Physics-Informed Neural Networks (PINNs) have recently received increasing attention, however, optimizing the loss function of PINNs is notoriously difficult, where the landscape of the loss function is often highly n...
详细信息
ISBN:
(数字)9798350308365
ISBN:
(纸本)9798350308372
Physics-Informed Neural Networks (PINNs) have recently received increasing attention, however, optimizing the loss function of PINNs is notoriously difficult, where the landscape of the loss function is often highly non-convex and rugged. Local optimization methods based on gradient information can converge quickly but are prone to being trapped in local minima for training PINNs. Evolutionary algorithms (EAs) are well known for the global search ability, which can help escape from local minima. It has been reported in the literature that EAs show some advantages over gradient-based methods in training PINNs. Inspired by the Memetic Algorithm, we combine global-search based EAs and local-search based batch gradient descent in order to make the best of both word. In addition, since the PINN loss function is composed of multiple terms, balancing these terms is also a challenging issue. Therefore, we also attempt to combine EAs with multiple-gradient descent algorithm for multi-objective optimization. Our experiments provide strong evidence for the superiority of the above algorithms.
Recent researches have shown that Large Language Models (LLMs) are susceptible to a security threat known as Backdoor Attack. The backdoored model will behave well in normal cases but exhibit malicious behaviours on i...
详细信息
Single Image Super Resolution (SISR) is a critical facet of image processing dedicated to reconstructing high-resolution (HR) images from low-resolution (LR) counterparts. Recently, the two major issues persist in dee...
详细信息
Single Image Super Resolution (SISR) is a critical facet of image processing dedicated to reconstructing high-resolution (HR) images from low-resolution (LR) counterparts. Recently, the two major issues persist in deep learning-based SISR methods: significant loss of high-frequency information in content loss based explicit methods and pattern collapse along with limited diversity in GAN based implicit methods. To overcome these obstacles, we present Edge Fusion Diffusion (EFD), a novel diffusion-based SISR model. EFD's innovative approach integrates image edge fusion to effectively capture and utilize local image details, thereby enhancing the reconstruction of high-frequency information. By incorporating our Edge Conditional block (EC block), EFD improves image generation while ensuring a stable training process. Our EC block replaces some residual blocks with SENet network structure blocks and adds additional SENet blocks to all upsampling and downsampling modules, enhancing the network architecture's depth and effectiveness in edge synthesis. Unlike other generative models, EFD significantly enhances image quality, specifically excelling in detailed edge synthesis. Compared to existing diffusion models, EFD stands out in extracting local details, improving edge effects, and substantially boosting the overall quality of synthesized images. Our comprehensive evaluation on the DIV2K dataset and six other benchmark datasets demonstrates that EFD achieves state-of-the-art performance in SISR tasks, delivering diverse and high-quality results.
The Ultra-Wideband (UWB) technology boasts notable attributes, including robust signal penetration capabilities, exceptional resistance to interference, precision in positioning, and admirable signal stability. These ...
详细信息
Causal relationships are a scientific research method used to describe relationships between variable data, allowing for the excavation of the deep logic and operational mechanisms behind phenomena. The integration of...
详细信息
We propose a novel decentralized federated learning framework called B2DFL. It decomposes the aggregation process of vanilla FL into layered and serialized sub-aggregation processes and offloads the communication and ...
详细信息
We propose a novel decentralized federated learning framework called B2DFL. It decomposes the aggregation process of vanilla FL into layered and serialized sub-aggregation processes and offloads the communication and computation from a single point to distributed nodes, thus addressing the single point of failure issue in centralized FL. The decentralization of B2DFL is based on the Butterfly, a distributed network topology, to organize and orchestrate the order and rules of node aggregation. Additionally, to mitigate potential risks such as dropouts or tampering, we leverage the blockchain and IPFS systems. Specifically, after each node completes its computation (including training and aggregation), it generates a hash value of the results as proof. We maintain a Tamper- evident Data Structure (TDS) on the blockchain, which records these proofs to ensure tamper-proofing and fast verification. To reduce the storage burden on the blockchain and improve throughput, we store the aggregated results on IPFS, a system that enables quick data location through hash values of data, for data backup. We also design a node replacement mechanism for quick dropout handling. We conduct a comprehensive performance evaluation and experimental results demonstrate that B2DFL presents a significant performance improvement while achieving privacy and decentralization.
The rapid advancements in time-series forecasting have led to the emergence of datasets with increasingly diverse characteristics. Researchers typically focus on designing robust algorithms to handle these datasets. H...
详细信息
In recent years, the use of CCTV footage for proactive crime prevention has surged, particularly in public places like airports, train stations, and malls. However, the efficacy of these surveillance systems becomes q...
详细信息
Intelligent traffic signal control plays a crucial role in reducing the escalating problem of traffic congestion. However, traditional methods of traffic signal control struggle to effectively adapt to the ever-changi...
详细信息
In this work,we develop energy stable numerical methods to simulate electromagnetic waves propagating in optical media where the media responses include the linear Lorentz dispersion,the instantaneous nonlinear cubic ...
详细信息
In this work,we develop energy stable numerical methods to simulate electromagnetic waves propagating in optical media where the media responses include the linear Lorentz dispersion,the instantaneous nonlinear cubic Kerr response,and the nonlinear delayed Raman molecular vibrational *** the first-order PDE-ODE governing equations considered previously in Bokil et al.(J Comput Phys 350:420–452,2017)and Lyu et al.(J Sci Comput 89:1–42,2021),a model of mixed-order form is adopted here that consists of the first-order PDE part for Maxwell’s equations coupled with the second-order ODE part(i.e.,the auxiliary differential equations)modeling the linear and nonlinear dispersion in the *** main contribution is a new numerical strategy to treat the Kerr and Raman nonlinearities to achieve provable energy stability property within a second-order temporal discretization.A nodal discontinuous Galerkin(DG)method is further applied in space for efficiently handling nonlinear terms at the algebraic level,while preserving the energy stability and achieving high-order *** with d_(E)as the number of the components of the electric field,only a d_(E)×d_(E)nonlinear algebraic system needs to be solved at each interpolation node,and more importantly,all these small nonlinear systems are completely decoupled over one time step,rendering very high parallel *** evaluate the proposed schemes by comparing them with the methods in Bokil et al.(2017)and Lyu et al.(2021)(implemented in nodal form)regarding the accuracy,computational efficiency,and energy stability,by a parallel scalability study,and also through the simulations of the soliton-like wave propagation in one dimension,as well as the spatial-soliton propagation and two-beam interactions modeled by the two-dimensional transverse electric(TE)mode of the equations.
暂无评论