Random deep photonic processor for high-throughput optical inference with cascaded MZI arrays

Abstract

We demonstrate a random deep photonic processor built from 10 layers of cascaded Mach-Zehnder interferometers with random-width tapers on a compact 29 mu m x 840 mu m silicon footprint. Operating with 8 input/output ports and wavelength-dependent phase shifts, the system performs nonlinear mappings via photodetection and a single-layer electronic readout. On a medical image classification task (BreastMNIST), the random deep photonic processor achieves 84.6% (phase encoding) and 82.0% (amplitude encoding) accuracy with only 1202 parameters, outperforming the 68.1% baseline achieved without the photonic processor. Similarly, on a gesture recognition task (Libras Movement), it reaches 88.9% and 80.6% using phase and amplitude encodings respectively, surpassing the 65.3% baseline using just 9015 parameters and enabling fast, scalable inference.