Title: Engineering Faithful and Interpretable AI Systems

Time: 11:30am 

Venue: Innovation Wing Two, G/F, Run Run Shaw Building

Speaker(s):  Prof. René Vidal

Remark(s): 

Abstract

Large Language Models (LLMs) and Vision Language Models (VLMs) have achieved remarkable performance across a wide range of tasks. However, their growing deployment has exposed fundamental limitations in faithfulness, safety, and transparency. In this talk, Prof. Vidal will present a unified perspective on addressing these challenges through principled model interventions and interpretable decision-making frameworks. He first introduces Parsimonious Concept Engineering (PaCE), an approach that improves faithfulness and alignment by selectively removing undesirable internal activations, mitigating hallucinations and biased language while preserving linguistic competence. Prof. Vidal then present Information Pursuit (IP), an interpretable-by-design prediction framework that replaces opaque reasoning with a sequence of informative, user-interpretable queries, yielding concise explanations alongside accurate predictions. Results across text, vision, and medical tasks illustrate how these ideas advance transparency without sacrificing performance. Together, these contributions point toward a broader direction for building AI systems that are powerful, faithful, and aligned with human values.

About the speaker

René Vidal is the Penn Integrates Knowledge and Rachleff University Professor of Electrical and Systems Engineering & Radiology, the Director of the Center for Innovation in Data Engineering and Science (IDEAS), and Co-Chair of Penn AI at the University of Pennsylvania. He is also an Amazon Scholar, an Affiliated Chief Scientist at NORCE, and a former Associate Editor in Chief of TPAMI. His current research focuses on the foundations of deep learning and trustworthy AI and its applications in computer vision and biomedical data science. His lab has made seminal contributions to motion segmentation, action recognition, subspace clustering, matrix factorization, deep learning theory, interpretable AI, and biomedical image analysis. He is an ACM Fellow, AIMBE Fellow, IEEE Fellow, IAPR Fellow and Sloan Fellow, and has received numerous awards for his work, including the IEEE Edward J. McCluskey Technical Achievement Award, D’Alembert Faculty Award, J.K. Aggarwal Prize, ONR Young Investigator Award, NSF CAREER Award as well as best paper awards in machine learning, computer vision, signal processing, controls, and medical robotics.

Title: Harmonizing the Trilemma: Orchestrating Privacy, Robustness, and Efficiency in Collaborative Intelligence

Time: 10:30pm 

Venue: CB 308

Speaker(s): Prof. Runhua Xu

Remark(s): 

Abstract

Collaborative intelligence, particularly Federated Learning, has emerged as a paradigm shift for decentralized knowledge discovery, promising to unlock data silos while safeguarding user privacy. However, real-world deployments face a critical "trilemma": the intrinsic tensions between rigorous privacy preservation, adversarial robustness, and system efficiency. In this talk, I will outline a roadmap to reconcile these challenges by exploring the intersection of efficiency, privacy and robustness -- focusing on methodologies that enable anomaly detection directly over encrypted models without compromising confidentiality, examining the security implications of communication-efficient FL, etc. Collectively, these insights pave the way for constructing a trustworthy, scalable, and secure collaborative AI ecosystem.

About the speaker

Runhua Xu is currently a Professor in the School of Computer Science and Engineering at Beihang University (BUAA). He is a recipient of the National Youth Talent Program. Prior to joining BUAA, he served as a Research Staff Member at IBM Research, leading multiple projects on federated learning security and privacy. His research interests encompass privacy-enhancing technologies, AI security/privacy, and trusted computing infrastructure. Dr. Xu has published extensively in top-tier conferences and journals, including ACM CCS, USENIX Security, NeurIPS, AAAI, IEEE TDSC, and IEEE TIFS. His work has been recognized with prestigious awards, including the ACM CCS 2023 Distinguished Paper Award and the IEEE CLOUD 2022 Best Paper Award. He serves as an Associate Editor for _IEEE TDSC_ and on the Youth Editorial Boards of Chinese Journal of Electronics and ELSP Blockchain. Additionally, he regularly serves on the program committees for premier conferences such as AAAI, ICDM, ESORICS, and ACM SACMAT.

Title: Optimal Decision Rules With Policy-Relevant Guarantees

Time: 02:00pm 

Venue: Room 301, Run Run Shaw Building

Speaker(s): Prof. Mats Julius Stensrud

Remark(s): 

Abstract

Policy makers desire to implement decision rules that, when applied to individuals in the population of interest, yield the best possible outcomes. For example, the current focus on precision medicine reflects the search for individualized decision rules, adapted to a patient's characteristics. In this presentation, I will study how to define, choose, and estimate effects that inform individualized decisions. A central difficulty, common to most existing approaches, is that as we include more covariates and aim for finer personalization, the required assumptions become stronger.

