Optimal Transport Theory for Domain Adaptation

Optimal transport (OT) theory is applied in deep learning for domain adaptation, aligning distributions between a source and target domain. This adaptation addresses performance decline due to domain differences. OT quantifies divergence using Wasserstein distance and transports source to target distributions, enhancing model performance on the target domain. However, existing methods struggle with domain shifts. This thesis introduces innovative OT-based approaches to address data and label shift problems, and achieve superior unsupervised domain adaptation results compared to recent baselines. This highlights OT's effectiveness in addressing domain adaptation and transfer learning challenges.