On the Choice of Interpolation Scheme for Neural CDEs
Authors: James Morrill, Patrick Kidger, Lingyi Yang, Terry Lyons
TMLR 2022 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We empirically benchmark our online Neural CDE model on three continuous monitoring tasks from the MIMIC-IV medical database: we demonstrate improved performance on all tasks against ODE benchmarks, and on two of the three tasks against SOTA non-ODE benchmarks. |
| Researcher Affiliation | Academia | James Morrill EMAIL Department of Mathematics University of Oxford Patrick Kidger EMAIL Department of Mathematics University of Oxford Lingyi Yang EMAIL Department of Mathematics University of Oxford Terry Lyons EMAIL Department of Mathematics University of Oxford |
| Pseudocode | No | The paper describes methods using mathematical formulations and textual explanations but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code All code is available at github.com/jambo6/online-neural-cdes. |
| Open Datasets | Yes | In addition to these, we additionally examine three tasks for the MIMIC-IV database (Johnson et al., 2021; Goldberger et al., 2000) related to continuous patient health monitoring. The database contains 76540 de-identified admissions to intensive care units at the Beth Israel Deaconess Medical Center. The data is highly irregular and channels have lots of missing data. |
| Dataset Splits | Yes | These all had a 70%/15%/15% train/validation/test split, with (as required by the Neural CDE formulation) time included as a channel. |
| Hardware Specification | Yes | All experiments were run on two computers. One was equipped with two Quadro GP100 s, the other with one NVIDIA A100-PCIE-40GB. |
| Software Dependencies | No | The paper mentions specific numerical solvers like 'Dormand-Prince 5(4) (dopri5)' and 'Runge-Kutta-4 (rk4)', the 'Adam' optimizer, and the 'torchcde' library. However, it does not provide specific version numbers for any of these software libraries or components. |
| Experiment Setup | Yes | Full experimental details including optimisers, learning rates, normalisation, architectures and so on can be found in Appendix C. Training details All models were run with batch size 1024 for up to 1000 epochs. If training loss stagnated for 15 epochs the learning rate was decreased by a factor of 10. If training loss stagnated for 60 epochs then model training was terminated and the model was rolled back to the point of lowest validation loss. |