Data repository of the paper "Machine vision with small numbers of detected photons per inference"

This repository contains the data, trained models, and code needed to reproduce the main figures from the paper: "Machine vision with small numbers of detected photons per inference" Shi-Yuan Ma, Jérémie Laydevant, Mandar M. Sohoni, Logan G. Wright, Tianyu Wang, and Peter L. McMahon The PANS Framework Photon-aware neuromorphic sensing (PANS) is a framework for optimizing sensing systems under extreme photon-budget constraints. PANS models the stochastic single-photon detection process exactly as it physically occurs, rather than relying on conventional surrogate approximations. By faithfully incorporating detection stochasticity into the forward pass and using customized gradient estimation techniques for backpropagation, PANS enables effective end-to-end optimization of both the physical optical front end and the digital back end, jointly optimized through the lossy, stochastic detection bottleneck under stringent physical resource constraints. Getting Started This repository is fully self-contained: all reported results can be reproduced end-to-end from the trained models and raw experimental data, not just replotted from saved numbers. The notebooks in `test/` run the trained PANS models on the provided datasets (and, for FashionMNIST/MNIST, on the experimentally collected photon counts) to regenerate the result files in `results/`. The notebooks in `main_figures/` then produce the plots in the main text (Figs. 2–5) from those results. For more information, please refer to 'README.txt'.