782 research outputs found
Slimmable neural networks for edge devices
No full textWhile methods based on deep learning have witnessed major breakthroughs in machine perception and generative modeling, the problem of how to run neural networks within latency budget for edge devices remains unsolved. This thesis presents a new approach to train a single neural network executable at arbitrary widths for instant and adaptive accuracy-efficiency trade-offs at runtime.
First a simple and general method is presented to train a single neural network executable at different widths (number of channels in a layer). The width can be chosen from a predefined widths set to adaptively optimize accuracy-efficiency trade-offs at runtime. Instead of training individual networks with different width configurations, we train a shared network with switchable batch normalization. At runtime, the network can adjust its width on the fly according to on-device benchmarks and resource constraints, rather than downloading and offloading different models. Our trained networks, named slimmable neural networks, achieve ImageNet classification accuracy similar to (and in many cases better than) that of individually trained models of MobileNet v1, MobileNet v2, ShuffleNet and ResNet-50 at different widths. We also demonstrate better performance of slimmable models compared with individual ones across a wide range of applications including COCO bounding-box object detection, instance segmentation and person keypoint detection without tuning hyper-parameters. We visualize and discuss the learned features of slimmable networks.
Further, we propose a systematic approach to train universally slimmable networks (US-Nets), extending slimmable networks to execute at arbitrary width, and generalizing to networks both with and without batch normalization layers. In addition, we propose two improved training techniques for US-Nets, named the sandwich rule and the inplace distillation, to enhance training process and boost testing accuracy. We show improved performance of universally slimmable MobileNet v1 and MobileNet v2 on ImageNet classification task, compared with individually trained ones and 4-switch slimmable network baselines. We also evaluate the proposed US-Nets and improved training techniques on tasks of image super-resolution and deep reinforcement learning. Extensive ablation experiments on these representative tasks demonstrate the effectiveness of our proposed methods. Our discovery opens up the possibility to directly evaluate a FLOPs-Accuracy spectrum of network architectures. Finally, we demonstrate an application to search for channel number configurations based on proposed slimmable networks.Submission published under a 24 month embargo labeled 'Closed Access', the embargo will last until 2021-05-01The student, Jiahui Yu, accepted the attached license on 2019-02-14 at 14:34.The student, Jiahui Yu, submitted this Thesis for approval on 2019-02-14 at 14:42.This Thesis was approved for publication on 2019-02-15 at 11:18.DSpace SAF Submission Ingestion Package generated from Vireo submission #13390 on 2019-08-22 at 16:19:49Made available in DSpace on 2019-08-23T20:44:31Z (GMT). No. of bitstreams: 2
YU-THESIS-2019.pdf: 1268760 bytes, checksum: c091ef8a839188e9d52d208dee832b8a (MD5)
LICENSE.txt: 4206 bytes, checksum: 1b6cf1c051b15c1073c51d0ad5e1abd0 (MD5)
Previous issue date: 2019-02-15Embargo set by: Seth Robbins for item 112252
Lift date: 2021-08-23T20:44:50Z
Reason: Author requested closed access (OA after 2yrs) in Vireo ETD systemEmbargo set by: Seth Robbins for item 112252
Lift date: 2021-08-23T20:46:41Z
Reason: Author requested closed access (OA after 2yrs) in Vireo ETD systemEmbargo set by: Seth Robbins for item 112252
Lift date: 2021-08-23T20:47:38Z
Reason: Author requested closed access (OA after 2yrs) in Vireo ETD systemEmbargo set by: Seth Robbins for item 112252
Lift date: 2021-08-23T20:48:32Z
Reason: Author requested closed access (OA after 2yrs) in Vireo ETD systemLimited Restriction Lifted for Item 112252 on 2021-08-24T09:15:34Z
Model-based myelin water fraction mapping: analyses and improvement
No full textIn this thesis, the problem of model-based myelin water fraction (MWF) mapping is addressed. We first focus on three of the most widely used signal models for T2*-myelin water imaging (MWI), i.e., the NNLS-multi-exponential model, the magnitude-3-exponential model, and the complex-3-exponential model, and investigate their sensitivities to practical perturbations such as random noise and field-related structured errors. We demonstrate through both Cramér-Rao lower bound (CRLB) analyses and Monte Carlo simulations that the three signal models are all very unstable inherently. Comparatively speaking, however, we demonstrate the theoretical advantage of the 3-exponential models over the multi-exponential model in handling noise, and the practical advantage of the magnitude models over the complex model in handling phase-related perturbations for T2*-MWI. We also illustrate the necessity and effects of incorporating various types of constraints for additional sensitivity gain.
