Variational Entropic Optimal Transport
Researchers have introduced Variational Entropic Optimal Transport (VarEOT), a novel method for domain translation problems. This approach reformulates the intractable log-partition term in Entropic Optimal Transport into a tractable minimization problem. VarEOT utilizes an exact variational reformulation, enabling optimization with stochastic gradients and eliminating the need for simulation-based training procedures. Experiments on synthetic data and image-to-image translation tasks show competitive results, with theoretical guarantees on generalization and approximation. AI
IMPACT Introduces a new optimization principle for domain translation, potentially improving image translation quality and offering theoretical guarantees.