1,721,051 research outputs found
A Deeply Pipelined CABAC Decoder for HEVC Supporting Level 6.2 High-tier Applications
Get PDFHigh Efficiency Video Coding (HEVC) is the latest video coding standard that specifies video resolutions up to 8K Ultra-HD (UHD) at 120 fps to support the next decade of video applications. This results in high-throughput requirements for the context adaptive binary arithmetic coding (CABAC) entropy decoder, which was already a well-known bottleneck in H.264/AVC. To address the throughput challenges, several modifications were made to CABAC during the standardization of HEVC. This work leverages these improvements in the design of a high-throughput HEVC CABAC decoder. It also supports the high-level parallel processing tools introduced by HEVC, including tile and wavefront parallel processing. The proposed design uses a deeply pipelined architecture to achieve a high clock rate. Additional techniques such as the state prefetch logic, latched-based context memory, and separate finite state machines are applied to minimize stall cycles, while multibypass- bin decoding is used to further increase the throughput. The design is implemented in an IBM 45nm SOI process. After place-and-route, its operating frequency reaches 1.6 GHz. The corresponding throughputs achieve up to 1696 and 2314 Mbin/s under common and theoretical worst-case test conditions, respectively. The results show that the design is sufficient to decode in real-time high-tier video bitstreams at level 6.2 (8K UHD at 120 fps), or main-tier bitstreams at level 5.1 (4K UHD at 60 fps) for applications requiring sub-frame latency, such as video conferencing
Emotional Context Influences Micro-Expression Recognition
No full text
Micro-expressions are often embedded in a flow of expressions including both neutral and other facial expressions. However, it remains unclear whether the types of facial expressions appearing before and after the micro-expression, i.e., the emotional context, influence micro-expression recognition. To address this question, the present study used a modified METT (Micro-Expression Training Tool) paradigm that required participants to recognize the target micro-expressions presented briefly between two identical emotional faces. The results of Experiments 1 and 2 showed that negative context impaired the recognition of micro-expressions regardless of the duration of the target micro-expression. Stimulus-difference between the context and target micro-expression was accounted for in Experiment 3. Results showed that a context effect on micro-expression recognition persists even when the stimulus similarity between the context and target micro-expressions was controlled. Therefore, our results not only provided evidence for the context effect on micro-expression recognition but also suggested that the context effect might result from both the stimulus and valence differences.</p
Emotion Recognition from Physiological Sensor Data - Learning and Applications
No full text情緒在人際溝通上與服務的提供上，是個常被遺忘的對象；但近年來由於溝通互動和個人化服務的迅速發展，情緒漸漸的被視為一個考慮、討論以及研究的對象。在心理學以及哲學的領域中，情緒算是較早被提出和被研究的，一些情緒的功能像是決策決定、學習上已證明有明確的影響與功能。
以前人之研究作為基石，本研究乃採用人體的胸腔擴張程度、肌膚導電度、肌膚表面溫度、以及血氧濃度等生理數值作為訊號來源，進而辨識喜、怒、哀、懼等四種情緒狀態，並透過較為隱蔽的表達方式展現結果，以利用這些資訊來幫助人際溝通與互動。在情緒的辨認部份，本文探討了情緒與生理訊號之關係，以及部分前人之研究情形，從而提出我的解決方式。除此之外，情緒資訊之應用亦為本文的另一個重點。藉由此珍貴的資訊，在此也提出了應用的雛形Cura，透過隱諱但隨手可得的方式，告知週遭的人自己的情緒狀態，以協助人際溝通與互動。Emotion as a concept is usually forgotten in service providing and interactive communication. As content service and communication develop, emotion becomes more and more important in these research fields. In this research, I use bio-sensors to collect physiological signals from human subjects in different emotion states. The signals include sensor data from blood volume pulse, skin conductance, skin temperature, and respiration. I extract the features from these signals, and then use support vector machine to learn a classifier. The recognition rate of emotion is about 97%.
