Inference from Real-World Sparse Measurements
Authors: Arnaud Pannatier, Kyle Matoba, François Fleuret
TMLR 2024 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct in-depth ablation studies that characterize this problematic bottleneck in the latent representations of alternative models that inhibit information utilization and impede training efficiency. We also perform experiments across various problem domains, including high-altitude wind nowcasting, two-day weather forecasting, fluid dynamics, and heat diffusion. Our attention-based model consistently outperforms state-of-the-art models in handling irregularly sampled data. Notably, our model reduces the root mean square error (RMSE) for wind nowcasting from 9.24 to 7.98 and for heat diffusion tasks from 0.126 to 0.084. |
| Researcher Affiliation | Academia | Arnaud Pannatier EMAIL Idiap Research Institute, Martigny, Switzerland École Polytechnique Fédérale de Lausanne, Lausanne, Switzerland Kyle Matoba Idiap Research Institute, Martigny, Switzerland École Polytechnique Fédérale de Lausanne, Lausanne, Switzerland François Fleuret Université de Genève, Geneva, Switzerland |
| Pseudocode | No | The paper describes the models using mathematical equations and text, but does not present any structured pseudocode or algorithm blocks. |
| Open Source Code | No | The paper states in section Q 'Reproducibility' that 'We provide the link to the dataset and the whole code base for processing it and running the experiments,' but no specific URL for the codebase of the described methodology is provided within the paper text. |
| Open Datasets | Yes | The wind nowcasting experiment uses the same dataset as (Pannatier et al., 2021). It is available at: https://zenodo.org/record/5074237. [...] This experiment uses the same dataset as (Alet et al., 2019). It is available at https://github.com/FerranAlet/graph_element_networks/tree/master/data. [...] This experiment adapts the dataset available in (Li et al., 2021). It is available at https://github.com/neural-operator/fourier_neural_operator [...] The requested data collection description focuses on climate reanalysis from the ERA5 dataset, which is available through the Copernicus Climate Data Store (CDS). To access the data, please visit the following URL: https://cds.climate.copernicus.eu/cdsapp#!/dataset/reanalysis-era5-single-levels?tab=form. |
| Dataset Splits | Yes | The training and validation set have respectively 10 000 and 1 000 pairs of sets of 64 examples. (Section 5.1). For training purposes, we selected the data from the year 2000. For validation, we use the data from the year 2010, specifically the months of January and September. (Appendix B.5). Additionally, Table 7, titled 'Description of the different datasets used in this study,' explicitly lists '# Train. points' and '# Val. points' for each dataset, providing absolute sample counts for the training and validation sets. For example, for Wind Nowcasting: '26.956.857' (train) and '927.906' (val). |
| Hardware Specification | Yes | Each model ran for 10, 2000, 1000, 100, and 100 epochs respectively on an NVIDIA GeForce GTX 1080 Ti. |
| Software Dependencies | No | The paper mentions using 'Hydra as a configuration manager' and 'PyTorch Geometric' but does not provide specific version numbers for these or any other software dependencies, which are necessary for reproducible setup. |
| Experiment Setup | Yes | We select model configurations with approximately 100,000 parameters and run each model using three different random seeds in both cases. (Section 4). Each model ran for 10, 2000, 1000, 100, and 100 epochs respectively on an NVIDIA GeForce GTX 1080 Ti. (Table 2 footnote). We focus on the number of layers [Figure 12a] and the hidden dimension [Figure 12c] in the case of the Poisson equation. In the Table 5, we report three different sizes of the different models, and we tried to keep them as balanced as possible. The exact configurations of the models can be found in the config folder of the code, however, for the sake of completeness, we grid-searched over the mentioned hyperparameters while randomly sampling the learning rate. (Appendix L). |