Reproduce “Generalized Prompt Tuning: Adapting Frozen Univariate Time Series Foundation Models for Multivariate Healthcare Time Series” 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
Mingzhu Liu, Angela H. Chen, George H. Chen
Gen-P-Tuning adapts frozen univariate time-series foundation models to multivariate healthcare data by learning input-dependent prefixes that encode relationships across channels. With MOMENT, it used only about 0.2% of model parameters and improved classification metrics on MIMIC-III mortality and phenotyping tasks, although its benefits were model- and task-dependent and did not consistently extend to forecasting or GPT4TS.
Key insights
- Cross-channel prompting can make a univariate foundation model more suitable for multivariate clinical data without modifying the backbone; for MOMENT, the adaptation required 619,777 trainable parameters, roughly 0.2% of the model.
- Performance depends strongly on the pretrained architecture and downstream task: Gen-P-Tuning was strongest for MOMENT-based MIMIC classification, while linear probing led on MOMENT forecasting and full fine-tuning or LoRA sometimes performed better with GPT4TS.
- Parameter efficiency does not imply computational speed: the prompt module added processing overhead, and the method took longer to train than full fine-tuning or linear probing in the reported mortality experiment.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.