Flow matching gains clearer insight with velocity uncertainty profile
Velocity Ambiguity Profiles: Time-Resolved Bayes-Risk Diagnostics for Flow Matching
Machine Learning
Summary
Predicting how things move over time can be tricky because some moments are naturally uncertain, while others just lack enough data. The authors introduce a new way to tell when errors in these predictions come from true ambiguity and when more data or smarter methods could help. They prove mathematical benchmarks and test their idea with controlled experiments and real image data. This approach helps find which time points are hard to predict because of intrinsic confusion versus insufficient information.
What this means in practice
- •For machine learning engineers: Identify exactly when velocity prediction problems are due to fundamental ambiguity versus lack of data to improve model training focus.
- •For computer vision developers: Use time-resolved diagnostics to better allocate resources for flow estimation in feature tracking and image sequence analysis.
Authors
Yunpeng Mei, Xiaowen Zhu, Chenyu Wang, Chenbo Xin, Hongjie Cao, Jiamin Wang, Jie Chen, Gang Wang
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. We introduce the Velocity Ambiguity Profile (VAP), $\mathcal{A}_v(t)$, the time-local Bayes-risk floor of velocity prediction, and decompose time-resolved loss as $L_N(t)=\mathcal{A}_v(t)+R(t;N)$, 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, we prove two statistical benchmarks. In an intrinsic local-regression benchmark, $\mathcal{A}_v(t)$ enters the time-dependent noise factor of the standard nonparametric rate. In separated isotropic Gaussian mixtures, the Bayes velocity becomes locally affine, and a label-oracle residual-regression benchmark has parametric excess-risk scale $K\mathcal{A}_v(t)/N$, matched within the same separated residual Gaussian-location subproblem. Controlled Gaussian-mixture experiments validate this diagnostic through a dense transition sweep, calibrated oracle-to-blind and empirical-proxy audits, and a local controlled time-sampler stress test in which a frozen reducible-error proxy reduces reducible loss while cross-cell sensitivity delimits aggressive sampling. On CIFAR-100 DINOv2 feature flow, guided time sampling reduces uniformly evaluated velocity loss by $6.78\%$ and global feature Fréchet distance (FD) by $23.6\%$, while class-conditional kernel inception distance (KID) shows no improvement. VAP turns an averaged flow-matching objective into a time-resolved diagnostic for locating ambiguity-limited and data-limited regions.