This repository is a structured research reading and synthesis workspace.
It supports the development of durable technical understanding, critical reading habits, reproducible research questions, and executable experimental ideas across reliable machine learning and related areas.
Paper -> Concepts and Assumptions -> Critical Analysis -> Open Question -> Testable Hypothesis -> Reproduction or Minimal Experiment -> Research Output
This repository supports my long-term research direction:
Mechanism-guided Trustworthy AI Systems.
The goal is to study how AI models can be made understandable, calibratable, intervenable, and controllable in complex dynamic environments through mechanism interpretability, concept representations, causal discovery, spatiotemporal graph learning, uncertainty calibration, and system-level monitoring.
Current flagship line:
Reliable Spatiotemporal Forecasting under Dynamic Distribution Shift: Calibration, Uncertainty Quantification, and Risk-Aware Decision-Making.
- Spatiotemporal graph forecasting: ChebNet, STGCN, DCRNN, Graph WaveNet, AGCRN, MTGNN.
- Mechanism discovery: SINDy, Neural Relational Inference, causal time-series discovery.
- Uncertainty and calibration: MC Dropout, Deep Ensembles, calibration, conformal prediction.
- Distribution-shift evaluation: robustness, stress testing, monitoring, failure analysis.
- Interpretable representation: concept representation, CKA, probing, representation stability.
- Decision reliability: selective prediction, risk-aware allocation, actionability.
The active public roadmap is tracked in
topics/reliable_spatiotemporal_forecasting/8-week-literature-and-reproduction-plan.md.
It records:
- core literature priority
- reproduction priority
- evaluation modules
- dynamic shift benchmark design
- uncertainty and calibration metrics
- risk-aware decision metrics
- 8-week technical deliverables
The reading curriculum includes foundational, canonical, frontier, and application-relevant work.
Papers are selected based on conceptual importance, methodological influence, evaluation value, reproducibility value, and relevance to active or likely future research questions.
literature-curriculum/: long-term reading tracks and verified paper indextopics/: project-specific reading maps and terminologypapers/: verified paper notes and critical reading recordsresearch-question-cards/: candidate questions, hypotheses, and minimal experimentsbacklog/: deferred research directions and activation gatestemplates/: reusable reading and research-question templates
Paper metadata, source links, and technical claims are added only after primary-source verification.
When metadata is not verified locally, title-level entries are marked
metadata-to-verify and are not given invented years, venues, links,
author lists, or paper IDs.