1,525 research outputs found
Slimmable neural networks for edge devices
Get PDFWhile 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
sj-docx-1-tct-10.1177_15330338211045510 - Supplemental material for Centromere Protein I (CENP-I) Is Upregulated in Gastric Cancer, Predicts Poor Prognosis, and Promotes Tumor Cell Proliferation and Migration
No full textSupplemental material, sj-docx-1-tct-10.1177_15330338211045510 for Centromere Protein I (CENP-I) Is Upregulated in Gastric Cancer, Predicts Poor Prognosis, and Promotes Tumor Cell Proliferation and Migration by Jiahui Wang, Xin Liu, Hong-jin Chu, Ning Li, Liu-ye Huang and Jian Chen in Technology in Cancer Research & Treatment</p
FlowTrack: Revisiting Optical Flow for Long-Range Dense Tracking
No full textIn the domain of video tracking, existing methods often grapple with a trade-off between spatial density and temporal range. Current approaches in dense optical flow estimators excel in providing spatially dense tracking but are limited to short temporal spans. Conversely, recent advancements in long-range trackers offer extended temporal coverage but at the cost of spatial sparsity. This paper introduces FlowTrack, a novel framework designed to bridge this gap. FlowTrack combines the strengths of both paradigms by 1) chaining confident flow predictions to maximize efficiency and 2) automatically switching to an error compensation module in instances of flow prediction inaccuracies. This dual strategy not only offers efficient dense tracking over extended temporal spans but also ensures robustness against error accumulations and occlusions, common pitfalls of naive flow chaining. Furthermore, we demonstrate that chained flow itself can serve as an effective guide for an error compensation module, even for occluded points. Our framework achieves state-of-the-art accuracy for longrange tracking on the DAVIS dataset, and renders 50% speed-up when performing dense tracking
Towards efficient, on-demand and automated deep learning
Get PDFIn the past decade, deep learning has achieved great breakthroughs on tasks of computer vision, speech, language, control and many others. The advanced and dedicated computing chips, like Nvidia GPU and Google TPU, largely contributed and broadened this success. However, the requirement of large computing power impedes the deployment of deep learning methods in many real scenarios, where cost, time and energy efficiency are critical -- for example, self-driving cars, AR/VR kits, internet-of-things devices and mobile phones. This thesis presents a series of in-depth research towards efficient, on-demand and automated deep learning.Submission original under an indefinite embargo labeled 'Open Access'. The submission was exported from vireo on 2020-08-25 without embargo termsThe student, Jiahui Yu, accepted the attached license on 2020-01-17 at 15:58.The student, Jiahui Yu, submitted this Dissertation for approval on 2020-01-17 at 16:04.This Dissertation was approved for publication on 2020-01-21 at 15:15.DSpace SAF Submission Ingestion Package generated from Vireo submission #14849 on 2020-08-25 at 17:03:08Made available in DSpace on 2020-08-26T21:53:55Z (GMT). No. of bitstreams: 3
YU-DISSERTATION-2020.pdf: 2498738 bytes, checksum: d6fc5cf2e3f6cfe94a0339c8e6a93444 (MD5)
LICENSE.txt: 4206 bytes, checksum: 9724ae490373d9c91b7c81ca6e090b30 (MD5)
PROQUEST_LICENSE.txt: 4552 bytes, checksum: f244f3fdd68510197b4e67e15280cec4 (MD5)
Previous issue date: 2020-01-2
Asymmetric [3+2] Photocycloadditions of Cyclopropanes with Alkenes or Alkynes through Visible-Light Excitation of Catalyst-Bound Substrates
Get PDFThe herein reported visible-light-activated catalytic asymmetric [3+2] photocycloadditions between cyclopropanes and alkenes or alkynes provide access to chiral cyclopentanes and cyclopentenes, respectively, in 63–99 % yields and with excellent enantioselectivities of up to >99 % ee. The reactions are catalyzed by a single bis-cyclometalated chiral-at-metal rhodium complex (2–8 mol %) which after coordination to the cyclopropane generates the visible-light-absorbing complex, lowers the reduction potential of the cyclopropane, and provides the asymmetric induction and overall stereocontrol. Enabled by a mild single-electron-transfer reduction of directly photoexcited catalyst/substrate complexes, the presented transformations expand the scope of catalytic asymmetric photocycloadditions to simple mono-acceptor-substituted cyclopropanes affording previously inaccessible chiral cyclopentane and cyclopentene derivatives
Global 3D radiation magnetohydrodynamic simulations of accretion onto a stellar-mass Black Hole at sub- and near-critical accretion rates
