講者介紹
下世代通訊
無線通訊
陳紹基
交通大學電子工程學系

近年來從事無線通信訊號傳收之研究,包括4G、5G、WiFi FFT IP設計、訊號同步、通道編解碼、新波形傳輸、通道估測、波束形成等。將針對FFT IP設計技術做較深入的討論。

下世代通訊
無線通訊
林大衛
交通大學電子工程學系

近年來從事LTE與IEEE 802.11p訊號傳收之研究,包括訊號同步、通道估計、預編碼等。也針對5G系統所擬採用的毫米波訊號傳輸進行研究。將摘要陳述以上研究的概況,並針對LTE與IEEE 802.11p的通道估計技術做較深入的討論。

下世代通訊
車用光學雷達
桑梓賢
交通大學電子研究所

運用單光子偵測器之光達(LiDAR)的測距功能 光達之測距功能係以測量光脈衝的飛行時間來達成。運用單光子偵測器的光達應有較佳的靈敏度,可以用較低的光量達成測距功能,但由於單光子偵測器極易受背景光干擾,且訊號量已經偏低,故正常工作環境下光達的訊噪比極不理想。我們針對光達工作環境,考慮光脈衝的強度機率模式,開發最大似然度(ML)的飛行時間估測方法。模擬以及初步測量結果顯示極佳的效果,未來有望成功用於車用3D景深描繪等用途。

下世代通訊
Statistical signal processing, digital communications, data modeling and analytics, optimization.
馮智豪
交通大學電子研究所

The success of signal processing increasingly relies on its linkage with other areas. It has become a basis for many applications because of its strong fundamentals. In particular, machine learning and digital communications have become more intertwined in recent years due to the imminence of a massively connected network, which allows for gathering (communications) AND processing (learning) of slew of information from distant locations. The gathering of data requires the data collecting agents to be communication ready, in which the accuracy and latency incurred during transmission are often mission critical to proper processing of the information collected. The processing of the collected data can no longer assumed to be done at a centralized location due to the diverse nature of data and constant improvement in processing power at the learning agents.
In the first half of this presentation, the emergence of heterogeneous wireless networks, which is seen as a paradigm shift from traditional cellular communications, and its role in supporting large amount of data collection will be highlighted. In particular, the importance of cell densification and the ensuing challenges will be discussed. The focus will then be shifted to data processing in which centralized learning and computational collective intelligence will be examined. Application such as autonomous driving will be discussed.

前瞻晶片
智慧監控、智慧穿戴式裝置、行動照護
李鎮宜
交通大學電子工程學系

結合智慧感測與數據分析,對於新興物聯網市場,將引發新一波的應用服務商機。此次的講題內容,將以醫療聯網的應用為案例,探討如何產生與收集一維到多維的生理數據,並透過深度學習模型與資料融合的新技術,取代傳統的人工檢測方案,進而研發智慧醫聯網裝置,有助於行動健康照護甚至個人化醫學的新興醫療服務與商機。

前瞻晶片
影像處理資訊安全
張錫嘉
交通大學電子研究所

While information security becoming an important issue, side-channel analysis (SCA) attacks, particularly power analysis attacks, are the biggest threats for Internet-of-Things (IoT) end nodes.  In this talk, we will illustrate how to build an Attribute Based Encryption (ABE) infrastructure for video streaming applications first.  The system will enforce access control and encrypt data using lightweight eSTREAM ciphers.  Second, we will illustrate how to build an ID-based system on end node like mobile phone device.  The system will directly use the phone-number or email address as user's public key.  Finally, we present a framework for side-channel evaluations which aims to unify the comparisons among different SCA attacks and countermeasures.

前瞻晶片
深度學習、影像辨識
張添烜
交通大學電子研究所

近年來深度學習類神經網路(CNN)在辨識、偵測以及相當多的電腦視覺應用上均有不錯的成果。但由於其具有高計算複雜度、大量待處理資料、以及高變異的網路結構特性,使得CNN在硬體實現上有相當大的困難。 本次提出一個支援完整網路的CNN IP generator。內部主要技術包含運行時可調整濾波器核心的結構來達到硬體使用率最大化,以及輸出優先策略使卷積層上的資料重複使用高達300到600倍,如此便可以降低資料頻寬。除此之外,針對考慮的網路結構,我們會根據設計上的限制來產生其最佳硬體和資料使用率的結果來實現適應網路變化以及即時的CNN加速器。

