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; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1]
Flowing Datasets with Wasserstein over Wasserstein Gradient Flows
Authors: Clรฉment Bonet, Christophe Vauthier, Anna Korba
ICML 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We apply our framework to transfer learning and dataset distillation tasks, leveraging our gradient flow construction as well as novel tractable functionals that take the form of Maximum Mean Discrepancies with Sliced-Wasserstein based kernels between probability distributions. ... We focus on image datasets, and show that the flow enables structured transitions of classes toward other classes, with applications to transfer learning and dataset distillation. ... Figure 1: Minimization of F(P) = 1/2MMD2(P, Q) with Q a mixture of 3 rings, and with kernels either the Gaussian SW kernel with bandwidth h = 0.05 or Riesz SW kernel, for a learning rate of ฯ = 0.1. We observe that they first form a ring for each distribution, and then each ring converges to a target ring. |
| Researcher Affiliation | Academia | 1ENSAE, CREST, IP Paris 2Universit e Paris-Saclay, Laboratoire de Math ematique d Orsay. Correspondence to: Cl ement Bonet <EMAIL>, Chrisophe Vauthier <EMAIL>. |
| Pseudocode | No | No explicit pseudocode or algorithm blocks are present in the paper. The methodology, including the forward scheme, is described using mathematical equations and descriptive text. |
| Open Source Code | Yes | Code available at https://github.com/clbonet/Flowing_Datasets_with_WoW_Gradient_Flows. |
| Open Datasets | Yes | We consider the *NIST datasets, i.e., MNIST (Le Cun & Cortes, 2010), Fashion-MNIST (FMNIST) (Wang et al., 2018), KMNIST (Clanuwat et al., 2018) and USPS (Hull, 1994). ... We also consider CIFAR10 (Krizhevsky et al., 2009) and SVHN (Netzer et al., 2011) which are of size 32 ร 32 ร 3. |
| Dataset Splits | Yes | Here, we first train a classifier on 5000 samples of MNIST with n = 500 images by class. Then, we flow the other datasets to MNIST (with ฯ = 0.1 and momentum m = 0.9), and measure the accuracy of the pretrained classifier on the flowed dataset. ... We use here n = 100 samples by class, and run the scheme for 500K steps with a step size of ฯ = 0.1 and m = 0.9. |
| Hardware Specification | Yes | On a Nvidia v100 GPU, the flow implemented in Jax (Bradbury et al., 2018) runs in around 10 minutes with the embedding and in 30 seconds without it. |
| Software Dependencies | No | The paper mentions software like Jax (Bradbury et al., 2018), Python Optimal Transport library (Flamary et al., 2021), and scikit-learn (Pedregosa et al., 2011) but does not provide specific version numbers for these components. |
| Experiment Setup | Yes | We run the scheme for 500K steps with a step size of ฯ = 0.1 and a momentum of m = 0.9. ... The classifier is the CNN used in the examples of the equinox library (Kidger & Garcia, 2021). It is trained for 5000 steps with the Adam W optimizer (Loshchilov & Hutter, 2019) and a batch size of 64. |