摘要 :
An enhanced method based on radial basis functions is proposed for large mesh deformation. The traditional RBF method only uses the displacements of boundary nodes to interpolate the displacements of nodes inside the computational...
展开
An enhanced method based on radial basis functions is proposed for large mesh deformation. The traditional RBF method only uses the displacements of boundary nodes to interpolate the displacements of nodes inside the computational domain. Because of the lack of control from the inner region, the elements squeeze around the moving boundaries in the deformed mesh. And inverted elements tend to appear over this region at an early stage as the deformation continues. To enhance the capability of the RBF method for large deformation, we investigate a way to add control from the inner region when transferring the displacements from boundary nodes to inner mesh nodes. Some virtual nodes inside the computational domain are created, and their displacements interpolated by nearby boundary nodes are also used to compute the displacements of inner mesh nodes. Numerical results of three test cases demonstrate the effectiveness of the proposed method.
收起
摘要 :
Fueled by the power of deep learning, stereo vision has made unprecedented advances in recent years. Existing deep stereo models, however, can be hardly deployed to real-world scenarios where the data comes on-the-fly without any ...
展开
Fueled by the power of deep learning, stereo vision has made unprecedented advances in recent years. Existing deep stereo models, however, can be hardly deployed to real-world scenarios where the data comes on-the-fly without any ground-truth information, and the data distribution continuously changes over time. Recently, Tonioni et al. proposed the first real-time self-adaptive deep stereo system (MADNet) to address this problem, which, however, still runs at a relatively low speed with not so satisfactory performance. In this paper, we significantly upgrade their work in both speed and accuracy by incorporating two key components. First, instead of adopting only the image reconstruction loss as the proxy supervision, a second more powerful supervision is proposed, termed Knowledge Reverse Distillation (KRD), to guide the learning of deep stereo models. Second, we introduce a straightforward yet surprisingly effective Adapt-or-Hold (AoH) mechanism to automatically determine whether or not to fine-tune the stereo model in the online environment. Both components are lightweight and can be integrated into MADNet with only a few lines of code. Experiments demonstrate that the two proposed components improve the system by a large margin in both speed and accuracy. Our final system is twice as fast as MADNet, meanwhile attains considerable superior performance on the popular benchmark datasets KITTI.
收起
摘要 :
Fueled by the power of deep learning, stereo vision has made unprecedented advances in recent years. Existing deep stereo models, however, can be hardly deployed to real-world scenarios where the data comes on-the-fly without any ...
展开
Fueled by the power of deep learning, stereo vision has made unprecedented advances in recent years. Existing deep stereo models, however, can be hardly deployed to real-world scenarios where the data comes on-the-fly without any ground-truth information, and the data distribution continuously changes over time. Recently, Tonioni et al. proposed the first real-time self-adaptive deep stereo system (MADNet) to address this problem, which, however, still runs at a relatively low speed with not so satisfactory performance. In this paper, we significantly upgrade their work in both speed and accuracy by incorporating two key components. First, instead of adopting only the image reconstruction loss as the proxy supervision, a second more powerful supervision is proposed, termed Knowledge Reverse Distillation (KRD), to guide the learning of deep stereo models. Second, we introduce a straightforward yet surprisingly effective Adapt-or-Hold (AoH) mechanism to automatically determine whether or not to fine-tune the stereo model in the online environment. Both components are lightweight and can be integrated into MADNet with only a few lines of code. Experiments demonstrate that the two proposed components improve the system by a large margin in both speed and accuracy. Our final system is twice as fast as MADNet, meanwhile attains considerable superior performance on the popular benchmark datasets KITTI.
收起
摘要 :
Low-Earth orbit satellite networks have received attention from academia and industry for their advantages in terms of wide coverage and low latency. Meantime deep learning can provide more accurate traffic anomaly detection and h...