No full textWe present global 3D radiation magnetohydrodynamic simulations of accretion onto a 6.62 solar-mass black hole, with quasi-steady-state accretion rates reaching 0.016–0.9 times the critical accretion rate, which is defined as the accretion rate for powering the Eddington luminosity, assuming a 10% radiative efficiency, in three different runs. The simulations show no sign of thermal instability over hundreds of thermal timescales at 10 rg. The energy dissipation occurs close to the mid-plane in the near-critical runs and near the disk surface in the low–accretion rate run. The total radiative luminosity inside ∼20 rg is about 1%–30% of the Eddington limit, with radiative efficiencies of about 6% and 3%, respectively, in the sub- and near-critical accretion regimes. In both cases, self-consistent turbulence generated by the magnetorotational instability leads to angular momentum transfer, and the disk is supported by magnetic pressure. Outflows from the central low-density funnel, with a terminal velocity of ∼0.1c, are seen only in the near-critical runs. We conclude that these magnetic pressure–dominated disks are thermally stable and thicker than the α disk, and that the effective temperature profiles are much flatter than those in the α disks. The magnetic pressures of these disks are comparable within an order of magnitude to the previous analytical magnetic pressure–dominated disk model
Initial failure surface for the cohesive slope after Huang et al. [1].
No full textInitial failure surface for the cohesive slope after Huang et al. [1].</p
MultiBodySync: Multi-Body Segmentation and Motion Estimation via 3D Scan Synchronization
Get PDFWe present MultiBodySync, a novel, end-to-end trainable multi-body motion segmentation and rigid registration framework for multiple input 3D point clouds. The two non-trivial challenges posed by this multi-scan multibody setting that we investigate are: (i) guaranteeing correspondence and segmentation consistency across multiple input point clouds capturing different spatial arrangements of bodies or body parts; and (ii) obtaining robust motion-based rigid body segmentation applicable to novel object categories. We propose an approach to address these issues that incorporates spectral synchronization into an iterative deep declarative network, so as to simultaneously recover consistent correspondences as well as motion segmentation. At the same time, by explicitly disentangling the correspondence and motion segmentation estimation modules, we achieve strong generalizability across different object categories. Our extensive evaluations demonstrate that our method is effective on various datasets ranging from rigid parts in articulated objects to individually moving objects in a 3D scene, be it single-view or full point clouds. Code at https: //github.com/huangjh-pub/multibody-sync
The semi-complementizer shuō and non-referential CPs in Mandarin Chinese
Get PDFThe empirical focus of this paper is the syntactic status of the semi-complementizer shuō grammaticalized from verbs of saying, in Mandarin Chinese. Such elements have been shown to exhibit atypical patterns compared to that in English, which triggers discussions of whether shuō should be analyzed as a complementizer (Paul, 2014; Huang, 2018). This paper presents novel data surrounding the distributional patterns of shuō and argues that shuō is a C head that introduces a subtype of CPs called non-referential CPs, following de Cuba (2017)
Application of biodegradable implants in pediatric orthopedics: shifting from absorbable polymers to biodegradable metals
No full textOver the past two decades, advances in pediatric orthopedics and closed reduction combined with percutaneous internal fixation techniques have led to significant growth in pediatric orthopedics surgery. Implants such as Kirschner-wires, cannulated screws and elastic stabilization intramedullary nails are commonly used in these procedures. However, traditional implants made of metal or inert materials are not absorbable, leading to complications that affect treatment outcomes. To address this issue, absorbable materials with excellent mechanical properties, good biocompatibility, and controlled degradation rates have been developed and applied in clinical practice. These materials include absorbable polymers and biodegradable metals. This article provides a comprehensive summary of these resorbable materials from a clinician's perspective. In addition, an in-depth discussion of the feasibility of their clinical applications and related research in pediatric orthopedics is included. We found that the applications of absorbable implants in pediatric orthopedics are shifting from absorbable polymers to biodegradable metals and emphasize that the functional characteristics of resorbable materials must be coordinated and complementary to the treatment in pediatric orthopedics
- …
