The optimal transport (OT) map offers a geometric lens for aligning probability distributions and has become a powerful, efficient, and reliable tool across diverse machine learning (ML) and artificial intelligence (AI) tasks.
In this thesis, we study...
The optimal transport (OT) map offers a geometric lens for aligning probability distributions and has become a powerful, efficient, and reliable tool across diverse machine learning (ML) and artificial intelligence (AI) tasks.
In this thesis, we study the estimation and real-world applications of the OT maps. First, we develop a practical estimator of the OT maps using deep neural networks (DNNs) and explore its applications to AI, particularly for visual learning tasks. Then, we introduce a novel algorithm for fair clustering based on the OT map and validate its strong empirical performance.
In the first part, we introduce a two-step scheme for the OT map estimation based on the optimal coupling. The proposed method first solves the Kantorovich problem to obtain the empirical optimal coupling, and then fits the an DNN to the coupled pairs obtained from the first step. Under standard regularity conditions, we show the minimax optimal convergence rate of the proposed estimator for Lipschitz OT maps up to logarithmic factors. Experiments on synthetic datasets and computer vision tasks show that the proposed estimator outperforms or performs competitive to existing baselines for estimating the OT map.
In the second part, we present a novel transport map-based fair clustering algorithm that achieves fairness by transporting protected groups into a common aligned space and then running standard clustering on the aligned space. Theoretically, we prove that the fair clustering can be obtained by jointly finding the coupling (i.e., matching) to build the aligned space and optimizing the cluster centers on the aligned space. Experimentally, across various benchmark datasets, the proposed algorithm can attain any given fairness level with competitive clustering cost, exhibits numerical stability, and scales easily.