A Microwave Photonic Neural Network in the Frequency Synthetic Dimension Using Multi-tone Single-Sideband Modulation in a Fiber Loop

Photonic synthetic dimension, an emerging concept in the field of photonics, introduces new dimensions within photonic systems, enabling more advanced and precise manipulation of high-speed optical signals. In this paper, we propose and experimentally demonstrate a novel microwave photonic neural network in the frequency synthetic dimension. By using a dual-parallel Mach-Zehnder modulator (DP-MZM) in an optical fiber loop, multi-tone carrier-suppressed single sideband (CS-SSB) modulation on multiple optical carriers is implemented which results in the dynamic coupling between different frequency modes in the loop, thus synthesizing additional frequency dimensions. In the synthetic dimension, the matrix-vector multiplication (MVM) between the neurons and the weights in a neural network is implemented by controlling the coupling coefficients between the optical carriers and the modulated sidebands, thus generating sidebands with specific intensities corresponding to the MVM results. By using a programmable optical filter after the DP-MZM to introduce attenuations and control the powers of the generated sidebands, the desired neuron biases can be also achieved. The multi-layer structure in a deep feedforward neural network (FNN) can be realized by refreshing the multi-tone radio frequency (RF) signals applied to the DP-MZM and the programmable filtering during each round trip in the optical loop. The approach is evaluated experimentally. A 3-2-8 three-layer FNN having 13 neurons is implemented to successfully tackle the task of 3-bit binary number recognition at a computational speed of 55×106 MAC/s (multiply-accumulate operations per second). An average prediction accuracy rate of 95.93% is demonstrated.