Using the insights obtained in the sensitivity analyses, we then propose a new MWF fitting scheme that leverages an improved signal model and a set of more effective constraints. In particular, a relaxed magnitude-3-exponential model with additional frequency compensation terms is introduced to better represent voxels with large field variations; a set of statistical distributions learned from in vivo training data is further imposed on the model parameters for additional constraints. Using phantom simulation and in vivo experiments, we then evaluate and compare the proposed method with several popular conventional MWF fitting schemes to demonstrate the improved accuracy and robustness of the proposed method.
In this thesis, a literature review on the study of myelin and the development of MWF mapping is provided at the start of the work. Background materials on the CRLB theories are also provided to facilitate reading.Submission published under a 24 month embargo labeled 'Closed Access', the embargo will last until 2022-12-01The student, Jiahui Xiong, accepted the attached license on 2020-11-17 at 12:52.The student, Jiahui Xiong, submitted this Thesis for approval on 2020-11-17 at 13:03.This Thesis was approved for publication on 2020-11-18 at 14:00.DSpace SAF Submission Ingestion Package generated from Vireo submission #15895 on 2021-03-04 at 16:31:59Made available in DSpace on 2021-03-05T21:45:32Z (GMT). No. of bitstreams: 2
XIONG-THESIS-2020.pdf: 11672242 bytes, checksum: 4f0060436c20d10f6c990dffb93b1d2f (MD5)
LICENSE.txt: 4209 bytes, checksum: 386f1d8975a06576e18f881ab10f5939 (MD5)
Previous issue date: 2020-11-18Embargo set by: Seth Robbins for item 117288
Lift date: 2023-03-05T21:45:47Z
Reason: Author requested closed access (OA after 2yrs) in Vireo ETD systemEmbargo set by: Seth Robbins for item 117288
Lift date: 2023-03-05T21:47:41Z
Reason: Author requested closed access (OA after 2yrs) in Vireo ETD systemAuthor requested closed access (OA after 2yrs) in Vireo ETD systemLimite
Architecture modeling for "Heterogeneous integration of spin-photon interfaces with a scalable CMOS platform"
No full textModeling file GitHub repo for Manuscript: Heterogeneous integration of spin-photon interfaces with a scalable CMOS platformAuthors: Linsen Li*, Lorenzo De Santis, Isaac Harris, Kevin C. Chen, Yihuai Gao, Ian Christen, Matthew Trusheim, Hyeongrak Choi, Yixuan Song, Carlos Errando-Herranz, Jiahui Du, Yong Hu, Genevieve Clark, Mohamed I. Ibrahim, Gerald Gilbert, Ruonan Han, Dirk Englund** linsenli@mit,edu; [email protected]</p
Data in "Heterogeneous integration of spin-photon interfaces with a scalable CMOS platform"
No full textManuscript: Heterogeneous integration of spin-photon interfaces with a scalable CMOS platformAuthors: Linsen Li*, Lorenzo De Santis, Isaac Harris, Kevin C. Chen, Yihuai Gao, Ian Christen, Matthew Trusheim, Hyeongrak Choi, Yixuan Song, Carlos Errando-Herranz, Jiahui Du, Yong Hu, Genevieve Clark, Mohamed I. Ibrahim, Gerald Gilbert, Ruonan Han, Dirk Englund** linsenli@mit,edu; [email protected] Data forFigure 3: https://doi.org/10.6084/m9.figshare.25374550.v1 Figure 4: https://doi.org/10.6084/m9.figshare.25374559.v1Figure 5: https://doi.org/10.6084/m9.figshare.25374571.v1</p
Lun zhong Tang yue fu de fu xing yu xin bian
No full text本文以「從禮樂到生民:論中唐樂府的復興與新變」為題，試圖將中唐樂府放入當時思想、政治與文化的新格局中，探討其創作及觀念之特徵，並對盛、中唐之際的文學發展與中唐樂府的革新提出新的解釋。已往之研究往往強調中唐樂府「刺時」的一面，而歸因於中唐的政治狀況。但本文認為，就題材來說，中唐樂府並非僅有「刺時」的內容，亦有大量的風俗描寫; 就原因而言，政治固然是因素之一，然絶非唯一因素。其背後還涉及詩學觀念的轉向與文人身份的變化。本文指出，盛、中唐之際文人對於「風雅」或者「詩教」的觀念發生了變化。盛唐以前對「教化」的理解總與「禮樂」相伴，認為只要行禮樂便可以達到教化的目的，強調「禮樂」的作用自然凸顯《詩經》「上以風化下」的一端，而非「下以風刺上」的一面。至盛唐末期，這種觀念受到挑戰，文人對詩教的理解從頌讚的雅頌走向諷刺的國風，而最後刺時之文也獲得了正當性。中唐樂府的興盛與這種詩學觀念的轉變有著密切的關係。本文認為，中唐文人的政治身分變化與中唐樂府的變化亦有密切關係。初盛唐文人少做地方官，至中唐，長期擔任地方官的文人比例大增。他們在觀念上調和「儒」、「吏」關係，更多地認同「循吏」的價值與傳統，使得地方官文人與漢良二千石的傳統銜接。他們以「循吏」的眼光在地方進行創作，為樂府的「觀風」職能提供良好的背景。觀念的轉向、文人身份的變化，加之當時的政治狀況，造成了中唐樂府的興盛，也造就其獨特之特徵。本文對張籍、王建樂府在盛中唐樂府演變中的作用與地位提出了新的論述，認為張、王的樂府唱和，繼承、改造了盛唐李白和杜甫新、古題樂府的創作形式，體現了盛唐末年所開啟的「風雅」詮釋方式，而以興諷刺時作為樂府的主要內容，為其後的元、白樂府提供了借鏡。本文指出，中唐樂府雖然繼承了漢樂府「觀風」的傳統，但在題材內容上有所擴大。漢代，「樂府」作為中央官署，其「采詩」的主要目的是考察地方吏治，對於地方風土風俗則沒有太大興趣，但中唐樂府對此一題材卻有大量的書寫。「風俗」題材的興起與當代儒學關注點從「禮樂」轉向「生民」有關。「生民」既成為評判政治優劣的標準，「生民」的活動自然被納入創作的視野。