雲端感知/分析/處理
工業4.0技術服務應用領域、產線自動化、產線監測、雲端計算
陳添福
交通大學資訊工程學系

國立交通大學產業用物聯網(Industrial IoT)與資料智慧產研技術中心,係由科技部經費支持之深耕工業基礎技術專案計畫而成立,研究主軸在於工業4.0技術。本團隊由交通大學、台灣大學、中正大學、台灣科技大學等學研團隊共計9位教授所組成,同時配合研華科技的經費贊助與技術支援,以「環境感知終端與伺服閘道」、「產業資料分析工具與雲端平台」、「大數據與機械學習之產業應用」三方面前瞻技術,進行有關產業用物聯網所需技術開發,同時,搭配國家高速電腦中心之雲端計算服務,建構IIoT共同所需之核心平台方案(vertical platform solution)。 本實驗室主要任務在於發展Intelligent Edge Systems (gateway system in IIoT),旨在強化分散式gateway的資料前處置與分析能力,以發揮IoT edge computing的精神。本次發表會將分享深耕計畫研究成果,以docker系統建置之分散式gateway系統,提供容錯、分散計算能力,以加強產業用預防性維護之資料分析。同時,我們針對資料分析R Engine之記憶體管理進行優化,提供有效edge端資料分析的計算能力與效率。

雲端感知/分析/處理
IoT, Industry 4.0, Predictive Maintenance, Embedded System, Big Data, Data Analytics
曹孝櫟
交通大學資訊工程學系

Industrial Internet of Things (IIoT) is considered as a key and enabling technology for Industry 4.0. However, data analytics and services for IIoT are very challenging due to noisy data, tightly coupling between target applications and possible sensing technologies, domain specific knowledge, etc. In this talk, we first present our experiences of some field studies for IIoT data analytic services. Based on these experiences and experiments, we identify important research issues on power, vibration, temperature, and acoustic sensing, edge intelligences, and deep learning methods for IIoT services. Finally, we share our research results and present our future work.

雲端感知/分析/處理
先進駕駛輔助系統(ADAS)、自動駕駛系統(Self-driving)、前瞻智能監控系統(Intelligent Surveillance)
郭峻因
交通大學電子研究所

本次技術發表主題設定為ADAS應用下的深度學習技術,此技術為下一代ADAS系統的關鍵技術,主要內容涵蓋以深度學習技術來實現ADAS系統中的多重物件偵測,包含行人偵測、車輛偵測、機車/自行車偵測等,其偵測準確度在FPPI=1時可達90%,並針對Faster R-CNN類神經網路架構進行優化,在不影響偵測準確率下大幅提升系統執行效能,且以嵌入式系統(nVidia Jetson TX-1)作為技術展示平台,而本團隊也針對上述ADAS應用下的機器學習/深度學習技術所需要各式物件資料庫建置了超過40萬筆,並且持續增加中,為將來訓練各式深度學習模型提供充足的資料內容。

雲端感知/分析/處理
High performance computing on throughput processors (e.g. multimedia, Big Data, machine learning, etc).
賴伯承
交通大學電子研究所

通量處理器平台(如:GPGPU)已成為當前高效能運算中廣泛使用的運算加速器。將從探討此通量處理器的架構特性,及在高通量運算時的運算行為,與在記憶體架構中遇到的效能瓶頸,並針對此效能瓶頸,提出有效的解決方案。此外也將介紹本研究團隊目前的研發成果以及可能應用的領域,如大資料處理,訊號處理,機器學習,以及其他高通量運算能力的應用。

EDA前瞻研究與應用
IC Design, Package and Board Codesign
陳宏明
交通大學電子研究所

This talk will introduce the power network synthesis tool for 2D and 2.5/3D IC designs. This tool is a multi-year work with ITRI.

EDA前瞻研究與應用
EDA、IC測試
趙家佐
交通大學電子研究所

探討機器學習技巧在EDA與IC測試上之應用,其中會用power-network design自動化、晶片速度分級、WAT測試量測、以及製程監控上等不同應用來做實例分享。

EDA前瞻研究與應用
Design for manufacturability
江蕙如
交通大學電子工程學系

Multiple patterning lithography has been recognized as one of the most promising solutions, in addition to extreme ultraviolet lithography, directed self-assembly, nanoimprint lithography, and electron beam lithography, for advancing the resolution limit of conventional optical lithography. Multiple patterning layout decomposition (MPLD) becomes more challenging as advanced technology introduces complex coloring rules. Existing works model MPLD as a graph coloring problem; nevertheless, when complex coloring rules are considered, layout decomposition can no longer be modeled accurately by graph coloring. Therefore, for capturing the essence of layout decomposition with complex coloring rules, we model the MPLD problem as an exact cover problem. We then propose a fast and exact MPLD framework based on augmented dancing links. Our method is flexible and general: It can consider the basic and complex coloring rules simultaneously, and it can handle quadruple patterning and beyond. Experimental results show that our approach outperforms state-of-the-art works on reported conflicts and stitches and is promising for handling complex coloring rules as well.