报告时间：10月30日 10:30

报告地点： 数学馆201

报告名称：AI Flow: Towards Ubiquitous Edge Intelligence with Large Models

报告摘要:

This talk explores how large AI models and communication systems can be co-designed to enable ubiquitous, low-latency intelligence at the network edge. In the first part, we present task-oriented communication for edge inference. Rather than transmitting or reconstructing raw sensor data, edge devices extract task-relevant features and communicate only what is needed for downstream inference. Guided by the information bottleneck principle, we jointly optimize feature extraction and the codec to balance communication cost against inference accuracy. In the second part, we turn to collaborative inference, detailing methods that accelerate inference by distributing computation across devices and edge servers. We focus on split inference and speculative decoding for large language models, and discuss practical co-design strategies in parameter placement, compute offloading, and feature extraction. Together, these advances address the tension between the high computational demands of large models and the constrained resources of low-end devices, charting a path toward scalable, efficient, and reliable edge intelligence.

报告人简介:

Jiawei Shao is a research scientist at the Institute of Artificial Intelligence (TeleAI), China Telecom. He is a principal investigator leading the AI Flow group. His research focuses on a wide range of topics, including large language models, edge AI, generative AI, and agentic AI. He received his Ph.D. from the Hong Kong University of Science and Technology (HKUST). He is a recipient of several awards, including the IEEE Communications Society Katherine Johnson Young Author Best Paper Award and the HKUST SENG PhD Research Excellence Award. He has published more than 30 research papers in top-tier journals and conferences, including Nature Communications and Nature Machine Intelligence. He is selected for inclusion in the World’s Top 2% Scientists list for single-year impact in 2024.