The Language of Dynamics:Learning Neural Representations of Time Series Data

The Hong Kong University of Science and Technology (Guangzhou)

Data Science and Analytics Thrust

Thesis Proposal Exam

By Mr ZHONG, Siru

Abstract

Learning neural representations of time series data treats temporal observations as a language,one spoken not by people but by physical systems.This survey examines the landscape of temporal representation learning through five lenses:tokenization,the definition of a dynamical vocabulary;architectural evolution from recurrent to attention-based and neuro-cognitive designs;spatiotemporal extensions with emphasis on earth science;agent-oriented world models progressing from passive forecasting to active decision-making;and neuro-cognitive perspectives that connect predictive coding,neural manifolds,and hippocampal replay to artificial temporal intelligence.Foundation models represent a prominent paradigm within this landscape,but they are not its entirety;we trace a broader arc from classical statistical methods through deep learning architectures to the emergent frontier of large-scale pretrained models.We identify critical bottlenecks.Time series benchmarks are orders of magnitude smaller than their NLP counterparts.Evaluation protocols remain simple,static,and blind to exogenous variables that shape real-world dynamics.The gap between correlation-based forecasting and causal understanding persists.We argue that the path forward lies in neuro-cognitively grounded representation learning at the scale of earth system science,where extreme events,multi-scale dynamics,and physical constraints converge to test whether learned representations can genuinely speak the language of dynamics.

TPE Committee

• Chair: Prof. Jeffrey Xu Yu
• Prime Supervisor: Prof. Yuxuan Liang
• Co-Supervisor: Prof. Yang Yue
• Examiner: Prof. Jia Li