<ul><li>Neural Estimation for Scaling Entropic Multimarginal Optimal Transport - a new computational framework called Neural Entropic MOT (NEMOT) has been proposed for multimarginal optimal transport (MOT), providing improved scalability.</li><li>NEMOT employs neural networks trained using mini-batches, transferring computational complexity from dataset size to mini-batch size, leading to substantial gains in performance.</li><li>Formal guarantees on the accuracy of NEMOT are provided via non-asymptotic error bounds, and numerical results demonstrate significant performance gains over Sinkhorn's algorithm for multimarginal entropic Gromov-Wasserstein alignment.</li><li>NEMOT offers orders-of-magnitude speedups compared to the state-of-the-art, increasing the feasible number of samples and marginals, making it suitable for large-scale machine learning pipelines and expanding the practical applicability of entropic MOT for multimarginal data tasks.</li></ul>

Neural Estimation for Scaling Entropic Multimarginal Optimal Transport

Discover more