Nonlocal spatial attention module for image classification

To enhance the capability of neural networks, research on attention mechanism have been deepened. In this area, attention modules make forward inference along channel dimension and spatial dimension sequentially, parallelly, or simultaneously. However, we have found that spatial attention modules mainly apply convolution layers to generate attention maps, which aggregate feature responses only based on local receptive fields. In this article, we take advantage of this finding to create a nonlocal spatial attention module (NL-SAM), which collects context information from all pixels to adaptively recalibrate spatial responses in a convolutional feature map. NL-SAM overcomes the limitations of repeating local operations and exports a 2D spatial attention map to emphasize or suppress responses in different locations. Experiments on three benchmark datasets show at least 0.58% improvements on variant ResNets. Furthermore, this module is simple and can be easily integrated with existing channel attention modules, such as squeeze-and-excitation and gather-excite, to exceed these significant models at a minimal additional computational cost (0.196%).

Short-Term Travel Time Prediction: A Spatiotemporal Deep Learning Approach

Traffic prediction is a complex, nonlinear spatiotemporal relationship modeling task with the randomness of traffic demand, the spatial and temporal dependency between traffic flows, and other recurrent and nonrecurrent factors. Based on the ability to learn generic features from history information, deep learning approaches have been recently applied to traffic prediction. Convolutional neural network (CNN) methods that learn traffic as images can improve the predictive accuracy by leveraging the implicit correlations among nearby links. Traffic prediction based on CNN is still in its initial stage without making full use of spatiotemporal traffic information. In this paper, we improve the predictive accuracy by directly capturing the relationship between the input sequence and the predicted value. We propose the new local receptive fields for spatiotemporal traffic information to provide the constraints in the task domain for CNN which is different from traditionally learning traffic as images. We explore a max-pooled CNN followed by a fully connected layer with a nonlinear activation function to convolute the new local receptive fields. The higher global-level features are fed into a predictor to generate the predicted output. Based on the dataset provided by Highways England, we validate the assumption that there exists direct relationship between the input sequence and the predicted value. We train the proposed method by using the backpropagation approach, and we employ the AdaGrad method to update the parameters of the proposed method. The experimental results show that the proposed method can improve the predictive accuracy.