Furthermore, a prototype application Cura is made. Cura is an ambient cube which shows emotion states depending on the recognition of emotion. It is a media to tell one's closer the emotion with privacy-concern. Cura is simple and nature so that human transmit the emotion directly. Furthermore, Cura also keeps the private of the emotion information. It leaves the decision to user themselves with who he/she shares the emotion.Acknowledgments ii
Abstract v
List of Figures xii
List of Tables xiv
Chapter 1 Introduction 1
1.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3
1.2 Research Objectives . . . . . . . . . . . . . . . . . . . . . . . . . . 4
Chapter 2 Related Work 5
2.1 View from Psychology . . . . . . . . . . . . . . . . . . . . . . . . 6
2.1.1 Emotion Models . . . . . . . . . . . . . . . . . . . . . . . . 6
2.1.2 Characteristics and Roles of emotion . . . . . . . . . . . . 9
2.2 Recognition in Artificial Intelligence . . . . . . . . . . . . . . . . . 9
2.2.1 Emotion Recognition . . . . . . . . . . . . . . . . . . . . . 10
2.2.2 Physiological Signals . . . . . . . . . . . . . . . . . . . . . 13
2.2.3 Learning Methods . . . . . . . . . . . . . . . . . . . . . . . 18
2.3 Relevance of Emotion Research for Affective Computing . . . . . 22
2.3.1 Evaluation . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
2.3.2 Health Care . . . . . . . . . . . . . . . . . . . . . . . . . . 23
2.3.3 Education and Tutor System . . . . . . . . . . . . . . . . . 24
Chapter 3 Problem Definition 26
3.1 System Overview . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
3.2 System Focus . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
3.2.1 Sensor Data from Physiological Signals . . . . . . . . . . . 29
3.2.2 Emotion States . . . . . . . . . . . . . . . . . . . . . . . . 31
3.2.3 Hypothesis . . . . . . . . . . . . . . . . . . . . . . . . . . . 31
3.3 Proposed Solution . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
Chapter 4 Experiment 36
4.1 Mood Induction and Data Collection . . . . . . . . . . . . . . . . 37
4.2 Feature Extraction and Learning . . . . . . . . . . . . . . . . . . 38
4.3 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 40
4.3.1 The Result of Emotion Recognition . . . . . . . . . . . . . 41
4.3.2 The Result from the Decision Tree . . . . . . . . . . . . . 42
4.3.3 Importance of Sensor Data . . . . . . . . . . . . . . . . . . 43
4.3.4 Cross Person Recognition Result . . . . . . . . . . . . . . 44
4.3.5 Summary . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
Chapter 5 Application 47
5.1 The Idea . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
5.2 Cura . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5.2.1 Structure of Cura . . . . . . . . . . . . . . . . . . . . . . . 49
5.2.2 Patterns in Cura . . . . . . . . . . . . . . . . . . . . . . . 50
5.2.3 Result . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
5.3 Music Service Providing . . . . . . . . . . . . . . . . . . . . . . . 51
Chapter 6 Conclusion 53
Bibliography 56
Appendix A Experiment Design Document 61
A.1 Preparing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 61
A.2 Methodology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
A.2.1 Equipment Set Up . . . . . . . . . . . . . . . . . . . . . . 62
A.2.2 Overview of the Whole Process . . . . . . . . . . . . . . . 62
A.2.3 Detail Description . . . . . . . . . . . . . . . . . . . . . . . 64
Appendix B Questionnaire for Mood Induction 70
B.1 Self Evaluation Form for Sad . . . . . . . . . . . . . . . . . . . . . 70
B.2 Self Evaluation Form for Anger . . . . . . . . . . . . . . . . . . . 71
B.3 Self Evaluation Form for Fear . . . . . . . . . . . . . . . . . . . . 71
B.4 Self Evaluation Form for Happy . . . . . . . . . . . . . . . . . . . 72
B.5 Self Evaluation Form for Personal Information . . . . . . . . . . . 7
Kinetic Energy Partition Method for Quantum Systems with Competing Potentials
No full text對於競爭型位能的量子力學系統，傳統上都是用微擾理論的級數近似解，但有可能此級數無法收斂，諸如此類問題，本研究室發展基於分離能量的全新方法，而非像是微擾理論的分解位能項。此方法是分離動能項裡的質量，我們稱此方法為分離動能法(Kinetic Energy Partition (KEP))，此方法不像是按照傳統方式把位能分成""微擾項""與""非微擾項""，因為在競爭型位能下，無法區分哪個位能為""微擾項""與""非微擾項""。分離後的動能項分別對應一個位能使形成子漢米爾頓系統，所以此系統的總漢米爾頓可拆解成個別的子系統漢米爾頓相加，總波函數可以由個別子漢米爾頓系統所對應的波函數線性疊加而成。 直覺上可以先把質量都分解成兩個正質量的動能項，此結構比較符合物理，在此我們有利用五個例子來驗證，把動能項裡的質量拆解成兩個都是正質量的粒子是否可行，利用分離動能法所得到的能量與波函數，會與正合解的能量與波函數作為比較。五個例子分別為:雙負delta函數位能、三負delta函數位能、簡諧振子在無限深位能井、氫原子模型、一維史塔克效應，利用分離動能法可以精準解出，簡諧振子在無限深位能井的基態與第一激發態能量，比某些微擾展開方法解來得更精準，氫原子模型與一維史塔克效應，利用分離動能法則須搭配某些邊界條件，則可解出與正合解基態能量誤差不到5%的精準度，對於兩個或兩個以上的競爭型負delta函數來說，分離動能法在delta函數脈衝較小情況下也是精準的。 