Neural Rough Differential Equations for Long Time Series

Authors: James Morrill, Cristopher Salvi, Patrick Kidger, James Foster

ICML 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We run experiments applying Neural RDEs to four real-world datasets. For each model, and each hyperparameter combination, we run the experiment three times and report the mean and standard deviation of the test metrics. We additionally report mean training times and memory usages.
Researcher Affiliation Academia 1Mathematical Institute, University of Oxford, UK 2The Alan Turing Institute, British Library, UK. Correspondence to: James Morrill <morrill@maths.ox.ac.uk>.
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes We make this approach available in the [redacted] open source project.
Open Datasets Yes Our first example uses the Eigen Worms dataset from the UEA archive from Bagnall et al. (2017). Next we consider three separate problems, using data from the TSR archive (Tan et al., 2020), coming originally from the Beth Israel Deaconess Medical Centre (BIDMC).
Dataset Splits No In practice, when choosing a final model, one would choose that with depth and step values that minimise the validation loss, as in any hyperparamter value selection. However, specific dataset split percentages for training, validation, and test sets are not provided in the main text.
Hardware Specification No The paper mentions 'GPU memory' in relation to memory usage limits, but does not provide specific details about the type or model of GPU, CPU, or any other hardware component used for the experiments.
Software Dependencies No The paper mentions the use of 'Signatory (Kidger & Lyons, 2020b)' and 'torchdiffeq (Chen, 2018)', but does not provide specific version numbers for these software dependencies.
Experiment Setup Yes Recall that the Neural RDE approach features two hyperparameters, corresponding to log-signature depth and step size. Accordingly for every experiment we run Neural RDEs for all depths in N = 2, 3 and all step sizes in 2, 4, 8, 16, 32, 64, 128, 256, 512, 1024.