Scalable Inference in SDEs by Direct Matching of the Fokker–Planck–Kolmogorov Equation

Authors: Arno Solin, Ella Tamir, Prakhar Verma

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

Reproducibility Variable Result LLM Response
Research Type Experimental The goals of the experiments are three-fold: We first provide a study of the computational complexity. Then, we look into properties of the GP-SDE model from Sec. 2.1, where the experiments are concerned with showcasing model specification rather than inference. Finally, we consider two benchmark problems with high-dimensional inputs for learning a latent SDE model, where we test the performance of the approximations presented when the model is not defined by GPs, as the SDE methods presented in Sec. 2 are model-agnostic.
Researcher Affiliation Collaboration Arno Solin Aalto University Espoo, Finland arno.solin@aalto.fi Ella Tamir Aalto University Espoo, Finland ella.tamir@aalto.fi Prakhar Verma Aalto University Espoo, Finland prakhar.verma@aalto.fi ET has been employed part-time at Sellforte Oy during the project.
Pseudocode No The paper does not contain any structured pseudocode or algorithm blocks.
Open Source Code Yes Codes for the methods and experiments in this paper are available at http://github.com/AaltoML/ scalable-inference-in-SDEs.
Open Datasets Yes Rotating MNIST ([25], available under CC BY-SA 3.0) The CMU walking data set ([1], CMU Mo Cap available under CC BYND 4.0)
Dataset Splits Yes we model the sequences of a single subject, 35, for which there are 16 train set, three validation set and four test set sequences.
Hardware Specification Yes GPU: NVIDIA Tesla V100 32 GB with Intel Xeon Gold 6134 3.2 GHz; CPU: Xeon Gold 6248 2.50GHz.
Software Dependencies No The paper mentions implementing models in Py Torch [33] but does not specify the version number for PyTorch or any other software dependencies.
Experiment Setup Yes The hyperparameters (ℓ= 0.2, σ2 = 0.1) are the same in each.