按照數學上分析，也可以把質量拆解成一個正質量與一個負質量，只要質量總合不變即可，負質量項的動能項搭配一項位能使形成子漢米爾頓系統，在某些位能下，漢米爾頓系統沒有束縛解，但是因為負質量關係，使得位能函數顛倒(對x軸鏡射)，而得到束縛解，我們利用四個例子來驗證:反對稱delta函數位能、正delta函數在無限深位能井、氦原子模型、一維氦原子，delta函數位能因為波函數簡單，且只有一個能量即為基態能量，對於KEP方法來說相當簡單，對於正delta函數位能是沒有束縛解，但利用分離動能法使得其中一項動能為負，得到E>0之束縛態。對於氦原子來說，是個最簡單的多體物理模型，也是許多多體理論裡的基礎例子，氦原子中有一項庫倫排斥能ke^{2}/r_{12}，因庫倫排斥能沒有束縛解，則可以利用負質量只得位能變為負值，是得能量有束縛解，利用分離動能法各個子漢米爾頓系統的波函數各取一個(基態波函數)，再利用橢圓座標轉換，可計算出氦原子基態能量為-70.2541eV，與實驗值-79.0143eV相差不遠，利用最佳化方法可得到基態能量為-78.9985eV則與實驗值非常接近。一維氦原子為簡化雙座標系統積分的近似，也可提供一維化學實驗的理論背景，而我們也利用KEP方法分析一維氦原子。For quantum systems with competing potentials, the conventional perturbation theory often yields an asymptotic series and the subsequent numerical outcome becomes uncertain. To tackle such kind of problems, we develop a general solution scheme based on a new energy dissection idea. Instead of dividing the potential energy into ""unperturbed"" and ""perturbed"" terms, a partition of the kinetic energy is performed. By distributing the kinetic energy term in part into each individual potential, the Hamiltonian can be expressed as the sum of the subsystem Hamiltonians with respective competing potentials. The total wavefunction is expanded by using a linear combination of the basis sets of respective subsystem Hamiltonians. We first illustrate the solution procedure using a simple system consisting of a particle under the action of double delta -function potentials and triple delta -function potentials . Next, this method is applied to harmonic oscillator in the infinite potential box , the hydrogen atom and 1D Stark effect. Compared with the exact solution, this new scheme converges much faster to the exact solutions for both eigenvalues and eigenfunctions. When properly extended, this new solution scheme can be very useful for dealing with strongly coupling quantum systems. In the mathematical allowed, we can shall dividing mass with ""negative"" term and ""positive"" term. Then the total Hamiltonian is conserved. Some kind of Hamiltonian system didn''t have discrete energy solution. But introduce with negative mass, we can regard as the potential energy inverse of x-axes. So the Hamiltonian system can be solved. We have 4 case to illustrate this method. The first case is antisymmetry delta -function potentials, this is a simply case to explain how to apply ""negative"" mass in the quantum system. The second case is ""positive"" $delta -function potential in the infinite potential box. The third case is Helium atom, we used negative mass to solved the electron repulsive potential energy. Optimize Helium system position we get -78.965eV Helium ground state energy, just 0.02% error with experimental value. The last case is 1D helium atom, it is a mathematical model. This model can reduce the double system integration, so there is a very persuasive example
Design and Implementation of CMOS Wireless Receiver Circuits for 24-GHz Communication Systems
No full text隨著對無線通訊頻寬需求的升高，低成本高效能的微波積體電路的發展也漸趨熱門。傳統上，由於較好的元件特性，高頻積體電路多利用III-V半導體製程和SiGe BiCMOS製程實現。CMOS製程擁有低成本及高度整合性的優勢，似乎非常適合作此種電路的應用。然而，由於嚴重的基板損耗及較差的高頻特性，利用CMOS製程來實現這樣的系統，仍舊是項挑戰。因此，本論文將描述CMOS製程在設計高頻電路時的一些設計考量，接著提出不同的電路技巧來實現。
首先，本論文將介紹一應用於24-GHz之0.18-µm CMOS 傳送端/接收端切換器。為了提高關閉時的隔離度，提出ㄧ增加隔離度之電路技巧，藉由增加一額外開關，可有較佳的隔離度及合理的穿透損耗。此外，採用阻抗轉換的技巧來提升功率特性。量測結果顯示此單刀雙擲切換器在24-GHz可達到2.5-dB的穿透損耗及20-dB的隔離度，同時輸入及輸出反射損耗皆大於14-dB。且因採用阻抗轉換技巧，此切換器之輸入P1dB為24.2-dBm。
接著，本論文提出ㄧ應用於24-GHz之接收機架構。採用CMOS 0.18-µm製程，此接收機前端電路整合了低雜訊放大器、混波器及中頻放大器而擁有良好的轉換增益及低雜訊表現。量測結果顯示，在中頻頻率為4.82-GHz的情況下，此電路擁有28.4-dB的轉換增益及6-dB的雜訊，同時輸入反射損耗大於14-dB；而輸入P1dB及IIP3分別為-23.2及-13.0-dBm。在1.8-V的工作電壓下，功率消耗為54mW。
最後介紹一個適用於之前提出之24-GHz接收機架構之本地信號產生器。該電路由ㄧ19-GHz低相位雜訊壓控振盪器、除四除頻器、及相位調整器構成。藉由雜訊濾除技巧及電容回授以增加gm，可有效降低振盪器相位雜訊。此外，信號經除四除頻器及相位調整器調整後，可提供精準的四相位信號。As the demands for wider bandwidth in wireless communication continue to expand, the development of low-cost and high-performance monolithic microwave integrated circuits (MMICs) have attracted great attention. Conventionally, the high-frequency integrated circuits were realized by III-V compound semiconductor and SiGe BiCMOS process due to their superior device performance. Owing to unparalleled advantages in both fabrication cost and denser integration, the CMOS technology appears to be well suited to implement these designs. However, it still remains a challenge for such RF building blocks to be integrated in a standard CMOS process due to its significant substrate losses and inferior high-frequency properties of the active devices. Consequently, this thesis describes the design consideration of the CMOS process for the individual building blocks then presents several novel circuit techniques to reach the requirements for the system designs.
