Neural Spatio-Temporal Point Processes
Authors: Ricky T. Q. Chen, Brandon Amos, Maximilian Nickel
ICLR 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We validate our models on data sets from a wide variety of contexts such as seismology, epidemiology, urban mobility, and neuroscience. and 5 EXPERIMENTS Data Sets Many collected data can be represented within the framework of spatio-temporal events. We pre-process data from open sources and make them suitable for spatio-temporal event modeling. |
| Researcher Affiliation | Collaboration | Ricky T. Q. Chen University of Toronto; Vector Institute rtqichen@cs.toronto.edu Brandon Amos, Maximilian Nickel Facebook AI Research {bda,maxn}@fb.com |
| Pseudocode | No | The paper does not contain any clearly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Code for preprocessing and training are open sourced at https://github.com/facebookresearch/neural_stpp. |
| Open Datasets | Yes | EARTHQUAKES For modeling earthquakes and aftershocks, we gathered location and time of all earthquakes in Japan from 1990 to 2020 with magnitude of at least 2.5 from the U.S. Geological Survey (2020). COVID-19 CASES We use data released publicly by The New York Times (2020) on daily COVID-19 cases in the New Jersey state, from March to July of 2020. BOLD5000 This consists of f MRI scans as participants are given visual stimuli (Chang et al., 2019). |
| Dataset Splits | Yes | EARTHQUAKES: This resulted in 950 training sequences, 50 validation sequences, and 50 test sequences. COVID-19 CASES: This resulted in 1450 training sequences, 100 validation sequences, and 100 test sequences. CITIBIKE: This resulted in 2440 training sequences, 300 validation sequences, and 320 test sequences. BOLD5000: We split the data into 1050 training sequences, 150 val sequences, 220 test sequences. |
| Hardware Specification | No | The paper does not provide specific hardware details such as exact GPU/CPU models or processor types used for running its experiments. |
| Software Dependencies | No | The paper lists several software tools in the acknowledgments, such as PyTorch, torchdiffeq, numpy, and SciPy, but it does not specify their version numbers. For example, it mentions 'Py Torch (Paszke et al., 2019)' where '2019' is the publication year, not the software version. |
| Experiment Setup | Yes | For the time-varying, jump, and attentive CNF models, we parameterized the CNF drift as a multilayer perceptron (MLP) with dimensions [d 64 64 64 d], where d is the number of spatial variables. We ultimately decided on using the time-dependent Swish for all experiments. We swept over the MLP for defining fh for the continuous-time hidden state in eq. (11) using hidden widths of [8 20], [32 32], [64 64], [32 32 32], and [64 64 64]. The majority of models used 32 32 as it provided enough flexibility while remaining easy to solve. We regularized the L2 norm of the hidden state drift with a strength of 1e-4, chosen from {0, 1e-4, 1e-3, 1e-2}. All integrals were solved using Chen (2018) to within a relative and absolute tolerance of 1E-4 or 1E-6, chosen based on preliminary testing for convergence and stability. |