Adaptive patching harder than expected for time-series Transformers

Arxiv CS.LG (machine learning)

Paper shows adaptive patching, which allocates finer patches to informative regions, often underperforms uniform patching in time-series forecasting. The study reveals that the adaptive operator's inductive bias can hurt generalization, challenging recent proposals.

transformer

Arxiv CS.LG (machine learning)·Federico ZucchiYi XieChao ZhangKeyuan LuoThomas LampertZiyue Li

7 days ago