論文認為，從文體角度說，中唐的樂府作者雖在主觀上排斥六朝傳統，但在樂府的命題以及表現方式上依然深受六朝的影響。論文還考察了與新樂府創作相應的一套詩學論述，發現中唐詩人提出了所謂樂府「正聲」的觀念，以對抗初、盛唐樂府的觀念。不僅如此，中唐的樂府觀念還與「史官」、「諫官」的意識相互滲透，使得本為詩歌所獨有的「采風」觀念擴散到詩歌以外的文類，而本為樂府所關注的獨特題材也進入其他文體的書寫當中。This thesis focusing on Yuefuin the Mid-Tang dynasty,discusses its characteristicsin the new pattern of thought, politics and culture. The thesis tries to provide a new explanation for the literature development in the Mid-Tang.Existing researches often emphasized the relation between yuefu and political reality. However, political reality was not the only theme reflected in yuefu, the description of customs was included. Politic was not the mere reason. Literati’s political identitywas changed and the existing tradition of "praise and criticize" from the Han dynasty was also transformed.The thesis pointed out that the idea of "Fengya" or "poetic education" was changed from the High-Tang to the Mid-Tang dynasty. "Civilization", usually relating to "rites and music", was considered to be successful in the High-Tang dynasty. Therefore, literature was encouraged to praise the dynasty while criticism was ignored. This idea faced challenged from the end of the High-Tang dynasty, and the "poetic education" was lead to "criticize" instead of "praise". The prosperity of yuefu in the Mid-Tang dynasty was closely allied to the new idea.Meanwhile, literati’s political identity changed. Most of them were local officials.Travelling within a wider area activated them to identify the "liangqiandan" tradition from the Han dynasty and to write in a "moral official’s" sight.The change of idea, literati identity as well as political reality provided an environment for yuefu creation in the Mid-Tang dynasty. The thesis provides a new exposition for Zhang Ji and Wang Jian’s positions in this literature development. Zhang and Wang wrote poems in replies. They continued or reformed Li Bai and Du Fu’s yuefu works and represented the new idea of "Fengya" in their works, which provided experience for Yuen Zhen and Bai Juyi later. At the same time, the concept of "Guan Feng" was broadened. Han "Yuefu", as a central office, collected poems to look at local politics with no interest in local customs. However, yuefu in the Mid-Tang dynasty viewed customs as a mean subject of writing. The literati subjectively excluded the tradition of the Six dynasties, but were in fact influenced by this tradition. Besides, customs as a literary subject related to the transform of contemporary Confucianism. When literati viewed "people", not "ritesand music" as the basic of civilization in the political field, people’s activities became a literary theme with justification.Finally, yuefu developed a standard for itself as poets concerned their works on the field. Poets raised "canon" of yuefu againstthe verse written in the Early and High-Tang dynasty. And the idea about yuefu penetrated with the sense of both historian and remonstrating officials when it came to mature. The idea of "folk song collection", coming down from yuefu was broadened and literary themes originally belonged to yuefu were written by other literary forms.Detailed summary in vernacular field only.Detailed summary in vernacular field only.Detailed summary in vernacular field only.Detailed summary in vernacular field only.Detailed summary in vernacular field only.Detailed summary in vernacular field only.呂家慧 = From rising "rites and music" to viewing "people": the research on yuefu poem in the Mid-Tang dynasty / Lu Chia Hui.Thesis (Ph.D.) Chinese University of Hong Kong, 2014.Includes bibliographical references (leaves 251-256).Abstracts also in English.Lü Jiahui = From rising "rites and music" to viewing "people" : the research on yuefu poem in the Mid-Tang dynasty / Lu Chia Hui
CHINA'S HOUSING AFFORDABILITY : MYTH OR REALITY? - A STUDY OF CHINA'S AFFORDABILITY SITUATION AFTER 1998
No full textBachelorBACHELOR OF SCIENCE (REAL ESTATE
Inbetween Context| A lobby for Hembrug: sustainable repurpose of industrial heritage buildings
