A Microwave Photonic Neural Network Using Recurrent Multi-Frequency Single-Sideband Modulation in an Optical Fiber Loop

Taking advantage of the high speed and wide bandwidth of photonics, microwave photonics (MWP) is well-suited for high-speed signal processing. In this paper, we propose and experimentally demonstrate a novel microwave photonic neural network based on recurrent carrier-suppressed single sideband (CS-SSB) modulation implemented using an amplified optical fiber loop incorporating a dual-parallel Mach-Zehnder modulator (DP-MZM). We show that the matrix-vector multiplication (MVM) between the neurons and weights can be implemented by multi-tone CS-SSB modulation on multiple optical carriers. By using a programmable optical filter to introduce attenuations and control the powers of the sidebands, it is also possible to achieve the desired neuron biases. Multiple layers in a deep neural network are realized by recurrent CS-SSB modulation and programmable filtering in the optical fiber loop. The approach is evaluated experimentally. A 2-2-4 feedforward neural network at a computational speed of $30\times 10^{6}\text{MAC}/\mathrm{s}$ (multiply-accumulate operations per second) is implemented to successfully tackle the task of 2-bit binary number recognition.