教学课程资源库 更多>>

关键词： Computer science   Medical imaging   Artificial intelligence  

摘要： Medical video and image analysis and understanding is important yet challenging. This is not only because of the complexity of the medical domain knowledge itself in understanding the manifestations, but also due to the big variety across the different input modalities with vastly different characteristics, which include the X-ray images, computed tomography (CT) scans, and RGB videos, etc. Each type of images and videos may require special design both in network structure and training strategy. Recently, in the wake of the application of deep learning (DL) in a variety of computer vision tasks, various DL models are widely used in medical image analysis. Thanks to its powerful capabilities for extracting feature, DL significantly improves the performance over previous methods in those medical image analysis tasks. In this dissertation, we focus on two tasks: 1) human pose estimation based in-home lower body rehabilitation system, which can help the patients to finish their physical therapy activities in home by themselves instead of under the guidance of physical therapist in clinic; 2) ileocecal valve detection in colonoscopy videos in real time, which can be used to evaluate the start time of withdraw from colonoscopy videos. Firstly, we design, develop and evaluate an in-home lower body rehabilitation system based on a novel lightweight human pose estimation model. After created a lower body rehabilitation dataset with each image annotated with the ground truth joint point locations, we design a lightweight but powerful neural network model, which runs on a smartphone, to estimate human pose. Using our system, patients can perform rehabilitation activities at home without the presence of physical therapists. Our system can facilitate in-home rehabilitation and reduce the cost for patients. Secondly, one significant landmark of a colonoscopy is when the endoscope begins to withdraw after reaching the cecum. We demonstrate that the withdrawal point is related to t

关键词： Computer engineering   Statistical physics   Computer science   Aeronomy   Aerospace engineering  

摘要： Very low Earth orbits (VLEO) with altitudes below 450 km are generally avoided for most applications. At such low altitudes, aerodynamic forces become the most significant source of attitude and orbit perturbation, whilst gravity gradient and residual magnetic dipole torques increase due to proximity with the Earth. In this scenario, the question of whether novel aerodynamic control strategies can be used to facilitate attitude and orbit control tasks arises. This thesis documents novel control methods that take advantage of aerodynamic torques to perform: 1) combined aerodynamic and reaction wheels pointing, 2) three-axis aerodynamic control and 3) reaction wheels momentum management below 300 km. A comprehensive review of the uncertainties associated with space weather and the mechanisms of interaction of the particles with the satellite is performed. These uncertainties are taken into account in the design of a novel real-time algorithm to determine the configuration of some control surfaces inducing the desired aerodynamic control torques. A representative 3U CubeSat with four motorized surfaces is then taken as reference to explore feasibility of the three control methods proposed. The analysis of the results benefits from the implementation of a realistic simulation environment, which allows to reproduce the inaccuracies affecting performance at the state of the art. A Monte Carlo analysis is also included to demonstrate the controller robustness against the identified sources of uncertainty and, more importantly, to acknowledge its limits. Comparison with results obtained in the presence of materials promoting quasi-specular reflection is presented with the purpose of discussing possible future developments. The analysis of the results show that active aerodynamic control is feasible and that the proposed control solutions, despite their relative simplicity, are robust enough to assure the control tasks completion under uncertain and highly variable orbit con

关键词： Computer science   Artificial intelligence  

摘要： The Internet of Things (IoT) is rapidly growing and has been widely recognized as the next revolution and promises to transform various realms of our daily lives. While the security and privacy threats (e.g., the IoT devices abuse, information theft, and sensor vulnerability) towards IoT devices are also increasing. These problems could lead to a series of catastrophic results, mainly severely increasing the risk of huge financial and health loss for both enterprises and individuals. However, due to the complexity of the surrounding occlusion, unavoidable variance in signal scaling, and privacy-preserving requirements, existing solutions yield unsatisfactory performance. First, most of the current wireless sensing approaches are based on the Huygens-Fresnel principle, which primarily focuses on the target exterior shape or object motion (e.g., appearance reflection information) and has limitations to assistant us to see-through the obstacles and explore the intrinsic secrets of the devices based on weak useful signal (e.g., structural components investigation and material characterization). Besides, the sensing data in the present methods (e.g., computer vision and voice recognition) inevitably contain sensitive information, which has a high risk of information leakage. In this dissertation, we propose a novel wireless meta-sensing technology to secure and identify the vulnerabilities of IoT devices and further empower a new paradigm of IoT with wireless inkables. This dissertation is the first to explore the interconnection between wireless sensing, material and component, and security and privacy analysis for IoT. To begin with, we notice that hidden electronic devices (e.g., spy camera, bomb package, and bug) can cause life-threatening hazards, eavesdropping, and cheating or intrusion in private zones. Thus, we exploit the feasibility of hidden electronic device recognition under mmWave and investigate the unique properties in the nonlinear responses of electroni