No full textMany industrial buildings had remained the situation of vacancy since the late 1970s, reusing has become a term to be used as a sustainable method of preserving the cultural identity of these industrial buildings in order to prolong their life-span.The Hembrug site is a plot which was formed in the 19th century by reclamation of land in the coastal area between Amsterdam and Zaandam. It was taken into use in 1895, as a weapon production base. However, the production came to a standstill in the 1990s, after which the factories were finally closed in 2003. The government decided to develop the mixed terrain for living, culture, nature and business in the coming years. However, industrial buildings on the site from industrial era are not adequate for the spatial and functional requirements in the contemporary society. Compare to demolition, renovation not only honors the past but also means looking into the future. It can help to create more space, reduce energy consumption and preserve the cultural value while bringing the building up to the latest technical standards.My research focus on finding the suitable renovation strategies to the industrial heritage buildings, case study was used as the fundamental methodology while the descriptive research helped to summarized them into three main prototypes with the perspective of the spatial order between the existing building and its new intervention. The research findings are applied to the Hembrug Site and the design process also follows the strategies as well.Second LifeArchitecture, Urbanism and Building Science
Impact of Exclusive Lanes for Autonomous Vehicles: A Cost-benefit Analysis of Exclusive AV Lanes in Eastern China
No full textThese days, autonomous driving has become a popular topic in different fields and showed its unavoidable developing tendency. With the increasing application of autonomous vehicles, a mixed period of autonomous vehicles (AVs) and conventional vehicles (CVs) is inevitable. Therefore, some traffic management was proposed. Exclusive lanes for AVs are a new way to improve the mixed road environment's traffic capacity. However, the effect of exclusive AV lanes was rarely studied. Therefore, the decision-makers could not clearly understand this new policy's impact. Besides, the existing research content did not provide a detailed quantitative analysis of this new policy, which is very important for decision-makers. Thus, this thesis focused on the exclusive AV lanes and proposed the following research question: Under which conditions are exclusive autonomous vehicle lanes societally beneficial?Management of Technology (MoT
Attention-Aware Age-Agnostic Visual Place Recognition
No full textA cross-domain visual place recognition (VPR) task is proposed in this work, i.e., matching images of the same architectures depicted in different domains. VPR is commonly treated as an image retrieval task, where a query image from an unknown location is matched with relevant instances from geo-tagged gallery database. Different from conventional VPR settings where the query images and gallery images come from the same domain, we propose a more common but challenging setup where the query images are collected under a new unseen condition. The two domains involved in this work are contemporary street view images of Amsterdam from the Mapillary dataset (source domain) and historical images of the same city from Beeldbank dataset (target domain). We tailored an age-invariant feature learning CNN that can focus on domain invariant objects and learn to match images based on a weakly supervised ranking loss. We propose an attention aggregation module that is robust to domain discrepancy between the train and the test data. Further, a multi-kernel maximum mean discrepancy (MK-MMD) domain adaptation loss is adopted to improve the cross-domain ranking performance. Both attention and adaptation modules are unsupervised while the ranking loss uses weak supervision. Visual inspection shows that the attention module focuses on built forms while the dramatically changing environment are less weighed. Our proposed CNN achieves state of the art results (99% accuracy) on the single-domain VPR task and 20\% accuracy at its best on the cross-domain VPR task, revealing the difficulty of age-invariant VPR
- …
