Low-depth optical neural networks

Optical neural network (ONNs) are emerging as attractive proposals for machine-learning applications. However, the stability of ONNs decreases with circuit depth, limiting scalability practical uses. Here we demonstrate how to compress depth scale only logarithmically in terms dimension data, leading an exponential gain noise robustness. Our low-depth (LD)-ONN is based on architecture, called CompuTing Of dot-Product UnitS (OCTOPUS), which can also be applied individually a linear perceptron solving classification problems. We present both numerical and theoretical evidence showing that LD-ONN exhibit significant improvement robustness, compared previous ONN singular-value decomposition.