First, a transmit/receive (T/R) switch for the 24-GHz ISM band application implemented in a standard 0.18-m CMOS process is demonstrated. A technique to enhance the isolation performance is proposed for the T/R switch. Reasonable isolation and insertion loss are achieved with an additional switch. Besides, impedance transformation networks (ITNs) is utilized to reduce the source and load impedance seen from the switch to increase the power handling capability. The fabricated SPDT (single-pole double-throw) switch at 24-GHz exhibits an insertion loss of 2.5 dB and an isolation of 20 dB while maintaining an input and output return loss better than 14 dB and 19 dB, respectively. Owing to the use of the ITNs, the measured Pin-1dB of the switch is 24.2 dBm.
Secondly, a receiver architecture is presented for operations at the 24-GHz frequency band. By monolithically integrating the LNA, the down-conversion mixer and the IF amplifiers, the receiver frontend is realized in a 0.18-m CMOS process, exhibiting enhanced circuit performance in terms of conversion gain and noise figure. With an IF frequency of 4.82-GHz, the fabricated circuit demonstrates a conversion gain of 28.4 dB and a noise figure of 6.0 dB while maintaining an input return loss better than 14 dB. The measured Pin-1dB and IIP3 of the receiver frontend are -23.2 and -13.0 dBm, respectively. The power consumption for the receiver frontend is 54mW at a supply voltage of 1.8 V,.
Finally, a signal generator which provides the required LO outputs for the proposed receiver architecture is presented. The proposed LO generator consists of a 19-GHz low-phase-noise voltage-controlled oscillator (VCO), a 4:1 frequency divider and a quadrature phase tuning circuit. With a noise filtering technique and gm-boosting capacitive feedback, the VCO exhibits low phase noise with adequate frequency tuning range. Besides, an interpolation technique, which is originally developed for ring oscillators, is adopted for the phase tuning the quadrature signal provided by the frequency divider to obtain accurate phases.Table of Contents
Chapter1 Introduction 1
1.1 Motivation 1
1.2 Overview of the thesis 2
Chapter2 RF Receiver Fundamentals 3
2.1 Linearity 3
2.1.1 Gain Compression 4
2.1.2 Input-referred Third Order Intercept Point (IIP3) 5
2.1.3 Input-referred Second Order Intercept Point (IIP2) 7
2.2 Noise Figure 7
2.3 Sensitivity 9
2.4 Dynamic Range 10
2.5 Receiver Architectures 12
2.5.1 Superheterodyne Receiver 12
2.5.2 Direct-Conversion Receiver 14
2.5.3 Low-IF Receiver 16
2.5.4 Wideband-IF Receiver 17
2.5.5 Half-IF Receiver 18
2.6 Design Challenges of High Frequency Operation 20
Chapter 3 A 24-GHz Fully Integrated Single-Pole Double-Throw Switch in a 0.18-m CMOS Process 23
3.1 Introduction 23
3.2 The Proposed Isolation-Enhanced Switch 24
3.2.1 Isolation issue 24
3.2.2 The Proposed Isolation-Enhanced Technique 28
3.3 Circuit Design 31
3.4 The Experimental Results 34
3.5 Conclusion 37
Chapter 4 CMOS Wireless Receiver Front-End for 24-GHz ISM band Applications 39
4.1 Introduction 39
4.2 The Proposed Receiver Architecture 40
4.3 The Proposed Receiver Design 43
