Continuous Temporal Domain Generalization

Authors: Zekun CAI, Guangji Bai, Renhe Jiang, Xuan Song, Liang Zhao

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

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments demonstrate the effectiveness and efficiency of our approach. The code can be found at: https://github.com/Zekun-Cai/Koodos. In this section, we present the performance of the Koodos in comparison to other approaches through both quantitative and qualitative analyses.
Researcher Affiliation Academia Zekun Cai1,4, Guangji Bai2, Renhe Jiang1*, Xuan Song3,4, and Liang Zhao2 1The University of Tokyo, Tokyo, Japan 2Emory University, Atlanta, GA, USA 3Jilin University, Changchun, China 4Southern University of Science and Technology, Shenzhen, China
Pseudocode No The paper does not contain a clearly labeled "Pseudocode" or "Algorithm" block. It describes the framework and optimization using text and mathematical equations.
Open Source Code Yes The code can be found at: https://github.com/Zekun-Cai/Koodos.
Open Datasets Yes We compare with classification datasets: Rotated Moons (2-Moons), Rotated MNIST (Rot-MNIST), Twitter Influenza Risk (Twitter), and Yearbook; and the regression datasets: Tropical Cyclone Intensity (Cyclone), House Prices (House). More details can be found in Appendix A.1.1. Rotated MNIST This dataset is a variant of the classic MNIST dataset [12]. Yearbook The Yearbook dataset [54]. Cyclone The Cyclone dataset [10].
Dataset Splits No All models were trained on training domains and then deployed on all unseen test domains. We use the last 30% domain of each dataset as test domains, which are marked with gray shading.
Hardware Specification Yes All experiments are conducted on a 64-bit machine with two 20-core Intel Xeon Silver 4210R CPU @ 2.40GHz, 378GB memory, and four NVIDIA Ge Force RTX 3090.
Software Dependencies No We use Adam Optimizer for all experiments, and we specify the architecture as well as other details for each dataset as follows: We use Neural ODEs [11] as the ODEs solvers.
Experiment Setup Yes Detailed experiment settings (i.e., dataset details, baseline details, hyperparameter settings, ablation study, scalability analysis, and sensitivity analysis) are demonstrated in Appendix A.1. The Predictive Model consists of 3 hidden layers, with a dimension of 50 each. We use a Re LU layer after each layer and a Sigmoid layer after the output layer. ... The learning rate for the Predictive Model is set at 1 10 2, while for the others is set at 1 10 3.