Zero-Flow Two-Sample Tests
The paper’s test looks for local directional misalignment between two sample sets, then uses that pattern as evidence they came from different distributions.
The authors introduce zero-flow discrepancy and a practical procedure called ZF2ST. They separate witness learning from hypothesis evaluation so neural networks can be used while keeping statistical calibration valid. They describe regression-based and power-maximized ways to learn the witness. Experiments on synthetic and image datasets report strong power on structured distribution shifts with calibrated type-I error. ArXiv · AI/CL/LG's note
The authors introduce zero-flow discrepancy and a practical procedure called ZF2ST. They separate witness learning from hypothesis evaluation so neural networks can be used while keeping statistical calibration valid. They describe regression-based and power-maximized ways to learn the witness. Experiments on synthetic and image datasets report strong power on structured distribution shifts with calibrated type-I error. ArXiv · AI/CL/LG's note
score 5