2026年7月23日（周四）下午两点半，深圳河套学院荣幸地邀请到意大利维罗纳大学创新医学工程系的马尔科・克里斯塔尼教授，在B401教室带来主题为“从可解释性到表达能力：多年来多模态数据生成建模研究”的学术报告分享。诚挚欢迎学院师生及对该领域感兴趣的产学研界同仁参与讲座。

Title

从可解释性到表达能力：

多模态数据生成建模研究

From Interpretability to Expressiveness:

Generative Modeling for Multimodal Data Through the Years

About the Speaker

马尔科・克里斯塔尼 教授
Prof. Marco Cristani

Marco Cristani教授现就职于意大利维罗纳大学创新医学工程系。2025至2026年在冰岛雷克雅未克大学担任访问教授，2026至2027年受聘为该校计算机科学系荣誉教授。Marco Cristani教授是意大利国家研究委员会（CNR）特伦托分部应用本体实验室副研究员，同时兼任热那亚意大利技术研究院（IIT）模式分析与计算机视觉学部外聘合作研究员。于2002年取得维罗纳大学计算机科学硕士学位，2006年获得该校计算机科学博士学位。其研究方向涵盖人工智能、计算机视觉、机器学习、生成模型、异常检测、社会信号处理、人类行为分析以及人工智能工业落地应用。

Marco Cristani教授担任智能实验室（INTELLIGO Labs）的负责人，领导包括22 名成员的科研团队。担任维罗纳大学工程教学委员会副主席，同时牵头设计并负责2024至 2025学年新开设的机器人与智能系统计算机工程本科专业。此外，他隶属于意大利国家人工智能博士项目工业4.0方向，也是欧洲学习与智能系统实验室（ELLIS）成员。

Marco Cristani教授已指导135篇本科毕业论文、52篇硕士学位论文及21篇博士学位论文。学术成果包含2本主编专著、59篇期刊论文、180篇会议论文与13篇书籍章节。谷歌学术引用量超14600次，H指数达53。他是国际模式识别协会会士，曾担任CVPR、ECCV、IJCAI、NeurIPS等顶会领域主席，并联合创办维罗纳大学两家衍生企业：Humatics s.r.l.与Qualyco s.r.l.。

Swipe up for EN version

Marco Cristani is Full Professor of Computer Engineering at the Department of Engineering for Innovation Medicine, University of Verona, Italy. He is also Honorary Professor at the Department of Computer Science, Reykjavik University, Iceland, for 2026-2027, following a Visiting Professorship at Reykjavik University in 2025-2026. He is an Associate Member of the National Research Council of Italy, CNR, Trento, within the Laboratory for Applied Ontology, and an External Collaborator of the Italian Institute of Technology, IIT, Genoa, within the Pattern Analysis and Computer Vision department.He received his Ph.D. in Computer Science in 2006 and his M.Sc. in Computer Science in 2002, both from the University of Verona. His research focuses on artificial intelligence, computer vision, machine learning, generative models, anomaly detection, social signal processing, human behavior analysis, and industrial applications of AI.

He is the founder and coordinator of INTELLIGO Labs, a 22-member research group, Vice Chair of the Engineering Teaching Board at the University of Verona, and designer and coordinator of the B.Sc. program in Computer Engineering of Robotic and Intelligent Systems, launched in 2024/25. He is also a member of the National Ph.D. Program in Artificial Intelligence, Industry 4.0 specialization, and of ELLIS, the European Laboratory for Learning and Intelligent Systems.

He has supervised 135 B.Sc., 52 M.Sc., and 21 Ph.D. theses. His publication record includes 2 edited books, 59 journal papers, 180 conference papers, and 13 book chapters. According to his CV, he has collected over 14,600 Google Scholar citations, with an h-index of 53. He is an IAPR Fellow and has served as Area Chair for major conferences including CVPR, ECCV, IJCAI, and NeurIPS. He co-founded the University of Verona spin-offs Humatics s.r.l. and Qualyco s.r.l.

Abstract

过去三十年间，生成模型发生了翻天覆地的变革。早期概率生成方法虽易于检视与解释，但对真实数据多样分布的表征能力存在局限。近年来出现的变分自编码器（VAE）、生成对抗网络（GAN）、扩散模型等模型大幅提升了生成效果，却也让底层运行机制更难分析与调控。本次报告将顺着这一发展脉络，围绕模型可解释性与表达能力之间的权衡关系展开论述。报告首先回顾经典生成模型，随后介绍现代分层架构——这类架构依靠多层隐变量逐步生成数据。该分层结构为研究两大问题提供了有效思路：信息是如何在生成流程的不同阶段注入模型，以及不同层级的参数变化会如何影响最终生成结果。报告最后从互信息与率失真理论的视角，探讨多隐变量生成模型的前沿研究成果；该理论视角为构建兼具强表达能力与高可控性的生成模型指明了可行路径。本次分享同时梳理了多模态人工智能、计算机视觉、生物医学影像及应用机器学习领域广阔的交叉合作研究方向。

Swipe up for EN version

Generative models have changed substantially over the last three decades. Early probabilistic approaches were relatively easy to inspect and explain, but had limited ability to represent the variability of real data. More recent models, including VAEs, GANs and diffusion models, have greatly increased generative quality, while making the underlying mechanisms harder to analyse and control. This talk follows this trajectory, focusing on the trade-off between interpretability and expressiveness. It first revisits classical generative models, then moves to modern hierarchical architectures in which data are generated progressively through multiple latent variables. This structure provides a useful way to study how information is introduced at different stages of the generation process, and how changes at different levels affect the final output. The last part of the talk discusses recent work on multiple-latent-variable generative models from the perspective of mutual information and rate-distortion theory. This view suggests possible routes towards generative models that are both expressive and more controllable. The talk will also outline broad opportunities for collaboration in multimodal AI, computer vision, biomedical imaging and applied machine learning.

Host

松井智子 教授

Prof. Tomoko Matsui

Date & Time

2026年7月23日（星期四）

下午14:30-16:00

July 23, 2026, Thursday,

2:30pm-4:00pm

Venue

深圳河套学院B401教室

（深圳市福田区福保街道红棉路6号，地图导航“深圳河套学院-南门”）

Room B401, Shenzhen Loop Area Institute

(6 Hongmian Rd, Fubao Sub-Street, Futian District, Shenzhen, navigate to "Shenzhen Loop Area Institute (South Gate)" on the map.)

Online link

扫码加入会议

Join the Meeting Online

腾讯会议号：948-1135-4043
Tencent Meeting：948-1135-4043