展开
Low-Earth orbit satellite networks have received attention from academia and industry for their advantages in terms of wide coverage and low latency. Meantime deep learning can provide more accurate traffic anomaly detection and has become an important class of methods for LEO satellite network security. However, deep learning is susceptible to adversarial sample attacks, and the LEO satellite network system has not been investigated to find a framework for adversarial sample attacks and defence systems, which poses a potential risk to network communication security. In this paper, we design a framework to generate and defend against adversarial samples in real time. By capturing traffic from LEO satellites, it can generate traffic adversarial samples to detect whether malicious traffic classification models are vulnerable to attacks, and defense against adversarial sample attacks in real time. In this paper, a simple LEO satellite simulation platform is built to generate traffic adversarial samples using four classical adversarial sample generation methods, and a two-classification deep learning model is trained to determine the effectiveness of the attack and defence. Experiments show that the framework proposed in the paper can crawl traffic and perform self-attack and defence tests.
收起
摘要 :
Deploying edge caching ability can significantly reduce the traffic volumes between edge networks and cloud, so as to save energy. However, how to intelligently use the communicating and caching capabilities of edge nodes/cloudlet...
展开
Deploying edge caching ability can significantly reduce the traffic volumes between edge networks and cloud, so as to save energy. However, how to intelligently use the communicating and caching capabilities of edge nodes/cloudlets is challenging because of users' mobility and limited edge compute power. This paper proposes an energy aware dynamic cache policy for virtualized mobile edge networks using Ant-Q learning algorithms. We first design a paradigm of software defined mobile edge networks with edge caching ability, which can more accurately and efficiently extract characteristics from network services than from mobile users' behavior, and build its system model. Then we formulate the energy-aware cache optimization problem into a reinforcement learning (RL) model, and solve the problem with a dynamic cache policy use Ant-Q algorithm, which can efficiently combine strong computing capacity of cloud and the responsiveness of edge networks. Simulation results show that the proposed dynamic cache policy can enhance energy efficiency and improve network performance.
收起
摘要 :
The Hyperloop system is a novel conceptual system aimed to provide a high-speed public transportation service in the future, featured with a reduced-pressure tube in which pressurized capsules ride on a cushion of air that is driv...
展开
The Hyperloop system is a novel conceptual system aimed to provide a high-speed public transportation service in the future, featured with a reduced-pressure tube in which pressurized capsules ride on a cushion of air that is driven by a combination of linear induction motors and air compressor. This work involves numerical simulations based on finite volume method to study the effects of different factors on the aerodynamic drag on a capsule running at subsonic and transonic speeds in Hyperloop system. Investigation includes the study of the effects of the internal tube pressure, operating speed, vehicle shape and air compressor on aerodynamic drag on a high-speed capsule. The compressible Navier-Stokes equations were solved by using k - w SST turbulent modelling. The simulated results show that the operating speed and different working vacuum pressure significantly affects the aerodynamic drag of the capsule. Investigations with respect to different shapes of the head as well as that of the tail indicate the optimum shape for minimum drag. Suction mechanism at the head was applied to study the additional reduction effect of the aerodynamic drag on the capsule. It was observed that aerodynamic drag was decreasing with the on-board air compressor, which is related to a means of increasing the maximum operating speed over a closed pod and provide a small amount of thrust. Those results provide guidelines for the initial design and construction of the Hyperloop system. Hyperloop; aerodynamic drag; vacuum degree; air compressor; capsule shape
收起
摘要 :
The Hyperloop system is a novel conceptual system aimed to provide a high-speed public transportation service in the future, featured with a reduced-pressure tube in which pressurized capsules ride on a cushion of air that is driv...
