Bridge-Preserving Gradient Calibration for Replay in Neural Schrödinger Bridges
Desk rejected at ICLR 2027, so left out of the paper lists and the map.
Abstract
Replay reduces simulation costs in neural Schrödinger bridge training, but stale endpoint pairs can distort gradient updates. Matching intermediate-state distributions alone does not resolve this mismatch and may alter conditional bridge laws. We propose Bridge-Preserving Gradient Calibration (BPGC), which reweights intact replay endpoint pairs through a bounded convex gradient-matching objective and generates independent training bridges. We establish conditional bridge preservation and a finite calibration certificate, and derive update-error bounds that expose sampling noise and unmeasured gradient components. The analysis distinguishes projected calibration from full-gradient accuracy and identifies limitations from replay support and parameter drift. We specify an evaluation protocol that separates gradient fidelity from downstream quality and compares calibration with direct fresh-data use under matched computational budgets.
- Primary area:
- generative models