关键词： Computer science   Artificial intelligence  

摘要： Recent democratization of fabrication techniques has revolutionized manufacturing. Desktop versions of devices like 3D printers, laser cutters, and CNC mills are now available at affordable prices making them increasingly common among end-users, hobbyists, and non-experts. While the reduced hardware cost of manufacturing devices has created new opportunity for casual makers to innovate and explore, the corresponding software pipeline has not advanced as dramatically. Design tools, mesh editing software, and CAD frameworks are developed primarily for trained experts and thus pose a steep learning curve for beginners. Even with crowd-sourced solutions, working with pre-existing models requires substantial expertise. Together with the fact that the fabrication process (typically multi-stage) itself is slow and hard to debug, these design level challenges frequently lead to delayed results, failed prints, wasted resources, and frustration. This thesis focuses on some of these challenges by viewing the entire computational fabrication pipeline as a compiler. First, it provide formal semantics to commonly used representations for 3D modeling. This formal foundation aids the development of a decompiler that automatically reverse engineers high-level computer-aided designs from low-level representations by recovering latent structure. An equality saturation based second decompilation phase discovers even high-level representations using dynamic rewrites and inverse transformations. A second application of dynamic rewrites in the domain of carpentry shows wider applicability of this approach for developing optimizing compilers for different kinds of fabrication techniques. Both these equality saturation based approaches require manually writing the rewrites which itself is tedious and error-prone. A novel framework for automatically inferring rewrites shows how the entire pipeline can be automated, and be applied to more traditional compilers.

关键词： Applied mathematics   Computer science   Artificial intelligence  

摘要： The first part of the thesis proves some theoretical results in deep learning. For the approximation problem, we prove that deep neural networks can approximate analytic functions exponentially fast. The number of parameters needed to achieve an error tolerance of epsilon is O((log 1/epsilon)^d), and it is exponential in the sense that the rate depends only on log 1/epsilon instead of epsilon itself. We also develop a general method to show that the deep networks never have worse approximation properties than shallow ones. For the optimization problem, we analyze the global convergence of gradient descent for deep linear residual networks by proposing a new initialization: zero-asymmetric (ZAS) initialization. It is motivated by avoiding stable manifolds of saddle points. We prove that under the ZAS initialization, for an arbitrary target matrix, gradient descent converges to an epsilon-optimal point in O(L^3 log 1/epsilon) iterations, which scales polynomially with the network depth L. It demonstrates the importance of the residual structure and the initialization in the optimization for deep linear neural networks. The second part focuses on continuous-time hidden Markov models (CT-HMM), where both the hidden states and observations occur in continuous time. We propose a unified framework that formally obtains the model parameter estimation by taking continuous-time limit of the classical discrete-time Baum-Welch algorithm, and recovers and extends several previous results in CT-HMM under different settings. Here two settings are illustrated: hidden jump process with a finite state space, and hidden diffusion process with a continuous state space. For each setting, we first estimate the hidden state given the observations and model parameters, showing that the posterior distribution of the hidden states can be described by differential equations in continuous time. Then we consider the estimation of unknown model parameters, deriving the continuous-time formulas f

关键词： Aeronautics   Receivers & amplifiers   Block codes   Telemetry   Bandwidths   Sensors   Antennas   Design   Spacetime   Transmitters   Codes   Algorithms   Radiation   Ground stations   Aerospace engineering   Computer science   Electrical engineering   Engineering  

摘要： This dissertation explores the application of Silvester's space-time block code to the multi-index CPM called "ARTM CPM" in the IRIG 106 standard to solve the "two antenna problem"---the use of two transmit antennas to provide full spatial coverage on an airborne test article and the accompanying self interference due to different delays between the two transmit antennas and the ground-based receive antenna. A symbol-level encoding scheme is derived that allows the burst-based space-time block code to operate in a continuously streaming mode. The results show that the space-time block code can solve the two antenna problem with differential delays, but that the differential delays generate a substantial increase in the computational complexity of the detector. Complexity-reducing techniques are applied and analyzed. The results show that the complexity reductions required to produce a practically realizable detector render the bit error probability performance sensitive to the differential delay. Numerical results are presented to quantify the performance loss due to the differential delay. The use of space-time coded ARTM CPM to solve the two-antenna problem in aeronautical mobile telemetry requires estimates of the parameters that define the propagation environment. The maximum likelihood estimator problem is defined and used to motivate reduced-complexity estimators suitable for use in a real system. A modified gradient descent algorithm performs the search required to find the delay parameters. An "inner" phase lock loop operating with an "outer" frequency lock loop computes decision-directed estimates of the frequency offset. Computer simulations were used to assess the impact on bit error rate performance introduced by the estimators. The simulation results show the combined joint estimator for the delays, channel gains, and frequency offset imposes a 1.15 dB loss in performance. This loss is approximately the same as the 1.1 dB loss due to the complexity-redu

