Search

Purpose: The purpose of this paper is to present a deep-learning-based beamforming method for phased array weather radars, especially whose antenna arrays are equipped with large number of elements, for fast and accurate detection of weather observations. Design/methodology/approach: The beamforming weights are computed by a convolutional neural network (CNN), which is trained with input–output pairs obtained from the Wiener solution. Findings: To validate the robustness of the CNN-based beamformer, it is compared with the traditional beamforming methods, namely, Fourier (FR) beamforming and Capon beamforming. Moreover, the CNN is compared with a radial basis function neural network (RBFNN) which is a shallow type of neural network. It is shown that the CNN method has an excellent performance in radar signal simulations compared to the other methods. In addition to simulations, the robustness of the CNN beamformer is further validated by using real weather data collected by the phased array radar at Osaka University (PAR@OU) and compared to, besides the FR and RBFNN methods, the minimum mean square error beamforming method. It is shown that the CNN has the ability to rapidly and accurately detect the reflectivity of the PAR@OU with even less clutter level in comparison to the other methods. Originality/value: Motivated by the inherit advantages of the CNN, this paper proposes the development of a CNN-based approach to the beamforming of PAR using both simulated and real data. In this paper, the CNN is trained on the optimum weights of Wiener solution. In simulations, it is applied on a large 32 × 32 planar phased array antenna. Moreover, it is operated on real data collected by the PAR@OU. [ABSTRACT FROM AUTHOR]

• Firstly, compared with the homogeneous system, the more general heterogeneous coupled neural network model is studied here. By processing the heterogeneous terms, the mean-square quasi-synchronization problem under the influence of stochastic disturbance is solved. • Secondly, on the basis of discussing the influence of communication time-delay, this paper also considers various time-delays such as distributed time-delay, transmission time-delay. According to the existing conclusions of time-delay systems, we extend it to heterogeneous CNNs, and obtain a sufficient condition for time-invariant delay. • Finally, in the reaction-diffusion environment, compared with the conventional controller, a more advantageous boundary controller is designed. At the same time, the corresponding boundary control under mixed boundary conditions is proposed, and the final system is uniformly ultimately bounded. In this paper, the quasi-synchronization problem of heterogeneous stochastic coupled neural networks (HSCNNs) is discussed. The effects of the mixed time-varying delay and diffusion phenomenon on the system are considered separately in time and space. Moreover, different from the previous distributed control, boundary control is introduced to realize network synchronization. This not only reduces the space cost of the controller, but also makes it easier to implement. Thus, the mean-square quasi-synchronization of HSCNNs is guaranteed by using matrix inequality and stochastic analysis tools. In addition to focusing on systems with Neumann boundary conditions, we briefly investigate HSCNNs with time-invariant delays and mixed boundary conditions respectively, and provide sufficient conditions to achieve the desired performance. Finally, the correctness of the conclusion is verified by several examples. [ABSTRACT FROM AUTHOR]

Traffic classification is essential in network management for a wide range of operations. Recently, it has become increasingly challenging with the widespread adoption of encryption in the Internet, for example, as a de facto in HTTP/2 and QUIC protocols. In the current state of encrypted traffic classification using deep learning (DL), we identify fundamental issues in the way it is typically approached. For instance, although complex DL models with millions of parameters are being used, these models implement a relatively simple logic based on certain header fields of the TLS handshake, limiting model robustness to future versions of encrypted protocols. Furthermore, encrypted traffic is often treated as any other raw input for DL, while crucial domain-specific considerations are commonly ignored. In this paper, we design a novel feature engineering approach used for encrypted Web protocols, and develop a neural network architecture based on stacked long short-term memory layers and convolutional neural networks. We evaluate our approach on a real-world Web traffic dataset from a major Internet service provider and mobile network operator. We achieve an accuracy of 95% in service classification with less raw traffic and a smaller number of parameters, outperforming a state-of-the-art method by nearly 50% fewer false classifications. We show that our DL model generalizes for different classification objectives and encrypted Web protocols. We also evaluate our approach on a public QUIC dataset with finer application-level granularity in labeling, achieving an overall accuracy of 99%. [ABSTRACT FROM AUTHOR]

