Seminars - Computer Science Department | Montana State University

Title: The Genesis of Time Series Foundation Models: From Generative Pre-training to Physically-Consistent Inference

Date/Time/Location: Wednesday, April 8th at 4:00 p.m. in Barnard 347

Speaker: Defu Cao

Abstract: While foundation models have transformed fields like Natural Language Processing and Computer Vision, their application to time series analysis has been limited by unique challenges inherent to temporal data, such as complex structures, pervasive data imperfections, and the need for task generality. This talk explores the emerging frontier of foundation models designed to overcome these obstacles, tracing a path through recent breakthroughs that are redefining time series analysis. We begin by examining TEMPO, a prompt-based generative pre-trained transformer that addresses the inability of standard attention mechanisms to disentangle the non-orthogonal trend, seasonal, and residual components of time series data. Building on these insights, we introduce TimeDiT, a more general-purpose diffusion transformer architecture that employs a unified masking strategy to robustly handle missing values and perform diverse tasks like forecasting, imputation, and anomaly detection from a single framework. Derived from TimeDiT, we introduce PINFDIT, a plug-and-play, physics-informed inference process to inject domain knowledge and ensure physical consistency without retraining. The primary contribution of this research will be a novel foundation model for time series that is physically consistent, advancing the field closer to the vision of a universal "world model" for temporal data. The discussion will cover the opportunities and challenges in scaling these models, their potential impact on real-world applications in finance, healthcare, and climate science, and future research directions, including multimodal integration and enhanced interpretability.

Brief Bio: Defu Cao is a Ph.D. candidate in the Department of Computer Science at the University of Southern California and a visiting scholar at Caltech. His research focuses on building practical time series foundation models and LLM-guided decision-making systems for real-world temporal data. He works on large-scale pretraining methods, model orchestration, and interpretable inference, with applications in domains such as finance, infrastructure monitoring, and large-scale forecasting systems. His work has been published in top machine learning conferences including NeurIPS, ICML, and ICLR, and has achieved state-of-the-art performance on large-scale time series benchmarks, including a top 1 result on the GIFT-Eval leaderboard. He is a recipient of the USC Best Research Assistant Award.

Title: Toward Reliable LLM Frameworks for Scientific Search Problems

Data/Time/Location: Friday, April 17th at 4:10 p.m. in Barnard 347

Speaker: Jungtaek Kim

Abstract: Large language models (LLMs) are widely used in applications such as chatbots, machine translation, and scientific discovery, and their performance can be improved through structured search over possible outputs. This talk focuses on sentence-level and idea-level search in LLMs. For sentence-level search, I show how process-supervised reward models (PRMs) can guide inference-time methods, such as weighted majority voting, beam search, and Monte Carlo tree search. Then, VersaPRM is introduced as a multi-domain PRM trained using synthetic data. For idea-level search, I isolate the search capabilities of LLMs by using them as search policies over tree-structured spaces. Using controllable synthetic benchmarks, both theoretical and empirical results demonstrate that Transformers are expressive enough to represent diverse search algorithms, along with evidence of generalization to unseen settings. The talk concludes with future directions for search-driven LLM frameworks and their applications to scientific discovery.

Bio: Jungtaek Kim is a research associate at the University of Wisconsin–Madison, working with Prof. Kangwook Lee. Previously, he was a postdoctoral associate at the University of Pittsburgh, working with Profs. Paul W. Leu, Satish Iyengar, Lucas Mentch, and Oliver Hinder. He received his Ph.D. in Computer Science and Engineering from POSTECH, under the supervision of Profs. Seungjin Choi and Minsu Cho. During his Ph.D. program, he interned at the Vector Institute and SigOpt (acquired by Intel). He has presented his work at top-tier machine learning conferences, including NeurIPS, ICML, AISTATS, ICLR, and UAI, and has served as a reviewer for several machine learning conferences. His main research interests include statistical machine learning, Bayesian optimization, large language models, and artificial intelligence for science.

PAST 2026 SEMINARS

Title: Toward Actionable and Reliable Decision Making by Sim-to-Real Framework and Trustworthy Machine Learning

Data/Time/Location: Monday, January 26th at 4:10 p.m. in Barnard 108

Speaker: Longchao Da

Abstract: Complex decision-making can be framed as a Markov Decision Process, and then solved by advanced policy learning such as Reinforcement Learning. However, policies learned in simulation often struggle to generalize to real, safety-critical environments due to distribution shift, partial observability, and uncertainty. This talk presents a line of work that addresses these challenges by developing high-fidelity simulation, introducing sim-to-real training paradigms, performing offline policy evaluation, and conducting uncertainty quantification to support actionable and trustworthy decision-making in real-world domains, with potential solutions in transportation, healthcare, and disaster monitoring and response.

Bio: Longchao Da is a Ph.D. Candidate in Computer Science at Arizona State University. His research interests include Sim-to-Real Policy Learning, Trustworthy AI, and Data Mining. He also leverages Generative AI with uncertainty quantification to detect and mitigate hallucinations for more trustworthy responses. His work has appeared in top venues like AAAI, KDD, NeurIPS, ICML, IJCAI, CIKM, ECML, CDC, etc. He is a 2025 Google PhD Fellowship nominee, a two-time ASU Ph.D. Fellowship recipient and the Best Poster Award winner at SDM 2025.

