Post by Robin Janssen

Very happy to share our paper “Enabling photonic Kolmogorov-Arnold networks for ultra-fast inference”, which was recently published in Frontiers in Photonics. In this work, Ravi Pradip and I show how a compact photonic nonlinear unit can be used as a building block for ultrafast, expressive neural networks with a structure resembling Kolmogorov-Arnold Networks. A central focus of the paper is hardware realism. We deliberately constrained the simulations to designs that could plausibly fit on a manufactured chip, rather than exploring network sizes or topologies that may look promising in simulation but are difficult to realize experimentally. This means: only a few hundred nonlinear units and chip-compatible network topologies. The nonlinear responses are based on a physical model fitted to experimental measurements, and our nanosecond-scale response time claims are grounded in measured device dynamics. Despite these constraints, the resulting networks can fit a range of nonlinear functions with compact architectures and very fast evaluation times. For me, one of the main takeaways is that hardware realism and manufacturability are not just implementation details. They shape the design space in nontrivial ways, and ignoring them can make proof-of-concept simulations much less meaningful. Many thanks to Ravi and all co-authors for the collaboration!