讲座题目 | When Views Disagree: Trustworthy AI with Bayesian Uncertainty Estimation for Multimodal Learning | ||
主讲人 (单位) | 杜岚 (蒙纳士大学) | 主持人 (单位) | 胡艺馨 (h片 ) |
点评人 (单位) | 厉雨婷 (h片 ) | ||
讲座时间 | 2026年6月11日(周四)上午10:00 | 讲座地点 | 经管楼B208 |
主讲人简介 |
杜岚博士现任澳大利亚蒙纳什大学数据科学与人工智能副教授,同时为IEEE高级会员(Senior Member)。其研究方向主要集中于机器学习、深度学习与自然语言处理的交叉领域,重点包括主动学习、不确定性估计、知识蒸馏以及分子生成等前沿方向。同时,他长期致力于推动人工智能技术在公共卫生、市场营销、生物化学等跨学科领域中的应用与落地。杜博士已在机器学习、自然语言处理与数据挖掘领域的国际顶级会议与期刊发表高质量学术论文100余篇,包括 NeurIPS、ICML、ACL、EMNLP、AAAI、TPAMI、IJCV、TMM 和 TNNLS 等,在相关研究领域具有广泛影响力。此外,他现担任《Journal of Artificial Intelligence Research(JAIR)》与《ACM Transactions on Probabilistic Machine Learning(TOPML)》副主编,《Machine Learning》和《Big Data Research》期刊编委,并多次担任 AAAI 与 ACL 等国际顶级学术会议的领域主席(Area Chair)。作为首席研究员(Chief Investigator),杜博士已成功获得多项澳大利亚政府科研资助,包括1项 NHMRC Ideas Grant、1项 ARC Discovery Project(DP)以及2项 MRFF 基金项目,总研究经费累计约500万澳元。 | ||
讲座内容摘要 | Modern AI systems increasingly make predictions using multiple sources of information, such as text, images, tabular records, and sensor measurements. While combining these sources can improve predictive performance, it also creates new risks: individual views may be noisy, sources may conflict, and deployment data may include categories that were never observed during training. A model can therefore produce a confident prediction precisely when caution is needed most.This seminar presents a research program for trustworthy multi-view and multimodal learning through uncertainty estimation. We will discuss Bayesian aggregation using multi-view Gaussian processes, scalable uncertainty estimation with multimodal neural processes, trust-aware fusion for resolving conflicting views, and open-set learning methods that calibrate ambiguity while reducing view-specific biases. The emphasis will be on the underlying ideas, practical implications, and evaluation principles rather than technical details alone. | ||