Title: Collaborative Active Learning for Robots

Data/Time/Location: Monday, February 2nd at 4:10 p.m. in Barnard 108

Speaker: Michelle Zhao

Abstract: Today, robot learning paradigms rely on human-provided data, (e.g. demonstrations, preference labels) to adapt their behavior and align with user intent. Yet in practice, this process of teaching robots is one of trial-and-error that places the burden on humans to decipher what the robot misunderstands, diagnose failures, and supply the “right” corrective data. My research develops user-centric active learning methods that learn by supporting human teachers. In this talk, I will first introduce uncertainty quantification tooling that extends conformal prediction to the human-robot interaction setting, enabling robots to rigorously “know when they don’t know” even when relying on black-box policies. I will then discuss how these uncertainty self-assessments enable robots to communicate insights with human teachers and proactively ask for targeted feedback within novel interactive learning paradigms. Coupling these ideas with cost-optimal planning algorithms, I will demonstrate how robots can interleave both learning and collaboration with human partners over multitask sequences. I will end this talk by taking a step back and examining the alignment process for robotics and discussing opportunities for how rethinking interactive learning as collaborative and continual accounts for not only task, but the nuanced interaction dynamics present during the teaching process.

Bio: Michelle Zhao is a Ph.D. candidate at Carnegie Mellon University in the Robotics Institute, working with Professors Henny Admoni and Reid Simmons. She studies human-robot interaction, with an emphasis on how robots can learn from and about people. Her research integrates methods from statistical uncertainty quantification, machine learning, and human-robot interaction to develop theoretical frameworks and practical algorithms for active learning from human feedback in domains like assistive robotic manipulation. Prior to her Ph.D., she earned her B.S. at the California Institute of Technology. She is the recipient of the Siebel Scholarship, Rising Stars in Computational and Data Sciences, the NDSEG Research Fellowship, HRI Pioneers 2025 Honorable Mention, and has worked at Toyota Research Institute.

Title: Towards Collaborative Intelligence: Learning from Decentralized Data at Scale

Date/Time/Location: Monday, February 9th at 4:10 p.m. in Barnard 108

Speaker: Yujia Wang

Abstract: As modern data increasingly comes from decentralized sources, e.g., phones, smart devices, and medical systems, learning must occur without centralizing sensitive data. Federated learning (FL) enables learning from decentralized data sources but faces significant challenges in real-world deployments, including data heterogeneity, system variability, and communication bottlenecks. In this talk, I will present the algorithmic and optimization foundations of collaborative intelligence, focusing on building efficient and scalable learning from decentralized data. My work addresses FL’s challenges both individually and in a more systematic, integrated way, depending on what the problem demands. I will first diagnose how stale updates and data heterogeneity jointly destabilize asynchronous FL and introduce a cached calibration mechanism that probably removes the harmful delay-heterogeneity interaction. I will then introduce a modularized and parallel block-coordinate framework for federated fine-tuning of large language models. Together, these results establish optimization-driven principles that enable efficient and scalable federated learning. The talk concludes with a vision for the next generation of collaborative AI, where models learn efficiently while respecting privacy, system constraints, and social trustworthiness

Bio: Yujia Wang is a Ph.D. candidate in the College of Information Sciences and Technology at The Pennsylvania State University, advised by Dr. Jinghui Chen. Her research spans the theories and applications of collaborative intelligence and privacy-preserving machine learning. Her work has been published in top venues such as ICML, NeurIPS, ICLR, AISTATS, ACL and TMLR. She has delivered technical talks at the SIAM-NNP Section Conference and IBM Research, and presented her work at the SDM Doctoral Forum. She actively serves as a reviewer for leading AI conferences and journals. Beyond academia, she gained industry experience as a Research Intern at IBM Research.

Title: Visualization Design and Artificial Intelligence for Scientific Inquiry

Date/Time/Location: Monday, February 23rd at 4:10 p.m. in Barnard 108

Speaker: Devin Lange

Abstract: Scientific inquiry seeks to understand how the world works, and data is a fundamental tool for representing underlying phenomena. In this talk, I discuss how well-designed visualizations, together with the integration of artificial intelligence, can help domain scientists interact with and reason about their data. Cancer research serves as a motivating application domain, as it involves complex, large-scale, and multimodal datasets. These data sets range from images of individual cancer cells to datasets assembled through collaborations across multiple institutions. Visualization plays a critical role in helping scientists navigate this complexity and extract meaningful insight from their data.

Bio: Devin Lange is a postdoctoral fellow in biomedical informatics at Harvard Medical School in the HIDIVE Lab, where he works with Nils Gehlenborg. He earned a PhD in computer science from the University of Utah under the supervision of Alexander Lex. His research investigates how visualization and artificial intelligence can support the understanding, discovery, and quality control of scientific data, with a particular focus on biological and biomedical data.His work has been recognized with multiple awards at IEEE VIS, including a Best Paper Award for Aardvark, a visualization system for integrated analysis of imaging, time-series, and cellular divisions. His research spans biological visualization, data forensics, and natural-language interfaces for interactive visual analysis.

Title: QCORE Tour

Date/Time/Location: Monday, March 30th at 4:00 p.m. at EngineWorks (2425 Technology Blvd)

Abstract: Learn about quantum initiatives at MSU and opportunities to engage with QCORE

Arrival: The Engine works building is located at 2425 Technology Blvd, Bozeman, MT 59718. Sometimes the address doesn't get you right to the entrance, but you park in the main parking lot, then enter the lobby under the "EngineWorks" sign. From there we have an IPAD that EngineWorks requires all members of your party to sign in on for security tracking purposes. It will prompt you to enter the name of the person you are meeting with, in this case, Jayne Morrow

Seminars from 2025.

Upcoming 2026 Seminars

PAST 2026 SEMINARS

More Information

Resources

Follow Us

Upcoming 2026 Seminars

PAST 2026 SEMINARS

Gianforte School of Computing

More Information

Resources

Follow Us