A Fourier Optoelectronic Synapse with Single-Wavelength Modulation.

Abstract

Analyzing spatial frequency domain features is essential for capturing diverse in-depth features of targets, thereby enhancing the adaptability to unstructured environments for embodied intelligence. Digital approaches suffer from limited efficiency due to frequent data transfer, while neuromorphic ones lack dedicated frequency-domain processing hardware. Here, we report a Fourier neuromorphic visual (FIVE) system integrating a Fourier optical system and Fourier optoelectronic synapses (FOSs). The FIVE system extracts frequency-domain cues optically with negligible time latency and computational energy consumption, and these cues are then filtered based on nonlinearity to the light intensity of FOSs. Furthermore, such a FOS device exhibits multilevel memory tunability by light intensity of single-wavelength. This property enables the implementation of multilayer perceptron for further classification of frequency domain features. The FIVE system achieves a high accuracy of ∼90% in the image noise classification task and outperforms convolutional neural network (CNN)-based approaches by orders of magnitude in parameter count.