Frontier of Artificial Network
A Series of Invited Talks @ FAN Group, CityU

Biography: Yang Cao is an Associate Professor at the Department of Computer Science, Institute of Science Tokyo (Science Tokyo, formerly Tokyo Tech), and directing the Trustworthy Data Science and AI (TDSAI) Lab. He is passionate about studying and teaching on algorithmic trustworthiness in data science and AI. Two of his papers on data privacy were selected as best paper finalists in top-tier conferences IEEE ICDE 2017 and ICME 2020. He was a recipient of the IEEE Computer Society Japan Chapter Young Author Award 2019, Database Society of Japan Kambayashi Young Researcher Award 2021. His research projects were/are supported by JSPS, JST, MSRA, KDDI, LINE, WeBank, etc.

Abstract: Large language models are often fine-tuned on small and sensitive datasets, where individual training examples can strongly influence model behavior and lead to memorization and privacy leakage. This talk focuses on differential privacy in LLM fine-tuning and explains how DP acts as a learning constraint that limits the influence of individual samples, thereby mitigating such risks. We introduce the core intuition behind DP without assuming prior background, review practical DP fine-tuning mechanisms, and discuss how privacy and utility should be evaluated together. We also clarify what DP fine-tuning can and cannot protect, and outline open challenges in privacy-aware LLM adaptation.

Biography: Jiancheng Yang is a Principal Investigator at the ELLIS Institute Finland and Assistant Professor at Aalto University. He received his Bachelor's and PhD degrees from Shanghai Jiao Tong University, was a visiting fellow at Harvard, and completed postdoctoral training at EPFL. His research advances AI for health, with a focus on spatial intelligence, generative AI, and multimodal deep learning. He has authored 60+ publications and is recognized for developing MedMNIST. He is a Forbes 30 Under 30 honoree and a recipient of the WAIC Yunfan Award.

Abstract: The next step for medical AI is scaling. However, scaling medical AI faces intrinsic challenges: medical data are often limited, complex, and highly heterogeneous, while many existing AI techniques are not yet designed to fully exploit such data. This talk presents contributions unified by a common goal: building scalable medical AI through maximal reuse. First, I introduce LeFusion, a constrained generative framework for medical data synthesis (ICLR 2025 Spotlight), which provides guarantees on synthetic data quality and allows downstream models to better exploit underutilized information in limited real datasets. Second, I describe DiffAtlas, a new perspective on multimodal medical data (MICCAI 2025 Spotlight), where traditional input–output medical image segmentation is reformulated as an unconditional multimodal diffusion model guided only at inference time. Finally, I introduce PlaneCycle, a training-free, adapter-free operator that lifts arbitrary 2D foundation models into 3D by cyclically distributing spatial aggregation across orthogonal planes, adding no parameters while preserving pretrained inductive biases. Together, these works suggest a broader principle: medical AI can scale without scaling cost.