New IPC framework benchmarks physical computers with photonic demo

Researchers Rahul Uma Ramachandran and Serge Massar have extended the Information Processing Capacity (IPC) framework to stationary physical computing systems, addressing a long-standing challenge in characterizing hardware-native machine learning. The theory establishes rigorous bounds: individual capacities lie between zero and one, their sum over a complete basis is bounded by the number of readouts, and noise strictly reduces this bound. To tackle finite-sample estimation, the authors derive the asymptotic form of systematic positive bias affecting naive estimators and propose data-efficient methods based on Richardson extrapolation and Sobol quasi-random sampling. These methods dramatically reduce the amount of data needed for accurate IPC estimation.

The framework was experimentally validated using a photonic computing system that sends picosecond laser pulses through a nonlinear optical fiber. By varying laser power and fiber length, the team observed systematic shifts in the IPC distribution toward higher-order nonlinear capacities, driven by the Kerr effect. Critically, total IPC strongly correlates with performance on benchmark machine-learning tasks and provides a reliable estimate of the system's effective dimensionality. These results establish IPC as a practical bridge between the intrinsic dynamics of physical computing systems and their machine-learning performance, offering a principled, task-independent way to evaluate and design next-generation hardware for AI.