展开
The Hyperloop system is a novel conceptual system aimed to provide a high-speed public transportation service in the future, featured with a reduced-pressure tube in which pressurized capsules ride on a cushion of air that is driven by a combination of linear induction motors and air compressor. This work involves numerical simulations based on finite volume method to study the effects of different factors on the aerodynamic drag on a capsule running at subsonic and transonic speeds in Hyperloop system. Investigation includes the study of the effects of the internal tube pressure, operating speed, vehicle shape and air compressor on aerodynamic drag on a high-speed capsule. The compressible Navier-Stokes equations were solved by using k - w SST turbulent modelling. The simulated results show that the operating speed and different working vacuum pressure significantly affects the aerodynamic drag of the capsule. Investigations with respect to different shapes of the head as well as that of the tail indicate the optimum shape for minimum drag. Suction mechanism at the head was applied to study the additional reduction effect of the aerodynamic drag on the capsule. It was observed that aerodynamic drag was decreasing with the on-board air compressor, which is related to a means of increasing the maximum operating speed over a closed pod and provide a small amount of thrust. Those results provide guidelines for the initial design and construction of the Hyperloop system. Hyperloop; aerodynamic drag; vacuum degree; air compressor; capsule shape
收起
摘要 :
This paper presents a symbolic model checking approach for Alternating Projection Temporal Logic (APTL). In our approach, the model of a system to be verified is specified by an Interpreted System IS, and a property of the system ...
展开
This paper presents a symbolic model checking approach for Alternating Projection Temporal Logic (APTL). In our approach, the model of a system to be verified is specified by an Interpreted System IS, and a property of the system is expressed by an APTL formula φ. To check whether φ is valid on IS or not: first, the system IS is symbolically represented and -φ is transformed into its normal form. Then, the set Sat(-φ), containing all the states from which there exists at least one computation such that -φ holds, is computed. Finally, whether the property is valid on the system is equivalently evaluated by checking the emptiness of the intersection of the set of initial states in the system and Sat(-φ). Supporting tool is also developed to show how the proposed approach works in practice.
收起
摘要 :
This paper presents a symbolic model checking approach for Alternating Projection Temporal Logic (APTL). In our approach, the model of a system to be verified is specified by an Interpreted System IS, and a property of the system ...
展开
This paper presents a symbolic model checking approach for Alternating Projection Temporal Logic (APTL). In our approach, the model of a system to be verified is specified by an Interpreted System IS, and a property of the system is expressed by an APTL formula φ. To check whether φ is valid on IS or not: first, the system IS is symbolically represented and ﹁φ is transformed into its normal form. Then, the set Sat(﹁φ), containing all the states from which there exists at least one computation such that ﹁φ holds, is computed. Finally, whether the property is valid on the system is equivalently evaluated by checking the emptiness of the intersection of the set of initial states in the system and Sat(﹁φ). Supporting tool is also developed to show how the proposed approach works in practice.
收起
摘要 :
The sequential recommendation is a prominent task that aims to provide accurate recommendations by leveraging users' historical behavior. However, the challenge of data sparsity poses a significant obstacle in achieving effective ...
展开
The sequential recommendation is a prominent task that aims to provide accurate recommendations by leveraging users' historical behavior. However, the challenge of data sparsity poses a significant obstacle in achieving effective sequential recommendations. In this paper, we propose a User Feedback-based Counterfactual data augmentation method for Sequential Recommendation (UFC4-SRec) to address this challenge. Our approach focuses on expanding the dataset for sequential recommendation tasks by employing counterfactual inference techniques. The UFC4-SRec method consists of two main components: a counter-factual generator and a recommender. The counterfactual generator is responsible for generating counterfactual examples based on users' feedback. By incorporating users' preferences for items, the generated counterfactual data are designed to be closer to their actual preferences. On the other hand, the recommender employs various sequential recommendation models to provide recommendation results. To guide the counterfactual generator, the recommender imitates reinforcement learning by computing reward values based on the quality of the generated data. To evaluate the effectiveness of our method, we conduct experiments on three real-world datasets. The experimental results demonstrate that our UFC4-SRec approach significantly improves the performance of sequential recommendation tasks. Moreover, it effectively addresses the data sparsity problem commonly encountered in sequential recommendations.
收起