Complex-valued photonic neural network empowered by FFT-based MZI mesh with balanced detection units

In this work, we demonstrate a scheme of complex-valued photonic neural networks enabled by an FFT-based MZI mesh with balanced detection units. To validate its feasibility, a silicon photonic chip with 4-mode FFT-based MZI mesh is fabricated to perform the classification task on the Vowel dataset, achieving an accuracy of 74.31% on the testing set.