Search

• The departure delay is predicted based on node sequence data of ground support services for transit flights. • The cases of Shanghai Pudong and Kunming Changshui airports are applied for model illustration. • A Bayesian network is constructed for interrelationship analysis between segments. • A GCN-GraphSAGE model is proposed for departure delay prediction. • The proposed GCN-GraphSAGE method is reported to outperform the other benchmark methods. Delays frequently occur during ground support services for transit flights. Due to the limitation of related data acquisition, the delay occurred during this process has seldom been explored. Toward the purpose of departure delay mitigation, the present research provides the node sequence data to identify the critical ground support service nodes that associate with departure delay. Machine learning techniques are used for accurate departure delay prediction of transit flights. First, the topological structure of the neural network is constructed according to the ground service support procedure of transit flights. The flight data from Shanghai Pudong and Kunming Changshui International airports are collected for delay description and model evaluation. Then, similarity of delay scenarios are illustrated and the typical delay patterns are explored based on clustering methods. It is found that there are some node sections that delays routinely occur. Third, a Bayesian network is constructed for delay probability analysis. The conditional probability table describing the impact of delay occurred during each section on the departure delay for transit flights is provided. Finally, the performance of two classes of approaches to predict departure delays are constructed and compared, including three traditional machining learning methods denoted as random forest, decision tree, XGBoost models, and four graph convolutional neural network models namely Graph convolutional network (GCN), Graph attention network (GAT), Graph sample and aggregate (GraphSAGE) and integrated Graph convolutional network and Graph sample and aggregate (GCN-GraphSAGE) models. In general, the proposed GCN-GraphSAGE method achieves the most stable and satisfactory performance for departure delay prediction. The research results can support the dynamic and accurate prediction of departure delays and the refined management of transit flights for delay alleviation. [ABSTRACT FROM AUTHOR]

The gap between Neural Architecture Search (NAS) and hardware embedded accelerators degrades the deployment efficiency, due to the absent of rethinking the applicability of the searched network layer characteristics and hardware mapping. Therefore, this work proposes a novel hardware-aware NAS framework in consideration of a deduced efficiency metric. Beside, a layer adaptive scheduler and a coarse-grained reconfigurable computing architecture are developed to deploy the searched networks by selecting the most appropriate dataflow pattern layer-by-layer. Evaluation results show that the proposed NAS framework can search networks with high accuracy and low inference latency, and the proposed architecture provides power-efficiency improvement. • Bridge the applicability gap between NAS networks and hardware embedded accelerators. • A novel hardware-aware NAS framework that incorporates a deduced efficiency metric. • A layer adaptive scheduler that generates the optimal dataflow and control sequence. • A flexible coarse-grained reconfigurable architecture to deploy searched networks. [ABSTRACT FROM AUTHOR]

As the deep neural networks (DNNs) with satisfied expression ability usually require large scale to gain adequate performance, and deploying large-scale DNNs on resource-limited environments is still a challenge, neural network compression becomes a hot topic nowadays. Among the multiple compression methods, tensor decomposition reveals many specific advantages, such as regular data structure comes from linear algebra, convenient approach of training from scratch, ideal compression ratio, etc. Nevertheless, for some compact neural modules such as two-dimensional/three-dimensional (2D/3D) convolutional kernels, traditional tensor decomposition in the way of approximation has encountered intractable obstacles. Fortunately, some works utilize the tensor mapping approach, which just regards the data structure of tensor decomposition as neural layers, to reconstruct the convolutional kernel into a new lightweight module with several thinner convolutional kernels. The only two flies in the ointment are there lack necessary theories of tensor mapping, and there is still no tensor mapping way to compress high-order three-dimensional convolutional neural networks (3DCNNs). In this paper, we first deeply analyze the tensor mapping theory including convergence and precision, which separately establishes the rationality of tensor mapping and its superiority over the traditional tensor approximation, according to the Lottery Ticket Hypothesis. Then we propose an efficient method termed as 3D-KCPNet, to compress 3DCNNs based on the novel Kronecker canonical polyadic (KCP) tensor decomposition. The proposed method can not only compress the 3D convolutional kernels, but also convert a 3D convolution to efficient 1 × 1 × 1 and 2D depthwise convolutions. The experiments on the video recognition datasets VIVA Challenge, UCF11, UCF50, and UCF101 show that the accuracy of 3D-KCPNet can surpass its original baseline model and the corresponding tensor approximation model. [ABSTRACT FROM AUTHOR]