RMLR: Extending Multinomial Logistic Regression into General Geometries

Authors: Ziheng Chen, Yue Song, Rui Wang, Xiaojun Wu, Nicu Sebe

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on different Riemannian backbone networks validate the effectiveness of our framework.
Researcher Affiliation Academia Ziheng Chen1, Yue Song2 , Rui Wang3, Xiao-Jun Wu3, Nicu Sebe1 1 University of Trento, 2 Caltech, 3 Jiangnan University
Pseudocode No The paper describes methods through mathematical equations and textual explanations, but does not include a distinct section or figure explicitly labeled as 'Pseudocode' or 'Algorithm'.
Open Source Code Yes The code is available at https://github.com/Git ZH-Chen/RMLR.
Open Datasets Yes Radar2: This dataset [10] consists of 3,000 synthetic radar signals. ... HDM053: This dataset [44] contains 2,273 skeleton-based motion capture sequences... Hinss20214: This dataset [28] is a recent competition dataset... Disease [2]: It represents a disease propagation tree... Cora [54]: It is a citation network... Pubmed [46]: This is a standard benchmark... G3D[6]: This dataset consists of 663 sequences...
Dataset Splits Yes On the Hinss2021 dataset, models are fit and evaluated with a randomized leave 5% of the sessions (inter-session) or subjects (inter-subject) out cross-validation (CV) scheme. On other datasets, K-fold experiments are carried out with randomized initialization and split.
Hardware Specification Yes All experiments use an Intel Core i9-7960X CPU with 32GB RAM and an NVIDIA GeForce RTX 2080 Ti GPU.
Software Dependencies No The paper mentions software like PyTorch and Pytorch3D, and optimizers like Riemannian AMSGrad, but does not provide specific version numbers for these software components.
Experiment Setup Yes On the Radar and HDM05 datasets, the learning rate is 1e-2, and the batch size is 30. On the Hinss2021 dataset, following [37], the learning rate is 1e-3 with a 1e-4 weight decay, and batch size is 50. The maximum training epoch is 200, 200, and 50, respectively. We use the standard-cross entropy loss as the training objective and optimize the parameters with the Riemannian AMSGrad optimizer [4].