Summary
This paper introduces the Velocity Ambiguity Profile (VAP), a time-local Bayes-risk floor for velocity prediction in flow matching. Instead of reporting losses averaged over interpolation time—which obscures whether model error stems from intrinsic ambiguity or data limitation—the authors decompose time-resolved loss as L_N(t) = A_v(t) + R(t;N), where only the excess term R(t;N) can be reduced by more data or a better estimator. They prove two statistical benchmarks: in an intrinsic local-regression setting, A_v(t) enters the time-dependent noise factor of standard nonparametric rates; in separated isotropic Gaussian mixtures, the Bayes velocity becomes locally affine with parametric excess risk scaling K*A_v(t)/N. Controlled Gaussian mixture experiments validate the diagnostic via transition scans, oracle-to-blind calibration audits, and time-sampler stress tests. On CIFAR-100 DINOv2 feature flows, guided time sampling reduces uniform-evaluated velocity loss by 6.78% and global feature Fréchet distance by 23.6%, though class-conditional KID shows no improvement. VAP transforms averaged flow matching objectives into a time-resolved diagnostic for locating ambiguity-limited versus data-limited regions. arXiv: 2610.30107.
Paper Overview
Field: Machine Learning
Authors: Yunpeng Mei, Xiaowen Zhu, Chenyu Wang, Chenbo Xin, Hongjie Cao, Jiamin Wang, Jie Chen, Gang Wang
Published: 2026-10-21
arXiv: 2610.30107
Abstract
Flow matching trains a continuum of velocity regression problems but usually reports a loss averaged over interpolation time, obscuring which regions are limited by intrinsic ambiguity and which remain data-limited. The authors introduce the Velocity Ambiguity Profile (VAP), the time-local Bayes-risk floor of velocity prediction, and decompose time-resolved loss as:
\[L_N(t) = A_v(t) + R(t;N)\]
where only the excess term R(t;N) is reducible by more data or a different estimator.
For N training samples and a K-component mixture in the structured example, the paper proves two statistical benchmarks:
- Intrinsic local-regression benchmark: A_v(t) enters the time-dependent noise factor of the standard nonparametric rate.
- Separated isotropic Gaussian mixtures: the Bayes velocity becomes locally affine, and a label-oracle residual regression benchmark has parametric excess risk scaling K*A_v(t)/N, matched within the same separation residual Gaussian location subproblem.
Controlled Gaussian mixture experiments validate the diagnostic through dense transition scans, calibrated oracle-to-blind audits, empirical surrogate audits, and local controlled time-sampler stress tests.
On CIFAR-100 DINOv2 feature flows, guided time sampling reduces uniform-evaluated velocity loss by 6.78% and global feature Fréchet distance (FD) by 23.6%, with no improvement in class-conditional kernel inception distance (KID).
VAP turns averaged flow matching objectives into a time-resolved diagnostic tool for locating ambiguity-limited and data-limited regions.
---
*Auto-collected on 2026-10-23*
This page is an English static mirror generated for search and AI citation.
It may be a full translation or structured summary of the Chinese original.
Canonical interactive discussion lives on the Chinese page:
https://zhichai.net/topic/178636347