08/09/2026
【IEEE Distinguished Talk】成功大學李國君教授專題演講：Machine Learning for Analytics Architecture: AI to Design AI
各位敬愛的教授、同仁與同學：
您好！
陽明交大電機學院非常榮幸邀請到成功大學電機工程學系 李國君 教授（Prof. Chris Gwo Giun Lee） 蒞臨本校進行專題演講。
李教授將針對輕量化 AI、軟硬體/演算法架構協同設計（Algorithm/Architecture Co-design）以及嵌入式系統/晶片系統（SoC）映射等關鍵前沿技術進行精彩分享。
【演講資訊】
• 演講題目： Machine Learning for Analytics Architecture: AI to Design AI
• 演講者： 李國君 教授 (Prof. Chris Gwo Giun Lee), 成大電機 Department of Electrical Engineering, National Cheng Kung University)
• 時 間： 2026 年 9 月 15 日 (二) 10:30 - 11:45
• 地 點： 陽明交大工程四館 108 室 (知新廳)
報名網址： https://forms.gle/AsSM3xcsaPhdyj2y9
【Abstract】 Recently, lean or lightweight AI, in real world applications ranging from cloud to edge, has witnessed transformative realizations and hence gained significant attention and popularity. Researchers and engineers from both academia and industry have thus been addressing the CHALLENGES in harnessing the prowess unleashed by AI. These exploration of new disruptive OPPORTUNITIES include flexible deployment of low-complexity AI algorithms onto high-performance computing platforms with high-accuracy and high-energy efficiency.
As such, based on the vertical integration design methodology this talk addresses how lightweight algorithms, based on software/hardware co-design and algorithm/architecture co-design (AAC) are mapped onto embedded systems or System-on-Chip (SoC) in AI. When crossing or traversing design spaces from algorithmic functionality to potentially synthesizable microarchitecture designs, algorithmic intrinsic complexity measures are characterized by potential computing in parallel, efficient data storage and data transfer rate. These platform independent features are extracted from different dataflow models using graph theory-based analytics algorithm, during joint exploration of algorithm and architecture co-design space. In this cross-level-of abstraction topic discussion, a case study in high level synthesis is introduced. In addition, a lightweight mobile edge AI for skin cancer detection with two layers CNN at 97% recognition rate, using limited training data, will also be discussed.
