Programmable Surface Plasmonic Neural Networks

Abstract Recently, some new forms of artificial intelligence computing hardware and chips have been presented. However, most of them have difficulties to simultaneously achieve advantages of light-speed computing, programmable weight matrix, and programmable nonlinear activation functions. Here, we propose a programmable surface plasmonic neural network (SPNN) with programmable weights and activation functions based on a spoof surface plasmon polariton (SSPP) platform, which can perform intelligent functions and sense electromagnetic (EM) waves at nearly light speed. We demonstrate a parallel coupling SSPP structure loaded with varactors to introduce four paths with tunable transmitting parameters. On this multi-port architecture, we further establish a real-time control and feedback method to enable arbitrarily designable activation functions under a detecting feedback loop. Experimental results show that a four-in and four-out fully-connected super-neuron can fulfill independently adjustable weights and programmable activation functions, where each input can be sensed for arbitrarily programming. To comprehensively show the above capabilities, we design and demonstrate experimentally a wireless communication system based on the SPNN for image decoding and recovery. We further illustrate a partially connected SPNN using the super-neurons with a high prediction accuracy. The proposed concept paves a new way for artificial intelligence devices, stimulating the fascinating fields like large-scale EM computing and communication systems in the future.