Continuous Spatiotemporal Transformer

Authors: Antonio Henrique De Oliveira Fonseca, Emanuele Zappala, Josue Ortega Caro, David Van Dijk

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We benchmark CST against traditional transformers as well as other spatiotemporal dynamics modeling methods and achieve superior performance in a number of tasks on synthetic and real systems, including learning brain dynamics from calcium imaging data.
Researcher Affiliation Academia 1Interdepartmental Neuroscience Program, Yale University, New Haven, CT, USA 2Department of Computer Science, Yale University, New Haven, CT, USA 3Department of Neuroscience, Yale University, New Haven, CT, USA 4Wu Tsai Institute, Yale University, New Haven, CT, USA 5Department of Internal Medicine (Cardiology), Yale University, New Haven, CT, USA 6Interdepartmental Program in Computational Biology & Bioinformatics, Yale University, New Haven, CT, USA.
Pseudocode Yes Algorithm 1 Implementation of the Sobolev loss.
Open Source Code Yes CST1 https://github.com/vandijklab/CST
Open Datasets Yes This data consists of 500 2D spirals of 100-time points each. Details about the data generation are described in Appendix C and an example of a curve from this dataset is shown in Figure 6.
Dataset Splits Yes The data was split into 70% of the spirals for training and 30% for validation.
Hardware Specification Yes All models were trained on an RTX 3090 NVIDIA GPU for up to 150 epochs or until convergence.
Software Dependencies No No specific ancillary software details (e.g., library or solver names with version numbers like Python 3.8, PyTorch 1.9) were provided. The paper mentions using 'Pytorch' but without a version number.
Experiment Setup Yes Both CST and the Transformer have 4 layers, 4 heads, and dmodel=32 (see Table 5 for more details).