Identifying Spatio-Temporal Drivers of Extreme Events

Authors: Mohamad Hakam Shams Eddin, Jürgen Gall

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our approach on three newly created synthetic benchmarks, where two of them are based on remote sensing or reanalysis climate data, and on two real-world reanalysis datasets. [...] Our evaluation shows that our approach outperforms approaches for interpretable forecasting, spatio-temporal anomaly detection, out-of-distribution detection, and multiple instance learning. Furthermore, we conduct empirical studies on two real-world reanalysis climate data.
Researcher Affiliation Academia Mohamad Hakam Shams Eddin Juergen Gall Institute of Computer Science, University of Bonn Lamarr Institute for Machine Learning and Artificial Intelligence {shams, gall}@iai.uni-bonn.de
Pseudocode No The paper describes the model components and training process in text and mathematical equations, but it does not include a dedicated 'Pseudocode' or 'Algorithm' block.
Open Source Code Yes The source code and datasets are publicly available at the project page https://hakamshams.github.io/IDE.
Open Datasets Yes The source code and datasets are publicly available at the project page https://hakamshams.github.io/IDE. [...] We conducted the experiments on two real-world reanalysis datasets; CERRA reanalysis [106] and ERA5-Land [107]. [...] The pre-processed data used in this study are available at https://doi.org/10.60507/FK2/RD9E33 [139].
Dataset Splits Yes More details regarding the variables and the domains along with the training/validation/test splits are provided in the Appendix Sec. H and Tables 20 and 21. [...] We synthesize overall 46 years of data; 34 years for training, 6 subsequent years for validation and the last 6 years for testing.
Hardware Specification Yes The training was done mainly on clusters with NVIDIA A100 80GB and NVIDIA A40 48GB GPUs.
Software Dependencies No The paper mentions using 'Adam optimizer [129]' and 'PyTorch Captum [128]' but does not provide specific version numbers for these or other key software components like Python, PyTorch, or CUDA.
Experiment Setup Yes Setup and implementation details. We set the hidden dimension K to 16 by default. The temporal resolution is T = 6 for the synthetic data and T = 8 for real-world data. [...] For the synthetic data we set the embedding dimension K = 16. We use one layer Video Swin Transformer with {depth=[2, 1], heads=[2, 2], window size=[[2, 4, 4], [6, 1, 1]]}. [...] We set \u03bb(ent) = \u03bb(div) = 0.1, \u03bb(anomaly) = 100, and \u03bb(commit) = 3. The models were trained with Adam optimizer [129] for 100 epochs with a batch size of 4. We use a linear warm up of 2 epochs and a cosine decay with an initial learning rate of 2 \u00d7 10\u207b\u00b3 and a weight decay of 3 \u00d7 10\u207b\u00b3.