中心名稱

裝置端之智慧晶片系統研究中心

AI-SoC CoE for

Scalable and Intelligent Edge Computing

願景

延續中心過去頂尖的研究成果，以及結合國內現有的半導體與資通訊的製造優勢，研發智慧終端裝置所需求的關鍵人工智慧晶片與系統應用，並透過學術論文的發表與創新產業智慧方案的提供，使國立交通大學成為全球人工智慧晶片的學術研究與新興系統應用的重鎮。

研究重點簡述

1. 透過機器學習最新技術，發展裝置端之智慧晶片與系統，導入監督式學習、非監督式學習及強化式學習，進行資料分析、分群、回歸、生成及分類等工作並解決裝置端與雲端運算中各式各樣的資料處理問題，研究重點還包括最新深度學習之技術開發，期望獲得系統最佳學習效能。

2. 結合跨領域的合作團隊，探討邊緣計算與智能系統所需求的新電子元件、節能電路、平行運算、學習演算法、到特定的系統與應用，尤其採用國內特有的半導體製造優勢（異質整合與高階製程），來驗證所提的人工智慧運算架構與提供特定場域所需求的有效數據集及深度學習模型。

最後藉由實際應用載具（自動駕駛輔助系統及快速醫學檢測系統）與業界合作成功案例，來驗證所提關鍵技術的特色及優勢，同時提供場域需求的智慧分析解決方案，尤其強調在終端裝置對於有效數據集的取得（採用0.35um/90nm製程的異質整合設計有效數據感測與分析晶片）與精準分析的掌握（採用16nm的FIN-FET製程設計人工智慧分析晶片與系統）。

中心成員

計畫主持人:黃威 李鎮宜

Group	Major Members	Research Content
分項一： 突觸元件	侯拓宏、李佩雯、張錫嘉	both Resistive-RAM and Charge-Trapped Transistor (CTT) will be explored and potential applications to neuromorphic computing will be further investigated. This study will provide some new approaches and potential solutions to solve memory access bottleneck in current memory-bus-computing architecture, and in the meantime, enhance energy-efficiency in battery-powered devices.
分項二： 智慧感測電路	陳科宏、董蘭榮、周世傑	valid datasets are quite important for model training and inference. As a result, exploiting signal processing and model-based techniques for valid datasets will be investigated in the front-end sensing and readout circuits. In addition, power consumption is a very crucial issue in sensing modules. This study will also exploit event-driven approach to enhance energy efficiency in data generation.
分項三： 節能與平行運算	張添烜、賴伯承、陳添福	data-driven system architecture dealing with data transfer between memory and computing kernels has to be further investigated, especially for deep models. In addition, distribution of datasets to reduce communication avoidance overhead has to be explored for large volume of datasets. Finally energy-efficient hardware accelerators have to be considered as well to meet both real-time inference and power constraints in edge devices.
分項四： 深度學習理論與演算法	簡仁宗、杭學鳴、盧鴻興	algorithms suitable for edge-cloud computing have to be developed. For edge site, we have to consider the constraint of both storage and computing capability with acceptable accuracy. For cloud site, we have to consider the communication bandwidth and latency between devices and clouds. As a result, a hybrid model consisting of shallow and deep models should be developed to meet emerging IoT applications.
分項五： 自駕車輔助系統(ADAS)	郭峻因、劉志尉、王聖智	this test vehicle is selected to evaluate the outcome from the previous 4 tasks. Moreover data fusion will be investigated to enhance inference accuracy to meet ADAS requirements, such as detection accuracy, path planning, obstacle avoidance, hardware resources, power budget, …etc. A demonstrator for field trial will be built up to evaluate the proposed solutions at NCTU campus.
分項六： 快速醫學檢測系統(HTCMS)	王崇智、王雲銘、黃俊達、何宗易、何盈杰、徐武輝	based on the recent results of Field-Programmable Lab-on-a-Chip (FPLoC), some fast medical tests are currently under investigation. The next goal is shooting for cell detection, especially CTC. Major research tasks include smart sensing, microfluidic architecture, and learning/detection models. A demo system based on CMOS bio chip will be developed for clinical trial.