4.3.1 The Low-Noise Amplifier 43
4.3.2 The Mixer and the IF Amplifiers 47
4.4 The Experimental Results 51
4.5 Conclusion 54
Chapter 5 CMOS LO Signal Generator for 24-GHz Receiver
Applications 57
5.1 Introduction 57
5.2 The Proposed LO Signal Generator Architecture 58
5.3 The Proposed LO Signal Generator Design 62
5.3.1 The Voltage-Controlled Oscillator 62
5.3.2 The Divide-by-Four Circuit with Phase Tuning 65
5.3.3 The testing passive mixer 69
5.4 The simulation results and measurement setup 70
5.4.1 The simulation results 70
5.4.2 The measurement setups 72
5.5 Discussion and Conclusion 74
Chapter 6 Conclusion 7
The Research on Using Standards for Protection of Confrontation to Replace the Hearsay RuleUse the Development of European Law as Example
No full text刑事被告享有對質詰問權，乃是普世價值之基本人權，歐洲人權公約第6條第3項d款及聯合國公民及政治權利公約第14條第3項f款皆有明文規定，主要是作為被告的程序參與權，在確保被告在程序中有主動的參與地位，免於淪為程序客體，然而各國對於對質詰問權的保障皆存有不同程度規範，究竟質問權保障範圍及限制的界限是否有一跨國性的標準得以遵循？當質問權的行使與證人的利益衝突時，應如何在二者間取得平衡，方不至侵害被告的質問權，影響公平程序的進行，皆為本文研究重點。歐洲人權法院關於對質詰問權的保障作出為數不少的裁判，更整合了歐洲不同法系間的相異規範，建立對質詰問權保障的歐洲標準，本文從比較法的觀點，以歐洲人權法院及歐洲人權委員會裁判為研究對象進行案例的分析，歸納對質詰問權保障的跨國性標準，並以英國及德國為例，說明質問保障的歐洲標準如何影響英國傳聞法則的操作，以及德國偵查中被告質問權的強化和證人保護措施的使用界限。國最高法院對於傳聞法則的操作，因欠缺質問例外的審查標準，故認為只要存在第159條之3的質問障礙事由，又符合傳聞例外不同程度的可信性及必要性要件，系爭未經質問的證詞即得例外作為證據使用，然而單純將傳聞例外的可信性或必要性作為質問例外，將導致被告質問權的侵害；其次，最高法院對於第159條之5的運用，未區分被告是否有辯護人的協助，即以被告未對證人審判外的陳述異議或未聲請傳喚證人，逕而認定被告拋棄質問權的行使，則產生不當減輕法院傳訊證人的調查義務，亦對無辯護人被告保障不足等問題；又當被告質問權的行使與證人的權利發生衝突時，釋字第636號解釋雖已明確指出必須考量採取其他限制較輕微之手段，例如以隔離方式為對質詰問，惟最高法院仍未能依釋字第636號之精神，具體在個案中審查是否有限制質問權行使之必要？有限制質問權行使之必要時，應優先考量哪些措施？因缺乏對次佳防禦措施優先性的審查標準，造成事實上其實尚存在其他較佳的防禦措施可以採行，卻捨棄不用而逕以對被告防禦權侵害較大的方式，作為證人訊問程序替代方案，實已不當侵害被告的質問權行使。針對上述問題，本文以歐洲人權法院所建立的質問權保障標準，及英國與德國如何以其標準調校內國實務操作為借鏡，認為應以質問權保障的審查標準取代我國傳聞例外的操作，避免傳聞法則的運用對被告造成質問權侵害。It is the basic international human right that the accused has the right to confront the witness against him, which is regulated in European Convention for the Protection of Human Rights and Fundamental Freedoms Article 6-3-d and International Covenant on Civil and Political Rights 14-3-f. However, the norms of the right to confront witnesses in every country are varying in degrees, whether a transnational standard of the protection and limitation of the right of confrontation exists to comply with? There are a large number of judgments of European Court of Human Rights(ECHR) regarding to the right of confrontation and ECHR integrats different legal systems in Europe between the different norms, then also establishes the European standard of the right of confrontation. This essay analyses the case law of ECHR and the European Commission of Human Rights(EComHR), then generalizes a transnational standard of the protection of the right to confront witnesses. Moreover, it uses the United Kingdom and Germany as examples to show how this standard affects the operation of the British rule of hearsay, as well as the right to confront witnesses in investigation stage and the limits of using witness protection measures in Germany.s to native law, the operation of the hearsay rule in Taiwan arises several problems. Due to a lack of standard of exceptions to the right of confrontation, the Supreme Court