关键词： Mechanical engineering   Aerospace engineering   Computer science  

摘要： Particle image velocimetry (PIV) techniques provide high fidelity measurements of mutliscale turbulent fluid motion. The research presented in this thesis broadly explores the utilization of Machine Learning for processing and analyzing non-time resolved PIV measurements of turbulent flows. The primary goal is to utilize bayesian inference to automate turbulent coherent structure identification in boundary layer flows. Automated identification was hitherto hindered by the lack of a priori information about the convecting velocities of the coherent structures. A Rankine vortex model is utilized to implement a Markov chain Monte Carlo-based vector matching to overcome this problem. Ultimately, a framework was developed for robust identification of vortices that is capable of visualizing the cumulative distributions of properties of all vortex structures in the flow to provide a rich description of multiscale turbulent coherent structures. Additionally, Bayesian inference is proven to outperform traditional optimization methods based on various loss metrics. The other major goal of this thesis is to implement super resolution in turbulent separated flows using neural network architectures. Machine learning based super resolution using PIV measurements is relatively unexplored, compared to its counterpart with CFD simulation. This is partly due to the unavailability of training datasets with sufficient magnitude. A large PIV dataset of non-time resolved measurements of turbulent separated flow with 50,000 snapshots was collected to overcome this problem. A neural network architecture was trained on this dataset to successfully outperform traditional interpolation based super resolution ***, it has been shown that the current approach also accurately reproduced the properties of turbulent derived quantities, like the stream wise variance of velocity fluctuations, with better accuracy in comparison with interpolation based super resolution methods.

关键词： Robotics   Mechanical engineering   Computer science  

摘要： Recent years, there has been many attempts with different approaches to the human-robot interaction (HRI) problems. In this paper, the multi-agent interaction is formulated as a differential game with incomplete information. To tackle this problem, the parameter estimation method is utilized to obtain the approximated solution in a real time basis. Previous studies in the parameter estimation made the assumption that the human parameters are known by the robot; but such may not be the case and there exists uncertainty in the modeling of the human rewards as well as human's modeling of the robot's rewards. The proposed method, empathetic estimation, is tested and compared with the ``non-empathetic'' estimation from the existing works. The case studies are conducted in an uncontrolled intersection with two agents attempting to pass efficiently. Results have shown that in the case of both agents having inconsistent belief of the other agent's parameters, the empathetic agent performs better at estimating the parameters and has higher reward values, which indicates the scenarios when empathy is essential: when agent's initial belief is mismatched from the true parameters/intent of the agents.

关键词： Biomechanics   Computer science   Medical imaging  

摘要： Digital volume correlation (DVC) is a computational tool used to measure a 3D displacement field between a pair of 3D images (from, for example, magnetic resonance imaging (MRI), computed tomography (CT), ultrasound, etc.). Studies in biomechanics have used DVC to quantify deformations in cells, tissues and organs, for the purpose of examining deformation and failure mechanisms, movement, and adaptation. The growing popularity of DVC has created increased demand for DVC algorithms that are computationally efficient, verified and validated. The goals of this project were to improve the efficiency of an existing DVC algorithm and to present a set of methods for robust verification and validation. This dissertation first introduces DVC through a series of 1D examples that illustrates the use of optimization to find the displacement field that produces the best match between the pair of images. Different methods of regularization are explored. The concept of downsampling of the images is introduced as a way to promote faster convergence and a better image match. In preparation for the move to 3D, the second part of the dissertation covers key concepts in 3D image acquisition and data preparation for the specific case of μCT imaging of human vertebrae. This section allows the reader to appreciate the use of DVC to enable study of failure mechanisms in the spine. The third section addresses the DVC method for 3D images. A custom process is introduced that uses rigid registration of the images to obtain an initial guess for the displacement field. The effect of the quality of the initial guess is then explored using test displacement fields. In the final section, new methods of verification and validation of DVC are presented. An “image-warping” code is presented that interpolates a given displacement field to every voxel of an image, producing a synthetic image. This code is used to warp one image of a pair that was analyzed by DVC, and the mismatch between the synthetic

关键词： Computer science   Artificial intelligence  

摘要： High-performance computing (HPC) resources at facilities such as Argonne National Laboratory's Leadership Computing Facility (ALCF) enable a wide array of scientific experiments and research applications. In day-to-day operation, these platforms collect copious amounts of system, performance, and debugging logs, capturing data about how jobs, individual tasks, and the system as a whole, operate and perform. This thesis builds on previous efforts to examine how these logs can be used to better understand user and application behavior and system resource usage, in addition to demonstrating machine-learning-based (ML-based) techniques for characterizing applications and predicting job behavior using log data. Five datasets collected from the operation of two supercomputers at the ALCF from 2014--2020 were used for the analysis. We first demonstrate that the usage of ALCF supercomputers is consistent, repetitive, and patterned, suggesting that it is suitable for training ML models. We next investigate the usage of workflow-based HPC jobs compared to ``traditional'' single-task HPC jobs, as well as the utilization of solid-state drive (SSD)-based cache drives on ALCF's Theta supercomputer. From these analyses, we enumerate potentially advantageous changes and adaptations the ALCF and other facilities might consider in current and future *** also show that hardware performance counters provide a viable alternative for application identity verification and resource-intensiveness classification using ML-based approaches, accomplishing near-parity in testing accuracy without overhead and coverage constraints faced by prior log-based approaches. Finally, we investigate methods to improve an ML-based technique for application runtime estimation, with implications for job scheduling on HPC systems.