Random Laser's Physics Does Feature Learning, Hitting 77.7% Accuracy From Just 10 Images

bravo_abad · x · 2026-09-08

Ng and coauthors show a random laser needs no training to act as a neural network — its physics already performs feature extraction.

• In a 150-μm InP network, many lasing modes compete for the same optical gain, producing excitatory and inhibitory nonlinear responses so different modes are sensitive to different image features in parallel
• The physical network stays fixed; only a simple linear readout is trained on hyperspectral responses — the representation comes from disorder, mode competition, and heterogeneous nonlinearities, not backpropagation
• On the BreakHis breast-cancer histology task, the system reaches 77.7% accuracy from just 10 training images and outperforms baselines

Original post →

More from Research

• Physicist: Navier-Stokes is physically meaningless at scales with possible singularities — MilesCranmer · 2026-09-08
• Denny Zhou: Least-to-Most Prompting Was an Early Agent Harness, Hitting 99% on SCAN — denny_zhou · 2026-09-08
• ECCV 2026 Workshop on Privacy-Preserving Visual Localization Features Brachmann, Pollefeys Talks — ducha_aiki · 2026-09-08
• MBZUAI launches Abu Dhabi deep phenotyping study for disease prediction — thamar_solorio · 2026-09-08
• Boundary-aware self-distillation precisely tunes LLM safety refusal boundaries — MultiverseComputingCAI · 2026-09-08
• Both Nvidia and Google run ARC synthetic data generators — GregKamradt · 2026-09-08

Research channel →