thinks exceptions of the hearsay rule are equal to exceptions of the right of confrontation, however, simply the credibility and necessity of the testimony, which are the elements of the exception of hearsay rule(§§159-1~159-3), are not enough to become the standard of exception to the right of confrontation. Besides, in what situation could be considered that the accused waives his right to confront witnesses? The application of §159-5 results in not only reducing the obligations of the court to summon witnesses, but also providing not enough protection to the accused who has no assistance of counsel. In addition, when the right of the witness conflicts with the right of the accused, the Supreme Court does not examine whether the relevant and sufficient reason to restrict confrontation exists or not, as well as what witness protection measures should be given priority. In response to these issues, this essay refers to the European standard of the right of confrontation established by ECHR and considers that we should use the standard of the protection of the right to confront witnesses to replace exceptions of the hearsay rule in practice for avoiding the violation of accused’s right to confront witnesses.第一章 緒論 1一節 研究動機與問題意識 1一項 刑事訴訟法傳聞例外之運用 1一款 傳聞例外等於對質詰問例外？ 1二款 偵查中質問權保障的架空 2三款 被告未聲請傳喚證人等於質問權的拋棄？ 3二項 證人保護與被告質問權的衝突 4二節 研究方法與本文架構 5一項 研究方法、對象及範圍 5二項 本文架構 6三節 用語說明 7二章 對質詰問權的保障範圍 9一節 概說 9一項 規範基礎 9二項 被告質問權的本質及保障目的 9二節 法理基礎：武器平等與對審程序 10一項 武器平等原則 10一款 在刑事程序體制框架內的當事人平等 11二款 相同的程序參與機會 12二項 對審程序 12三節 對質詰問的對象 13一項 歐洲人權法院－自主性解釋的證人概念 13二項 美國聯邦最高法院－Crawford v. Washington 14一款 質問法則取代傳聞法則 14二款 質問條款適用的限縮 15三款 學者見解 15三項 美國質問保障的後續發展 16四節 對質詰問權的保障內涵 17一項 有效質問的前提：閱卷權的保障 18二項 質問權行使的方式及範圍 19一款 被告與證人在場面對面的質問方式 19二款 質問範圍：證人的所有指控 19三項 質問權行使之主體 20一款 基本立場：Isgrò v. Italy 20二款 偵查中被告受律師協助權利的重要性 21五節 對質詰問權行使的程序階段 23六節 歐洲人權法院見解之檢討 24一項 被告受律師協助與有效質問之關係 24二項 偵查中質問權的行使 25一款 偵查中質問證人與公平審判的矛盾 25二款 偵查中質問證人與公平審判的調和 25七節 小結 26三章 對質詰問權的容許例外審查 29一節 歐洲人權法院審查方法及基準概說 29一項 審查方法：整體程序中是否對被告防禦權進行補償 29二項 審查基準 30二節 限制質問權行使的正當化事由 31一項 限制質問權行使之事由 31二項 正當化事由的判斷標準 31三項 證人匿名的必要性 32一款 一般標準 32二款 特殊證人身分的嚴格審查 33四項 秘密偵查者的保護必要審查 33五項 遭性侵害證人的保護必要審查 33六項 歐洲人權法院見解之檢討 34一款 秘密偵查者保護必要的審查標準未明 34二款 檢討案例：Sapunarescu v. Germany 35三節 義務法則 36一項 促使被告與證人直接質問的義務 36一款 法院傳喚拘提證人的義務：Delta v. France 36二款 偵查機關促成質問的義務：A.M. v. Italy 37三款 與相關判決的比較 37二項 反覆持續調查證人所在的義務 38一款 Artner v. Austria 38二款 Gossa v. Poland 39三項 義務的解除 39一款 Andandonskiy v. Russia 39二款 相關判決的意見 40四項 小結 41四節 歸責法則 41一項 可歸責於國家之事由 42一款 故意或過失導致證人無法被質問 42二款 先前程序的瑕疵 42三款 相關判決的意見 43二項 可歸責於被告之事由 43一款 被告的指摘義務（Rügepflicht） 43一目 Pullar v. the United Kingdom 43二目 Solakov v. the Former Yugoslav Republic of Macedonia 44三目 與相關判決的比較 45二款 被告的缺席 45一目 偵查中被告的缺席 45、Artner案與Solakov案 45、Demski v. Poland 46二目 被告缺席審判對辯護權的影響 47三目 被告缺席審判對質問權的影響 48三項 歐洲人權法院見解之評析 48一款 先前程序瑕疵的判斷標準不一 48一目 判斷標準的延伸發展－學說見解 49二目 本文意見 50二款 被告未盡指摘義務而失權的運用界限 51三款 以被告缺席作為可歸責於被告事由之限制 52一目 Artner案的不同意見 52二目 法院的特殊照料義務 53四項 小結 53五節 防禦法則 54一項 最理想質問保障方式的優先性 54二項 限制質問權行使的補償 55三項 最小侵害手段的審查 56四項 次佳防禦措施的優先性 56一款 僅限制質問範圍而維持直接質問的方式 56一目 Doorson v. the Netherlands 56二目 相關判決的意見 57二款 被告間接在場直接詰問證人 58一目 被告透過同步的聲音及影像傳送詰問證人 58二目 證人以化裝或變聲之方式接受被告之詰問 58三目 被告透過同步的聲音傳送詰問證人 59、Liefveld v. the