Sequential Latent Variable Models for Few-Shot High-Dimensional Time-Series Forecasting

Authors: Xiajun Jiang, Ryan Missel, Zhiyuan Li, Linwei Wang

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

Reproducibility Variable Result LLM Response
Research Type Experimental We compared the presented framework with a comprehensive set of baseline models 1) trained globally on the large meta-training set with diverse dynamics, 2) trained individually on single dynamics with and without fine-tuning to k-shot support series, and 3) extended to few-shot meta-formulations. We demonstrated that the presented framework is agnostic to the latent dynamic function of choice and, at meta-test time, is able to forecast for new dynamics given variable-shot of support series.
Researcher Affiliation Academia Golisano College of Computing and Information Sciences Rochester Institute of Technology Rochester, NY 14623, USA {xj7056,rxm7244,zl7904,Linwei.Wang}@rit.edu
Pseudocode No The paper describes the methodology in prose and mathematical equations but does not include any explicit pseudocode blocks or algorithm figures.
Open Source Code Yes Source code available at https://github.com/john-x-jiang/meta_ssm.
Open Datasets Yes Data: We first considered benchmark images generated with controllable physics, including bouncing ball Fraccaro et al. (2017), Hamiltonian pendulum (Botev et al., 2021), and Hamiltonian mass-spring systems (Botev et al., 2021). Details of data generation are available in Appendix G... All ball data can be found here: https://drive.google.com/drive/ folders/1Tm3DNrugc Sb WXSNye GL3j QKR8y3i Xx0m?usp=sharing. The heart data can be found here: https://drive.google.com/drive/folders/ 12S579V0KWMgb HGXDQZt0r Qyfz F1Ay NCu?usp=sharing.
Dataset Splits Yes For gravity-16 data, we used 10 gravity in meta-training, 2 in meta-validation, and 4 in meta-testing.
Hardware Specification Yes All experiments were run on NVIDIA Tesla T4s with 16 GB memory.
Software Dependencies No The paper specifies the optimizer (Adam) and neural network architectures (e.g., GRU-res, NODE, RGN-res) but does not provide specific version numbers for software dependencies like Python, PyTorch, TensorFlow, or CUDA.
Experiment Setup Yes Detailed hyperparameter settings are shown below. Meta Model Architecture on Mixed-Physics and Gravity-16: Domain Input: 20 observation timesteps of 32 32 dimensions, Initialization Input: 3 observation timesteps of 32 32 dimensions, Optimizer: Adam, 5 10 4 learning rate, Batch size: 50, Number of epochs: 200, Latent Units: 8, Transition Units: 100, Domain Encoder Filters: [8, 16, 8], Domain Time Units: [10, 5, 1], Initial Encoder Filters: [8, 16, 8], Emission Filters: [32, 16, 8, 1], KL term initialization: λ1 = 10 2, KL term set-embedding: λ2 = 10 2.