Skip to content

Repository for ICLR 2023 work, "Sequential Latent Variable Models for Few-Shot High-Dimensional Time-Series Forecasting"

Notifications You must be signed in to change notification settings

john-x-jiang/meta_ssm

Repository files navigation

Sequential Latent Variable Models for
Few-Shot High-Dimensional Time-Series Forecasting
(ICLR 2023 Top-25%)
[OpenReview, Presentation]

*Xiajun Jiang, *Ryan Missel, Zhiyuan Li, Linwei Wang (*equal contribution)

framework schematic

Figure 1: Sequential latent-variable models (SLVM) framework for forecasting
high-dimensional sequences, based on the underlying PGM.

Description

Modern applications increasingly require learning and forecasting latent dynamics from high-dimensional time-series. Compared to univariate time-series forecasting, this adds a new challenge of reasoning about the latent dynamics of an unobserved abstract state. Sequential latent variable models (LVMs) present an attractive solution, although existing works either struggle with long-term forecasting or have difficulty learning across diverse dynamics. In this paper, we first present a conceptual framework of sequential LVMs to unify existing works, contrast their fundamental limitations, and identify an intuitive solution to long-term forecasting for diverse dynamics via meta-learning. We then present the first framework of few-shot forecasting for high-dimensional time-series: instead of learning a single dynamic function, we leverage data of diverse dynamics and learn to adapt latent dynamic functions to few-shot support series. This is realized via Bayesian meta-learning underpinned by: 1) a latent dynamic function conditioned on knowledge derived from few-shot support series, and 2) a meta-model that learns to extract such dynamic-specific knowledge via feed-forward embedding of support set. We compared the presented framework with a comprehensive set of baseline models trained 1) globally on the large meta-training set with diverse dynamics, and 2) individually on single dynamics, both with and without fine-tuning to k-shot support series used by the meta-models. 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.

Citation

Please cite the following if you use the data or the model in your work:

@inproceedings{
jiang2023sequential,
title={Sequential Latent Variable Models for Few-Shot High-Dimensional Time-Series Forecasting},
author={Xiajun Jiang and Ryan Missel and Zhiyuan Li and Linwei Wang},
booktitle={The Eleventh International Conference on Learning Representations },
year={2023},
url={https://openreview.net/forum?id=7C9aRX2nBf2}
}

Requirements

  • Python >= 3.8
  • PyTorch >= 1.7
  • ipdb
  • matplotlib
  • numpy
  • scipy
  • Pillow
  • pymunk
  • pygame
  • torchdiffeq

All Datasets

  1. Bouncing ball dataset generated using the code provided with the original paper.
  2. Mixed-physics datasets generated using the code provided with the original paper
  3. Check here for generated datasets.

Cardiac Electrophysiology Model

  1. Check EP_model for the model specific to cardiac data.

About

Repository for ICLR 2023 work, "Sequential Latent Variable Models for Few-Shot High-Dimensional Time-Series Forecasting"

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published