As an alternative, I propose a strategy for detecting and estimating group-level effects,with statistical guarantees that the estimated groups truly differ. I then show that, in realistic settings, group-based decision rules can substantially outperform state-of-the-art optimal-regime methods, even when those methods rely on correctly specified models and are implemented with modern doubly robust machine-learning estimators.

Title: Optimal No-Regret Learning in Repeated First-Price Auctions

Time: 10:30am 

Venue: CB 308

Speaker(s): Prof. Zhengyuan Zhou, New York University Stern School of Business

Remark(s): 

Abstract

First-price auctions have very recently swept the online advertising industry, replacing second-price auctions as the predominant auction mechanism on many platforms for display ads bidding. This shift has brought forth important challenges for a bidder: how should one bid in a first-price auction, where unlike in second-price auctions, it is no longer optimal to bid one’s private value truthfully and hard to know the others’ bidding behaviors? In this paper, we take an online learning angle and address the fundamental problem of learning to bid in repeated first-price auctions. We discuss our recent work in leveraging the special structures of the first-price auctions to design minimax optimal no-regret bidding algorithms.

About the speaker

Zhengyuan Zhou is currently an associate professor in New York University Stern School of Business, Department of Technology, Operations and Statistics. Before joining NYU Stern, Professor Zhou spent the year 2019-2020 as a Goldstine research fellow at IBM research. He received his BA in Mathematics and BS in Electrical Engineering and Computer Sciences, both from UC Berkeley, and subsequently a PhD in Electrical Engineering from Stanford University in 2019. His research interests lie at the intersection of machine learning, stochastic optimization and game theory and focus on leveraging tools from those fields to develop methodological frameworks to solve data-driven decision-making problems.

Title: The Promise of Procedural Synthetic Data and Environments

Time: 04:00pm 

Venue: CB 308

Speaker(s): Prof. Jia Deng, Princeton University

Remark(s): 

Abstract

Data, especially large-scale labeled data, has been a critical driver of progress in AI. However, many important tasks remain starved of high-quality data. Synthetic data from computer graphics is a promising solution to this challenge, but still remains in limited use. This talk will present our work on Infinigen, a procedural generator designed to create unlimited high-quality 3D data and environments for computer vision and robotics. Infinigen is entirely procedural: every asset, from shape to texture, is generated from scratch via randomized mathematical rules. I will provide an overview of Infinigen along with its latest developments and applications. 

About the speaker

Jia Deng is a Professor of Computer Science at Princeton University. His current research interests include computer vision and robotics. He received his Ph.D. from Princeton University and his B.Eng. from Tsinghua University, both in computer science. He is a recipient of the Sloan Research Fellowship, the NSF CAREER award, and the ONR Young Investigator award. 

Title: The Promise of Procedural Synthetic Data and Environments

Time: 04:00pm 

Venue: CB 308

Speaker(s): Prof. Jia Deng, Princeton University

Remark(s): 

Abstract

Data, especially large-scale labeled data, has been a critical driver of progress in AI. However, many important tasks remain starved of high-quality data. Synthetic data from computer graphics is a promising solution to this challenge, but still remains in limited use. This talk will present our work on Infinigen, a procedural generator designed to create unlimited high-quality 3D data and environments for computer vision and robotics. Infinigen is entirely procedural: every asset, from shape to texture, is generated from scratch via randomized mathematical rules. I will provide an overview of Infinigen along with its latest developments and applications. 

About the speaker

Jia Deng is a Professor of Computer Science at Princeton University. His current research interests include computer vision and robotics. He received his Ph.D. from Princeton University and his B.Eng. from Tsinghua University, both in computer science. He is a recipient of the Sloan Research Fellowship, the NSF CAREER award, and the ONR Young Investigator award. 

Title: LLM based Zero Shot Speech Synthesis

Time: 02:00pm 

Venue: CB 308

Speaker(s): Dr. Shujie Liu

Remark(s): 

Abstract

With the rapid development of large language models (LLMs) in natural language processing, speech LLMs have also begun to receive increasing attention. In this talk, we will introduce VALL‑E, a zero‑shot text‑to‑speech (TTS) synthesis approach built upon large language models. Leveraging the in‑context learning capabilities of LLMs, VALL‑E can generate high‑quality, personalized speech using only a three‑second audio prompt from an unseen speaker. Building upon this foundation, we will further introduce several extensions of VALL‑E, including: VALL‑E X (the multilingual version), VALL‑E 2(addressing stability issues), PALLE(combining AR and NAR modelling), MELL‑E and FELLE (based on continuous speech representations).

