Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Tree-Sliced Entropy Partial Transport
Authors: Viet-Hoang Tran, Thanh Tran, Thanh Chu, Tam Le, Tan M. Nguyen
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In Section 4, we empirically evaluate Partial TSW on challenging tasks, such as enhancing noise robustness for generative models and addressing class imbalance in image-to-image translation. This section details the empirical evaluation of Partial TSW against other methods in tasks requiring noise robustness for point cloud alignment, outlier rejection in generative modeling, and effective handling of class imbalance in image to image translation. |
| Researcher Affiliation | Academia | Viet-Hoang Tran Department of Mathematics National University of Singapore EMAIL Thanh Tran College of Engineering & Computer Science Vin University EMAIL Thanh Chu School of Computing National University of Singapore EMAIL Tam Le Department of Advanced Data Science Institute of Statistical Mathematics EMAIL Tan M. Nguyen Department of Mathematics National University of Singapore EMAIL |
| Pseudocode | Yes | Algorithm 1 Partial Tree-Sliced Wasserstein distance. Input: Measures µ and ν in M(Rd), number of tree systems L, number of lines in tree system k, space of tree systems T, splitting maps α, parameters a, b, λ, total mass µ(T ), ν(T ). |
| Open Source Code | Yes | Our code is publicly available at https://github.com/thanhqt2002/Partial TSW. |
| Open Datasets | Yes | First, an Autoencoder pre-trained on MNIST digits provides 2D latent representations for digit 0 (the target class) and digit 1 (the outlier class)... |
| Dataset Splits | Yes | The generator is subsequently trained using an observed dataset, Xobs, which is a mixture composed of 90% samples drawn from X0 and 10% samples (outliers) drawn from X1. This dataset presents a significant imbalance, with 38K Young images and 10.5K Adult images. We collect samples until we can form a dataset of Nobs = 50,000 latent points, with the 90/10 proportion. This process yields an imbalanced dataset comprising approximately 38,000 Young latent vectors and 10,500 Adult latent vectors. |
| Hardware Specification | Yes | The experiments were conducted on a single NVIDIA H100 GPU. The runtime comparisons for all methods were conducted with an Intel Xeon Platinum 8580 CPU and an NVIDIA H100 GPU. The experiments for all methods were conducted on a system equipped with an Intel Xeon Platinum 8580 CPU and one NVIDIA H100 GPU. |
| Software Dependencies | No | modern deep learning frameworks like Py Torch, when using compilation tools (e.g., torch.compile ) can perform kernel fusion. For LPIPS calculations, we use the Alex Net backbone with pre-trained weights. |
| Experiment Setup | Yes | The AE is trained for 50 epochs using the Adam optimizer with a learning rate of 3 10 5 and a batch size of 256. The WAE-MMD loss uses a λ hyperparameter of 500.0 to balance reconstruction and MMD regularization terms. Training is performed for 30 epochs using the Adam optimizer with a learning rate of 2 10 4 and a batch size of 256. |