Netherlands 60、Kok v. the Netherlands 60四目 歐洲人權法院意見小結 61三款 被告間接在場間接詰問證人 61一目 S.N. v. Sweden 61二目 Bocos-Cuesta v. the Netherlands 62三目 歐洲人權法院意見小結 62四款 被告僅能間接詰問證人而不得在場 63五款 無詰問直接證人之機會 63一目 僅能詰問間接證人 63、違反公約之案例 64、不違反公約案例 65二目 僅能對未能詰問之證言表示意見 67五項 次佳防禦措施及相關判決的評析 67一款 直接詰問方式作為補償程序的優先考量 67二款 剝奪被告在場權之審查標準檢討 67一目 學者對Sapunarescu v. Germany案之評析 68二目 本文意見 69三款 以間接證人代替直接證人詰問的審查標準不一 69一目 學者對Haas v. Germany案之評析 69二目 本文意見 71四款 以賦予被告表示意見機會作為補償程序的正當性 71六項 歐洲人權法院對於防禦法則運用的總結及檢討 71一款 被告提問及在場權的保障 72二款 防禦法則規避及濫用的可能性 72三款 無機會詰問匿名證人的補償 73六節 佐證法則 74一項 其他法則的補充適用 74二項 未經質問證詞之證據能力 75一款 Unterpertinger v. Austria 75二款 Asch v. Austria 75三項 歐洲人權法院意見及檢討 76七節 其他考量因素 77八節 小結 79四章 對質詰問權保障的歐洲標準對各內國法影響－以英國法及德國法為例 81一節 英國法規範 81一項 對質詰問權保障之依據 81二項 傳聞法則之規定 81一款 1988年刑事司法法案（CJA 1988） 82二款 2003年刑事司法法案（CJA 2003） 82一目 法定傳聞例外規定 83二目 本法所承認之普通法傳聞例外規則 83三目 經法院裁量許可使用之傳聞證據 84四目 被告的防禦保護措施 84三款 英國傳聞法則與歐洲人權公約之衝突 84二節 英國實務對於傳聞法則之運用 85一項 歐洲質問權保障標準的內化 85二項 義務法則 86一款 CJA 1988及CJA 2003之規定 86二款 實務判決 86一目 R. v. Radak 86二目 Attorney-General’s Reference (No. 1 of 2003) 87三項 歸責法則 88一款 CJA 2003之規定 88二款 實務判決 88一目 質問不能可歸責於國家 88二目 質問不能可歸責於被告 90四項 防禦法則 90一款 CJA 1988及CJA 2003之規定 90二款 實務判決 91一目 寬鬆的權衡標準 91二目 向歐洲人權法院標準調校 92、似是而非的歐洲標準運用：R. v. M (KJ) 92、未竟全功的調校：R. v. Sellick 93、匿名證人的審查標準：R. v. Davis 95三款 小結 98五項 佐證法則 98一款 CJA 2003之規定 99二款 實務判決 99六項 傳聞法則運用的其他問題 101三節 偵查階段質問權保障的強化－以德國法為例 102一項 德國刑事訴訟法相關規定 102二項 德國聯邦最高法院判決 102一款 BGHSt 46, 93 103二款 BGHSt 51, 150 104三款 BGH, Beschluss vom 27.01.2005, 1 StR 396/04 106三項 判決評析 106四項 參與式偵查程序的改革 109一款 概說 109二款 與質問權的衝突與調和 110四節 證人保護與被告質問權的保障 111一項 秘密偵查者與審判程序的衝突 111二項 行政機關的資訊封鎖（Behördliche Sperrung） 112三項 德國聯邦最高法院判決發展 112一款 早期實務見解及1983年大刑庭意見 113二款 近年實務見解 113四項 秘密證人在審判中的證據調查方式 114一款 與歐洲人權法院一致的標準 114一目 歐洲人權法院標準 114二目 德國證人保護之規範及其適用基準 114二款 尚待解決的問題 115五項 證人保護措施運用的界限 116一款 證人訊問的同步視訊轉播 116二款 訊問證人錄影帶之使用界限 117三款 與歐洲人權法院標準的比較分析 118五節 小結 119五章 我國實務發展與檢討 121一節 大法官會議解釋解決及遺留的問題 121一項 肯認對質詰問權憲法上之地位 121二項 尚待解決的問題 122二節 傳聞法則與對質詰問權之關係 123一項 未經質問之證詞無證據能力 124二項 質問證人為合法證據調查程序 125三項 二者實際運用上的差異？ 126四項 共通問題：欠缺不能質問的容許例外審查標準 127三節 國家機關促使質問的義務內涵 128一項 傳喚拘提證人並通知被告到場質問 129二項 促成質問的程序階段 131一款 少數見解－分段保障法 131二款 多數見解－整體觀察法 132三款 衍生問題：對整體觀察法的扭曲濫用 132四節 被告失權的運用 134一項 第159條之5解釋適用 134二項 衍生問題 136一款 不當減輕法院傳訊證人的調查義務 136二款 對無辯護人被告保障不足 137五節 次佳防禦措施的運用 137一項 刑事訴訟法第177條 137二項 證人保護措施 138一款 秘密證人保護的相關規定及適用 138二款 性侵害案件被害人的相關保護規定及適用 139三項 衍生問題 140一款 次佳防禦措施的濫用 140二款 欠缺次佳防禦措施優先性的審查標準 141六節 我國實務的檢討－向國際人權標準調校 142一項 質問權容許例外標準的建立 142一款 歐洲人權法院標準 143二款 英國法之發展 143三款 我國法的檢討 144二項 質問不能的例外審查標準 145一款 歐洲人權法院標準 145二款 德國審查標準 146三款 最高法院判決的檢討與修正 146一目 補行詰問說－偵查中質問權保障的障礙 146二目 修正方向 147三項 被告失權的運用界限 149一款 歐洲人權法院標準 149二款 最高法院判決的檢討與修正 150四項 次佳防禦措施的運用界限 150一款 限制質問權正當化理由的實質檢驗 150二款 防禦法則的運用 151三款 德國法規範 152四款 最高法院判決的檢討與修正 153一目 嚴格檢驗限制質問權的理由是否存在 153二目 證人保護與質問權保障的調和 155五項 質問權限制的法律效果 156一款 歐洲人權法院及德國實務見解 156二款 本文意見 157七節 小結 158六章 結論 159一節 質問保障的歐洲標準 159一項 歐洲人權法院審查標準 159二項 英國法與德國法的調校 160二節 我國實務審查標準的修正 161一項 我國實務傳聞法則操作衍生之問題 161二項 建立質問保障的審查標準 162三節 法制面的檢討修正 164一項 偵查階段質問權的保障 164二項 組織犯罪防制條例第12條第1項 165三項 性侵害犯罪防治法 166參考文獻 167案例索引 17
Eyeriss: A Spatial Architecture for Energy-Efficient Dataflow for Convolutional Neural Networks
