Reproduce “Arbitrarily Conditioned Hierarchical Flows for Spatiotemporal Events” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Keyan Chen, Qiwei Yuan, Zhitong Xu, Bin Shen, Shandian Zhe
Researchers at the University of Utah developed Arbitrarily Conditioned Hierarchical Flows (ARCH), a framework that models spatiotemporal events by combining expressive generative modeling with tractable conditional intensity computation. ARCH demonstrates superior intensity recovery on synthetic data and competitive predictive performance across diverse inference tasks on real-world datasets.
Key insights
- Decomposing spatiotemporal event generation into separate conditional flows for time and space allows for the tractable computation of conditional intensities, resolving a key limitation of many expressive generative models.
- A unified framework with a flexible masking strategy can effectively train a single model to perform diverse tasks, including next-event prediction, inverse inference, missing data imputation, and trajectory forecasting.
- Augmenting flow ODEs and utilizing the invariance of CDFs in 1D flows provides a method for accurately recovering log-densities and marginal CDFs, essential for calculating conditional intensities.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.