SigGPDE: Scaling Sparse Gaussian Processes on Sequential Data

Authors: Maud Lemercier, Cristopher Salvi, Thomas Cass, Edwin V. Bonilla, Theodoros Damoulas, Terry J Lyons

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

Reproducibility Variable Result LLM Response
Research Type Experimental We showcase the significant computational gains of Sig GPDE compared to existing methods, while achieving state-of-the-art performance for classification tasks on large datasets of up to 1 million multivariate time series.
Researcher Affiliation Academia 1University of Warwick and Alan Turing Institute 2University of Oxford and Alan Turing Institute 3Imperial College London and Alan Turing Institute 4CSIRO s Data61 and The University of Sydney.
Pseudocode Yes Algorithm 1 Backpropagation for kθ(X, X) via PDE (41)
Open Source Code No The paper states 'All code is written in Tensor Flow using GPFlow', but does not explicitly provide a link or statement that the code for Sig GPDE itself is open-source or available.
Open Datasets Yes We use a mixture of UEA & UCR time series datasets (timeseriesclassification.com) and real world data for the final example.
Dataset Splits Yes For each dataset all models are trained 3 times using a random training-validation split. The validation split is used to monitor the NLPP when optimizing the hyperparameters of the models.
Hardware Specification No No specific hardware details (like GPU/CPU models, memory, or cloud instance types) used for experiments are mentioned in the paper.
Software Dependencies No The paper states 'All code is written in Tensor Flow using GPFlow', but does not provide specific version numbers for these software dependencies.
Experiment Setup Yes For GPSig-IS, we use inducing sequences of length ℓ= 5 as recommended in Toth & Oberhauser (2020). We made use of M = 500 inducing features.