Get PDFDeep convolutional neural networks (CNNs) are widely used in modern AI systems for their superior accuracy but at the cost of high computational complexity. The complexity comes from the need to simultaneously process hundreds of filters and channels in the high-dimensional convolutions, which involve a significant amount of data movement. Although highly-parallel compute paradigms, such as SIMD/SIMT, effectively address the computation requirement to achieve high throughput, energy consumption still remains high as data movement can be more expensive than computation. Accordingly, finding a dataflow that supports parallel processing with minimal data movement cost is crucial to achieving energy-efficient CNN processing without compromising accuracy. In this paper, we present a novel dataflow, called row-stationary (RS), that minimizes data movement energy consumption on a spatial architecture. This is realized by exploiting local data reuse of filter weights and feature map pixels, i.e., activations, in the high-dimensional convolutions, and minimizing data movement of partial sum accumulations. Unlike dataflows used in existing designs, which only reduce certain types of data movement, the proposed RS dataflow can adapt to different CNN shape configurations and reduces all types of data movement through maximally utilizing the processing engine (PE) local storage, direct inter-PE communication and spatial parallelism. To evaluate the energy efficiency of the different dataflows, we propose an analysis framework that compares energy cost under the same hardware area and processing parallelism constraints. Experiments using the CNN configurations of AlexNet show that the proposed RS dataflow is more energy efficient than existing dataflows in both convolutional (1.4x to 2.5x) and fully-connected layers (at least 1.3x for batch size larger than 16). The RS dataflow has also been demonstrated on a fabricated chip, which verifies our energy analysis
Eyeriss: An Energy-Efficient Reconfigurable Accelerator for Deep Convolutional Neural Networks
No full textDeep learning using convolutional neural networks (CNN) gives state-of-the-art
accuracy on many computer vision tasks (e.g. object detection, recognition,
segmentation). Convolutions account for over 90% of the processing in CNNs
for both inference/testing and training, and fully convolutional networks are
increasingly being used. To achieve state-of-the-art accuracy requires CNNs with
not only a larger number of layers, but also millions of filters weights, and varying
shapes (i.e. filter sizes, number of filters, number of channels) as shown in Fig.
14.5.1. For instance, AlexNet [1] uses 2.3 million weights (4.6MB of storage) and
requires 666 million MACs per 227×227 image (13kMACs/pixel). VGG16 [2] uses
14.7 million weights (29.4MB of storage) and requires 15.3 billion MACs per
224×224 image (306kMACs/pixel). The large number of filter weights and
channels results in substantial data movement, which consumes significant
energy.United States. Defense Advanced Research Projects Agency (DARPA YFA grant N66001-14-1-4039)Intel CorporationMassachusetts Institute of Technology. Center for Integrated Circuits and System
- …
