Geometric Optimal Transport for Unsupervised Domain Adaptation
Authors: Gal Maman, Ronen Talmon
TMLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We showcase the effectiveness of GOT through comprehensive experiments, demonstrating its superior performance compared to recent methods across various benchmarks and datasets. |
| Researcher Affiliation | Academia | Gal Maman EMAIL Viterbi Faculty of Electrical and Computer Engineering Technion Ronen Talmon EMAIL Viterbi Faculty of Electrical and Computer Engineering Technion |
| Pseudocode | Yes | Algorithm 1 GOT... Algorithm 2 GOT on the Riemannian manifold of SPD matrices |
| Open Source Code | Yes | The source code is available here1. 1https://github.com/Gal Maman/GOT |
| Open Datasets | Yes | We apply our method to three benchmark domain adaptation datasets: Digits, Office-Home, and Vis DA. One of the key advantages of our method is its adaptability to any OT formulation. We apply our approach to two Motor Imagery (MI) benchmarks from the BCI Competition IV. The first dataset, dataset I from Blankertz et al. (2007), is referred to as MI1, and the second dataset, IIa from Tangermann et al. (2012), is referred to as MI2. |
| Dataset Splits | Yes | For the cross-subject task, we employed the Sinkhorn algorithm with entropy regularization λ = 0.02. For the first dataset, MI1, we used the following scale parameters for the probability kernels to produce the results shown in Table 3a: ϵs = 0.4, ϵc = 1, and ϵt = 1. For the second dataset, MI2, the scale parameters for the probability kernels were set as follows: ϵs = 0.9, ϵc = 2, and ϵt = 2. In this experiment, each subject, in turn, serves as the target domain, with the remaining subjects alternately acting as the source domain. |
| Hardware Specification | Yes | All the experiments were conducted on Nvidia DGX A100. |
| Software Dependencies | No | Unless specified otherwise, we employed the Sinkhorn algorithm (Cuturi, 2013) to solve the OT problem, utilizing the publicly available POT package (Flamary et al., 2021). |
| Experiment Setup | Yes | Full details of the GOT parameters are provided in Appendix C.1, and the parameters used in each experiment are reported in Appendix C. For the proposed method, we use a learning rate lr = 0.001. We set the hyperparameters η1 = 1, η2 = 1 for the objective function, ϵ = 1 for all the Gaussian kernels constructing our diffusion operator, and λ = 0.02, τ = 1 for the unbalanced Sinkhorn problem. |