ZF2ST enables powerful two-sample testing with neural networks by learning directional misalignment patterns between distributions, while keeping statistical validity through separated witness learning and hypothesis evaluation.
This paper introduces a new statistical test for determining whether two datasets come from the same distribution. The method learns how samples from each distribution are locally misaligned, then uses this pattern as evidence of difference. By separating the learning phase from the testing phase, it can use flexible neural networks while maintaining valid statistical guarantees.