Convolutional Neural Networks (CNNs) have performed very well on image classification tasks, but CNNs is insensitive to detailed image information and requires a large amount of training data and time. Capsule Networks(CapsNets) can solve this problem very well, but the Baseline CapsNet model is very shallow, and the extraction of low-level features is not enough. We propose a Multi-Scale Capsule Network (Multi-Scale CapsNet), by extracting the low-level features of images with multi-channel convolution of multiple convolution kernels, so extracted features are more diverse, then passing from the bottom layer to the upper layer in the form of a "capsule", which encapsulats the multidimensional features of the image in the form of a vector, thus the features are saved in the network, rather than being recovered after being lost. In the German Traffic Sign Recognition Benchmark(GTSRB), we obtained competitive results with the accuracy of 99.4%, which is better than the human performance of 98.81% and the Multi-Scale Convolutional Neural Network(MS-CNN) of 97.33%.

Information aggregation is an essential component of text encoding, but it has been paid less attention. The pooling-based (max or average pooling) aggregation method is a bottom-up and passive aggregation method, and loses a lot of important information. Recently, attention mechanism and dynamic routing policy are separately used to aggregate information, but their aggregation capabilities can be further improved. In this paper, we proposed an novel aggregation method combining attention mechanism and dynamic routing, which can strengthen the ability of information aggregation and improve the quality of text encoding. Then, a novel Leaky Natural Logarithm (LNL) squash function is designed to alleviate the “saturation” problem of the squash function of the original dynamic routing. Layer Normalization is added to the dynamic routing policy for speeding up routing convergence as well. A series of experiments are conducted on five text classification benchmarks. Experimental results show that our method outperforms other aggregating methods.