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]
Disentangled Representation Learning with the Gromov-Monge Gap
Authors: ThƩo Uscidda, Luca Eyring, Karsten Roth, Fabian Theis, Zeynep Akata, marco cuturi
ICLR 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate the effectiveness of our approach for disentanglement across four standard benchmarks, outperforming other methods leveraging geometric considerations. |
| Researcher Affiliation | Collaboration | 1CREST-ENSAE 2Helmholtz Munich 3TU Munich 4Munich Center of Machine Learning 5Tubingen AI Center 6Apple |
| Pseudocode | Yes | Algorithm 1 GMG(x1, . . . xn, T, Īµ). |
| Open Source Code | Yes | To facilitate reproducibility, we provide the implementation code for computing the GMG on source and target batches in Appendix E. |
| Open Datasets | Yes | We benchmark over four 64 Ć 64 image datasets: Shapes3D (Kim and Mnih, 2018), DSprites (Higgins et al., 2017), Small NORB (Le Cun et al., 2004), and Cars3D (Reed et al., 2015). |
| Dataset Splits | No | The paper mentions using well-known datasets and evaluating with DCI-D, but it does not explicitly state the training, validation, or test splits used for these datasets within its text. It refers to standard hyperparameter choices from other works, which might implicitly include splits, but no specific percentages or methods are detailed here. |
| Hardware Specification | Yes | All experiments described in this paper can be conducted using a single RTX 2080TI GPU, ensuring accessibility and replicability of our results. |
| Software Dependencies | No | All our experiments build on python and the jax-framework (Babuschkin et al., 2020), alongside ott-jax for optimal transport utilities. |
| Experiment Setup | Yes | Using an Adam optimizer ((Kingma and Ba, 2014), Ī²1 = 0.9, Ī²2 = 0.999, Ļµ = 10ā8) and a learning rate of 10ā4. Following Locatello et al. (2020); Roth et al. (2023) we utilize a batch-size of 64, for which we also ablate all baseline methods. The total number of training steps is set to 300000. |