Characteristic Neural Ordinary Differential Equation
Authors: Xingzi Xu, Ali Hasan, Khalil Elkhalil, Jie Ding, Vahid Tarokh
ICLR 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical results demonstrate the improvements provided by the proposed method for irregularly sampled time series prediction on Mu Jo Co, Physio Net, and Human Activity datasets and classification and density estimation on CIFAR-10, SVHN, and MNIST datasets given a similar computational budget as the existing NODE methods. |
| Researcher Affiliation | Academia | 1 Department of Electrical and Computer Engineering, Duke University 2 School of Statistics, University of Minnesota |
| Pseudocode | Yes | Algorithm 2 C-NODE algorithm using the forward Euler method (Appendix F); Algorithm 3 Algorithm for training CNFs defined with C-NODE (Appendix G); Algorithm 4 Algorithm for sampling CNFs defined with C-NODE (Appendix G). |
| Open Source Code | Yes | Implementation details of this paper can be found at https://github.com/Xingzi Xu/Neural PDE.git. |
| Open Datasets | Yes | Empirical results demonstrate the improvements provided by the proposed method for irregularly sampled time series prediction on Mu Jo Co, Physio Net, and Human Activity datasets and classification and density estimation on CIFAR-10, SVHN, and MNIST datasets... and Physio Net dataset... (Goldberger et al., 2000 (June 13)); Human Activity dataset in the UCI dataset... (Dua & Graff, 2017). |
| Dataset Splits | No | The paper uses standard benchmark datasets like MNIST, CIFAR-10, SVHN, PhysioNet, and Human Activity, and describes the experimental setups, but does not explicitly provide the specific percentages or counts for train/validation/test dataset splits used for all experiments. |
| Hardware Specification | Yes | All experiments were performed on NVIDIA RTX 3090 GPUs on a cloud cluster.; All experiments were performed on NVIDIA RTX 3080 ti GPUs on a local machine. |
| Software Dependencies | No | The paper mentions optimizers like 'Adam W' and 'Adamax', and numerical integration methods like 'Runge-Kutta method of order 5 of the Dormand-Prince-Shampine solver', but does not provide specific version numbers for these software components or the libraries implementing them (e.g., PyTorch version, torchdyn version). |
| Experiment Setup | Yes | Table 4: Training hyperparameters for image classification. Learning Rate 1.00E-3, Weight Decay 5.00E-04; Table 9: Training hyperparameters for time series analysis. Learning Rate 1.0E-2, Weight Decay 0.0; Table 14: Training hyperparameters for continuous normalizing flow. Learning Rate 1.00E-3, Weight Decay 0.0. |