About the speaker

Shujie Liu is a Principal Researcher at MSRA Hong Kong. His research focuses on natural language processing, speech processing, and machine learning. He has published over 100 papers in top-tier conferences and journals in NLP and speech, co‑authored the book Machine Translation, and contributed to Introduction to Artificial Intelligence. He has won multiple first‑place awards in international NLP and speech evaluation campaigns and has served as a reviewer and area chair for several major conferences. His research has been widely deployed in Microsoft products, including Microsoft Translator, Skype Translator, Microsoft IME, and Microsoft Speech Services.

Title: Towards Scalable Serverless LLM Inference Systems

Time: 10:30am 

Venue: CB 308

Speaker(s): Prof. Minchen Yu

Remark(s): 

Abstract

Serverless computing has become a compelling cloud paradigm for model inference due to its high usability and elasticity. However, current serverless platforms suffer from significant cold-start overhead---especially for large models---limiting their ability to deliver low-latency, resource-efficient inference. In this talk, I will present three systems we built for scalable serverless inference. First, Torpor proposes node-level GPU pooling that enables fine-grained GPU sharing and fast model swapping. Second, LambdaScale leverages high-speed interconnects to scale models across nodes and performs pipelined inference for lower latency. Third, for emerging large mixture-of-experts (MoE) models, we design fine-grained expert scheduling with elastic scaling to improve the cost-effectiveness of MoE inference.

About the speaker

Minchen Yu is an Assistant Professor at the School of Data Science, The Chinese University of Hong Kong, Shenzhen. He received his Ph.D. from Hong Kong University of Science and Technology. His research interests cover cloud computing and distributed systems, with a recent focus on serverless computing and machine learning systems. His research has been published at various prestigious venues, such as NSDI, ATC, EuroSys, INFOCOM, and SoCC, and has been applied in leading cloud platforms, such as Alibaba Cloud. He received the Best Paper Runner-Up Award at IEEE ICDCS 2021.

Title: Towards Consistent and Physically Plausible Visual Generation

Time: 11:00am 

Venue: CB 308

Speaker(s): Departmental Seminar by Prof. Jianfei Cai, Monash University

Remark(s): 

Abstract

Recent advances in large language models (LLMs) and multimodal large language models (MLLMs) have significantly enhanced the understanding and encoding of textual information. Leveraging these capabilities, a growing number of diffusion-based generative models have emerged for text-conditioned visual generation — spanning text-to-image, text-to-video, and text-to-3D tasks. While these models offer remarkable flexibility and produce increasingly realistic content, they still face fundamental challenges: aligning precisely with user intent, maintaining spatial, view, and temporal consistency, and adhering to the laws of physics. In this talk, I will present several recent research projects from my group that attacks these challenges. PanFusion enforces global consistency in text-to-panorama image generation; MVSplat360 uses image conditions and explicit 3D representation to enhance view consistency of 3D generation. VLIPP integrates physics-informed priors to ensure physically plausible text-to-video generation. I will conclude by pointing out the limitations and discussing future directions such as developing world models.

About the speaker

Jianfei Cai is a Professor at Faculty of IT, Monash University, where he had served as the inaugural Head for the Data Science & AI Department. Before that, he was Head of Visual and Interactive Computing Division and Head of Computer Communications Division in Nanyang Technological University (NTU). His major research interests include computer vision, deep learning and multimedia. He is a co-recipient of paper awards in ACCV, ICCM, IEEE ICIP and MMSP, and a winner of Monash FIT’s Dean's Researcher of the Year Award and Monash FIT Dean's Award for Excellence in Graduate Research Supervision. He serves or has served as an Associate Editor for TPAMI, IJCV, IEEE T-IP, T-MM, and T-CSVT as well as serving as Senior/Area Chair for CVPR, ICCV, ECCV, ACM Multimedia, ICLR and IJCAI. He was the Chair of IEEE CAS VSPC-TC during 2016-2018. He had served as the leading TPC Chair for IEEE ICME 2012, the best paper award committee chair & co-chair for IEEE T-MM 2020 & 2019, and the leading General Chair for ACM Multimedia 2024. He is a Fellow of IEEE.