Principles and metrics of photonic learning machines

Photonic systems are emerging as a compelling alternative to electronic computing, offering unmatched parallelism, energy efficiency, and speed. Recent advances demonstrate that nonlinear optical platforms, such as highly nonlinear fibers and semiconductor lasers, can be configured as physical neural networks, leveraging their intrinsic physics for computation. Here, we characterize two photonic platforms: highly nonlinear fiber (HNLF) and vertical-cavity surface-emitting lasers (VCSELs). Using metrics like dimensionality (independent degrees of freedom) and consistency (reproducibility), we show how input power and physical parameters shape computational capacity. HNLF achieves up to 100 principal components and 87% MNIST classification accuracy, outperforming linear baselines. VCSELs reveal similar dependencies, with performance scaling laws analogous to classical AI. These findings highlight photonics’ potential to revolutionize computing architectures and introduce relevant application agnostic metrics.