Certified Parallel-in-Time Sinkhorn for Dynamic Entropic Optimal Transport
TemporalSinkhorn speeds up repeated entropic transport solves while certifying that accepted outputs meet tolerance.
The paper introduces a parallel-in-time Sinkhorn executor that batches future candidates but only releases a deterministic safe prefix after certificate checks. Its prediction machinery can change where work is placed, but the author says it cannot approve an inaccurate result. Reported tests show speedups from 1.42x to 3.55x over a sequential warm start on synthetic streams, 3.054x to 3.632x on Flow Matching minibatches at n = 2048, and no marginal-tolerance violations. The paper notes that end-to-end Flow Matching integration, optimized-solver comparisons, and multi-node validation are still open. ArXiv · AI/CL/LG's note
The paper introduces a parallel-in-time Sinkhorn executor that batches future candidates but only releases a deterministic safe prefix after certificate checks. Its prediction machinery can change where work is placed, but the author says it cannot approve an inaccurate result. Reported tests show speedups from 1.42x to 3.55x over a sequential warm start on synthetic streams, 3.054x to 3.632x on Flow Matching minibatches at n = 2048, and no marginal-tolerance violations. The paper notes that end-to-end Flow Matching integration, optimized-solver comparisons, and multi-node validation are still open. ArXiv · AI/CL/LG's note
score 4