Modeling complex spatial and temporal dependencies in multivariate time series data is crucial for traffic forecasting. Graph convolutional networks have proved to be effective in predicting multivariate time series. Although a predefined graph structure can help the model converge to good results quickly, it also limits the further improvement of the model due to its stationary state. In addition, current methods may not converge on some datasets due to the graph structure of these datasets being difficult to learn. Motivated by this, we propose a novel model named Dynamic Correlation Graph Convolutional Network (DCGCN) in this paper. The model can construct adjacency matrices from input data using a correlation coefficient; thus, dynamic correlation graph convolution is used for capturing spatial dependencies. Meanwhile, gated temporal convolution is used for modeling temporal dependencies. Finally, we performed extensive experiments to evaluate the performance of our proposed method against ten existing well-recognized baseline methods using two original and four public datasets.

Despite the great success of Convolutional Neural Networks (CNNs) in various visual recognition tasks, the high computational and storage costs of such deep networks impede their deployments in real-time remote sensing tasks. To this end, considerable attention has been given to the filter pruning techniques, which enable slimming deep networks with acceptable performance drops and thus implementing them on the remote sensing devices. In this paper, we propose a new scheme, termed Pruning Filter with Attention Mechanism (PFAM), to compress and accelerate traditional CNNs. In particular, a novel correlation-based filter pruning criterion, which explores the long-range dependencies among filters via an attention module, is employed to select the to-be-pruned filters. Distinct from previous methods, the less correlated filters are first pruned after the pruning stage in the current training epoch, and they are reconstructed and updated during the next training epoch. Doing so allows manipulating input data with the maximum information preserved when executing the original training strategy such that the compressed network model can be obtained without the need for the pretrained model. The proposed method is evaluated on three public remote sensing image datasets, and the experimental results demonstrate its superiority, compared to state-of-the-art baselines. Specifically, PFAM achieves a 0.67% accuracy improvement with a 40% model-size reduction on the Aerial Image Dataset (AID) dataset, which is impressive.

Temporal convolution networks (TCNs) are recently proposed to be used in the short-term load forecasting (STLF) tasks in modern smart grids, however, TCNs have two shortcomings, i.e., redundant convolutional operation and equal input importance problems. Therefore, we propose a novel TCN-based backbone model, called PhaCIA-TCNs, to achieve a more accurate short-term load forecasting, where parallel hybrid activated convolution (PhaC) and input attention (IA) are proposed to resolve the above problems, respectively. Specifically, IA is proposed to highlight important input elements while depressing irrelevant ones, which thus rises the model's forecasting accuracies but also brings additional time-cost; then PhaC is further proposed to remedy the efficiency problem and to further enhance the forecasting accuracies by shortening the convolutional learning path to overcome the redundant convolutional operation problem. Extensive experimental results show that (i) PhaCIA-TCNs significantly outperform all state-of-the-art RNN-based and TCNs-based baselines in forecasting-error-based evaluation metrics on all datasets; (ii) ablation studies show that PhaC and IA are both effective and essential for PhaCIA-TCN to achieve the superior forecasting accuracies in STLF tasks, and by integrating IA and PhaC with TCN, the proposed PhaCIA-TCN not only greatly outperforms TCN in forecasting accuracies but also keeps similar (sometimes even better) learning efficiency.

Medical image segmentation is an important and complex task in clinical practices, but the widely used U-Net usually cannot achieve satisfactory performances in some clinical challenging cases. Therefore, some advanced variants of U-Net are proposed using multi-scale and attention mechanisms. Different from the existing works where multi-scale and attention are usually used independently, in this work, we integrate them together and propose a collaborative attention guided multi-scale feature fusion with enhanced convolution based U-Net (EC-CaM-UNet) model for more accurate medical image segmentation, where a novel collaborative attention guided multi-scale feature fusion (CoAG-MuSFu) module is proposed to highlight important (but small and unremarkable) multi-scale features and suppress irrelevant ones in model learning. Specifically, CoAG-MuSF uses a multi-dimensional collaborative attention (CoA) block to estimate the local and global self-attention, which is then deeply fused with the multi-scale feature maps generated by a multi-scale (MuS) block to better highlight the important multi-scale features and suppress the irrelevant ones. Furthermore, an additional supervision path and enhanced convolution blocks are used to enhance the deep model's feature learning ability in both deep and shallow features, respectively. Experimental results on three public medical image datasets show that EC-CaM-UNet greatly outperforms the state-of-the-art medical image segmentation baselines. The codes will be released after acceptance.

Edge computing is a popular paradigm in solving the problems of long time delay and high energy consumption in Internet of Things (IoT) network, which can effectively realize the IoT task offloading by collaboration of multiple edge servers. Nevertheless, how to choose the appropriate edge servers for offloading the dependent subtasks is still a big challenge, considering the limited resources and computing power of the edge servers as well as the start and end execution time of each subtask. These factors have a great impact on the execution efficiency of the whole task. At present, most of the research works focus on single-hop or multi-hop task offloading, where the edge servers farther away are not considered in the offloading decision. Such task offloading strategy is not optimal, and difficult to achieve high parallel execution of tasks, resulting in some delay-sensitive tasks not being completed within the specified time. In this paper, a two-stage optimization method is proposed to solve the resource allocation problem between edge servers and tasks. In the first stage, we group tasks according to their priorities, and the group with a higher priority is given the preference to resource allocation, thereby ensuring the timeliness of delay-sensitive tasks. Within the same group, resources are competed according to the game theory, and the total delay of all tasks is optimized. In the second stage, we aim to optimize the energy consumption of each task without increasing its completion time by allocating the computing resources to its subtasks based on their maximum completion time. For group resource allocation, we propose a spatial index tree to store the information of all edge servers for optimal server selection. During the selection process, an online learning based double prediction model is utilized to reduce the energy consumption caused by information transmission. We have evaluated the performance of the experiment on iFogSim simulator, and the experimental results show that our proposed method can achieve better performance in terms of time delay and energy consumption.