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

Neuromorphic optoelectrical synapses have shown greatpotentialin edge artificial intelligence (AI) for energy-efficient sensorycomputing. However, environmental noise and device nonidealitiespose immense challenges to the inference accuracy and robustness ofneuromorphic devices and networks. Here, inspired by the hierarchicalbiological vision system, self-powered organic optoelectronic synapticdevices, which benefit from a simple asymmetric-electrode structure,are demonstrated. The multifunctional synapses can be configured asself-powered, denoising vision sensors for noise-suppressed visualperception and as photostimulated synapses for implementing a quantizedneural network (QNN) by the differential synaptic structure. An exsitu training method by weight mapping from trained real-valued weightsto binarized weights has been proposed to increase the robustnessof organic synapses. The binary neural network (BNN) with binarizedweights mitigates the impact of the synaptic nonlinearity with 97.0%inference accuracy and exhibits only 1.7% accuracy reduction withlarge synaptic weight variations due to device nonidealities comparedwith the accuracy using backpropagation algorithms. Therefore, thisself-powered organic synapse enables simultaneous optical sensingand processing for more robust and energy-efficient neuromorphic systems.