Optimal Transport for Machine Learning: Distances, Algorithms, and Domain Adaptation
par
UTC - GI
Optimal Transport (OT) offers a principled framework for domain adaptation by aligning source and target data distributions through cost-minimizing “mass transport.” In this talk, we introduce OT from the Monge formulation and show how transport plans yield interpretable couplings between samples while preserving underlying geometry. We then focus on practical OT-based adaptation pipelines—using regularized, scalable solvers (e.g., Sinkhorn-type methods) to reduce domain shift, reweight or map source data toward the target, and improve performance with minimal labeling effort. We conclude with key implementation takeaways and brief pointers to extensions such as Gromov–Wasserstein for cases where feature spaces or structures differ across domains.