simple asymmetric-electrode structure, are demonstrated. The multifunctional synapses can be configured as self-powered, denoising vision sensors for noise-suppressed visual perception and as photostimulated synapses for implementing a quantized neural network (QNN) by the differential synaptic structure. An ex situ training method by weight mapping from trained real-valued weights to binarized weights has been proposed to increase the robustness of organic synapses. The binary neural network (BNN) with binarized weights mitigates the impact of the synaptic nonlinearity with 97.0% inference accuracy and exhibits only 1.7% accuracy reduction with large synaptic weight variations due to device nonidealities compared with the accuracy using backpropagation algorithms. Therefore, this self-powered organic synapse enables simultaneous optical sensing and processing for more robust and energy-efficient neuromorphic systems.