2 cos

A.1 Mach-ZehnderInterferometers(MZIs) A basic coherent optical component used in this work is an MZI. To optimize the MZI meshes, a straightforward idea is to use first-order methods to optimize all rotationsphases ΦU,ΦV,andΦΣ. A similar sampling matrixPx is applied to input features. According to the experience from the field of structured/subspace neural networks, e.g., block-circulant neural nets, the block size is typically set to a number around 8. Here we add newresults onL2ight-SL (αW=αC=0.6, In conclusion, we recommend using multiple interconnected 9 9 PTCs for parallel computing, since this choice of 9 9 block balances both systematic performance, hardware complexity,robustness,andon-chiptrainability.