ABSTRACT Cyber‐Physical‐Social System (CPSS) which refers to the complex interaction of cyber, physical and social systems, has the important purpose to provide personalized intelligent services. CPSS data, generated from every aspect of people living life, are mainly in the form of time series multimodal data with characteristic of high order and high dimension. How to efficiently process these CPSS data is one of the fundamental ways for the intelligent services. In this paper, a highly‐efficient attention driven Quantized Tensor Train Recursive Neural Network is proposed, in which the CPSS data is decomposed into the form of tensor train cores. In this way, the proposed method is composed of lightweight high‐order neural network units, which better preserves the multi‐attribute features of the original data and the correlation between different dimensions by using tensors with its calculations, and implicitly trims the dense vector‐matrix connections in the fully connected network by using the form of quantization tensor train decomposition, which greatly reduces the model parameters, shortens the training time and improves the efficiency. Also, an effective attentional feature enhancement module is constructed to assist the high‐order neural network, so that the overall model can achieve a balance between low parameter number and accuracy. The network structure proposed in this paper realizes an efficient and lossless high‐order tensor recurrent neural network model with a small number of parameters. Finally, experiments on the UCF50 action video dataset, CWRU bearing dataset, and image generation tasks are conducted. Comparative analyses with vanilla LSTM and other tensorized LSTMs in terms of training time, accuracy, error, and compression ratio validate the reliability of the proposed model.

Convolutional Neural Network (CNN) and Transformer architectures have been extensively applied in the domain of remote sensing image super-resolution. However, to achieve optimal performance, many existing methods are designed with a large number of parameters, thereby increasing the complexity of the model and hindering practical deployment. To address this issue, we propose an innovative model that synergizes CNN and Transformer architectures while employing a minimal number of their respective modules. This approach significantly reduces the parameter count while maintaining superior performance. Firstly, by constructing additional shallow feature representations as input, we enhance the feature extraction capabilities for individual images. Secondly, we utilize residual connections between various modules to integrate multi-scale, high-dimensional feature information, thus ensuring efficient transmission. Finally, the image reconstruction module is employed to restore the high-resolution image. Experimental results show that FCTNet significantly outperforms existing methods while maintaining a substantially lower parameter count, as demonstrated through evaluations on two public datasets.

Abstract The continuous enhancement of optimization algorithms and their parameters has spurred the expansion of AI into novel application domains such as image recognition and smart home technology. This paper employs the system response curve (SRC) to the adaptive learning rate optimizer, addressing challenges associated with the establishment of the optimizer control model and parameter adjustments affecting the dynamic performance of the system. These insights offer theoretical support for the optimizer's application in deep learning models. To begin, the adaptive learning rate optimizer is a time‐varying system. Based on the intrinsic relationship between the network optimization and the control system, the time domain expression and approximate transfer function of the adaptive learning rate optimizer are derived, and the system dynamic model is established. Furthermore, based on the system control model of the optimizer, it is proposed to explain the performance impacts of different optimizers and their hyperparameters on the deep learning model through the SRC. Finally, experiments are performed on the MNIST, CIFAR‐10, UTKinect‐Action3D, and Florence3D‐Action datasets to validate the control theory of explaining optimizers through system response curves. The experimental results show that the recognition performance of the Adaptive Moment Estimate (Adam) is better than that of the Adaptive Gradient (AdaGrad) and Root Mean Square Propagation (RMSprop). Additionally, the learning rate affects the model training speed, and the practical application aligns with the theoretical analysis.

Abstract Hair modeling plays an important role in computer animation, virtual reality, and other applications. This paper proposes an encoder‐decoder network, named HAO‐CNN, to recover 3D hair strand models from a single image. Specifically, HAO‐CNN generates a volumetric vector field (VVF) from the oriented map of hairstyles. However, instead of directly working on the full resolution VVFs, we introduce the adapted O‐CNN to predict the adaptive representation of VVFs in order to greatly reduce the memory cost. In addition, we fuse the features from different layers of the encoding stage for both capturing the global structure and being aware of hair filaments. Considering the difficulty of acquiring true three‐dimensional (3D) hair models, we augment the dataset with 340 3D hair models by 1,800 hair models via interactive editing using the software and render their oriented maps as training data. Then given a hair photo associated with human head, we segment out the hair region, compute its two‐dimensional oriented map using Gabor filter, and feed it into the network to produce a hair volumetric vector field which is then converted into hairline models using an improved VVF‐to‐strands algorithm. This greatly decreases the time cost of approaches based on volumetric vector fields.