Five research outputs across predictive coding, Bayesian sampling, controllable generation, scientific imaging, and anomaly-detection benchmarks.五項研究成果，涵蓋預測編碼、貝氏取樣、可控生成、科學影像與異常偵測基準。
文以載道
L3
Research in public研究，走到現場
Workshop · research cohortWorkshop · 研究實習
01Event photograph現場影像Presenting the NSF HDR challenge work at AAAI-25於 AAAI-25 發表 NSF HDR 挑戰成果AAAI-25 · 2025.03 · PODIUM
02Event photograph現場影像The 2025 summer research cohort at Academia Sinica IIS中研院資訊所 2025 暑期研究實習合影ACADEMIA SINICA IIS · 2025.08.28
L4
Entries條目
2026
Langevin Temporal Predictive Coding Lifts Modes to Multimodal Densities
C.-W. Lee (co-first author) et al.
NeurIPS 2026 · under review
Replaces the deterministic MAP inference of temporal predictive coding with preconditioned Langevin dynamics. Lifts posterior mode coverage from 19.4% to 71.3% (3.67×) and cuts predictive MAE by 16.7% (22.3% on naturally multimodal tasks) over twelve UCI benchmarks.把時間預測編碼的確定性 MAP 推論換成 preconditioned Langevin 動力學。在十二個 UCI benchmark 上，後驗模式涵蓋率從 19.4% 提升到 71.3%（3.67 倍），預測 MAE 降低 16.7%（在本質上呈多峰分布的任務中降低 22.3%）。
Under review審查中
2026
Predictive Coding with a Sequential Memory-Driven Bayesian Denoiser
C.-W. Lee (co-first author) et al.
NeurIPS 2026 · under review
A Tweedie–Hopfield correspondence identifies associative memory as the Bayesian denoiser predictive coding implicitly requires, yielding HOPE. It outperforms predictive-coding and Hopfield baselines on Bouncing Moving-MNIST and UCF-101 recall using only local Hebbian updates — no BPTT, no weight transport.以 Tweedie–Hopfield 對應關係指出，聯想記憶就是預測編碼所需的貝氏去噪器，並由此導出 HOPE。在 Bouncing Moving-MNIST 與 UCF-101 的序列回憶上勝過兩類基準模型，而且只使用局部 Hebbian 更新——不需 BPTT，也不需權重傳輸。
Under review審查中
2025
Building Machine Learning Challenges for Anomaly Detection in Science
E. G. Campolongo, Y.-T. Chou, E. Govorkova, … C.-W. Lee … (150+ authors)
arXiv:2503.02112 · preprint, under submission
A community benchmark paper arising from the NSF HDR ML Challenge. Contributors are listed alphabetically after the organising authors, so position carries no seniority. My contribution was feedback on benchmark positioning, baselines, metrics and clarity.源自 NSF HDR ML Challenge 的社群基準論文。核心組織者之後的貢獻者依姓氏字母排序，因此作者順序不代表貢獻程度。我的貢獻是針對基準定位、基準模型、評估指標與表述清晰度提供意見。
The two co-first-author manuscripts are under review at NeurIPS 2026; author lists remain abbreviated here while review is active.兩篇共同第一作者手稿正投稿 NeurIPS 2026 審查中；審查期間作者名單於此採簡寫。