Forgery of a signature with the aim of deception is a serious crime. Machine learning is often employed to detect real and forged signatures. In this study, we present results which argue that robotic arms and generative models can overcome these systems and mount false-acceptance attacks. Convolutional neural networks and data augmentation strategies are tuned, producing a model of 87.12% accuracy for the verification of 2,640 human signatures. Two approaches are used to successfully attack the model with false-acceptance of forgeries. Robotic arms (Line-us and iDraw) physically copy real signatures on paper, and a conditional Generative Adversarial Network (GAN) is trained to generate signatures based on the binary class of 'genuine' and 'forged'. The 87.12% error margin is overcome by all approaches; prevalence of successful attacks is 32% for iDraw 2.0, 24% for Line-us, and 40% for the GAN. Fine-tuning with examples show that false-acceptance is preventable. We find attack success reduced by 24% for iDraw, 12% for Line-us, and 36% for the GAN. Results show exclusive behaviours between human and robotic forgers, suggesting training wholly on human forgeries can be attacked by robots, thus we argue in favour of fine-tuning systems with robotic forgeries to reduce their prevalence. • Development of a computer vision-based system for signature spoofing attack detection. • A Conditional GAN can generate "real" and "fake" signatures. • Two robots can physically replicate human signatures with pen and paper. • The GAN and both robots can fool the model and mount false-acceptance attacks. • Verification model can be defended by fine-tuning on generative and robotic forgeries. [ABSTRACT FROM AUTHOR]

Data augmentation strategies have always been important in machine learning techniques and play a unique role in model performance optimization processes. Therefore, in recent years, these techniques have become popular in the artificial intelligence field. In this paper, a new data augmentation strategy is proposed based on the interpretation algorithm of deep convolutional neural networks, i.e., constructing new training samples by deeply exploiting key features extracted from interpretable networks to achieve sample augmentation. Thus, a novel supervised data augmentation approach known as Supervised Data Augmentation–Key Feature Extraction (SDA-KFE) was proposed. By introducing the Neural Network Interpreter-Segmentation Recognition and Interpretation (NNI-SRI) algorithm, an augmentation strategy is proposed that can balance the high accuracy and high robustness of the final model while ensuring a large amount of data augmentation. The advantages of the SDA-KFE algorithm are mainly reflected in the following aspects. First, it is easy to implement. This algorithm is implemented based on the lightweight NNI-SRI algorithm, which lays the foundation for the implementation of SDA-KFE so that it can be easily implemented on convolutional neural networks. Second, this model, which is widely applicable, can be applied to almost any deep convolutional network. Through research and experiments on this proposed algorithm, SDA-KFE can be applied in graphical image binary classification and multiclassification models. Third, SDA-KFE can rapidly construct data samples with diverse variations. Under the premise of determining the classification labels of the generated samples, the distribution of the feature unit composition of the samples can be controlled. Compared with traditional data augmentation methods, SDA-KFE can control the direction of the model performance, i.e., the balance between the pursuit of high accuracy and robust performance of the model. Therefore, the novel supervised augmentation approach proposed in this paper is relevant for optimizing deep convolutional neural networks, solving model overfitting, augmenting data types, etc. The data augmentation algorithm proposed in this paper can be regarded as a useful supplement to traditional data augmentation methods, such as horizontal or vertical image flipping, cropping, color transformation, extension and rotation. [ABSTRACT FROM AUTHOR]

Moisture diffusion is a common phenomenon in geotechnical engineering, and its induced deformation seriously affects the stability of the engineering structure, such as embankment slope instability and tunnel surrounding rock deformation. Numerical simulation is an effective method for moisture diffusion-deformation coupling computation, but it has a large computational cost and a high learning threshold for ordinary engineers. In this paper, a surrogate model based on deep convolutional neural networks is presented for moisture diffusion-deformation coupling computation. First, designing the neural network structure includes the dense blocks and transition layers, and hyperparameters. Then, the moisture diffusion-deformation coupling model in the finite discrete element method (FDEM) software package MultiFracS is used to obtain the high-fidelity simulation data for the moisture diffusion-deformation examples. The simulation data and the key parameter (elastic modulus and Poisson's ratio) are processed into the image data structure (matrix) for training the surrogate model. Finally, the root means square error and the correlation coefficient are used to evaluate the effectiveness of the surrogate model. The results reveal that, rather than taking several hours to run a numerical model, the surrogate model only takes a few seconds to obtain the deformation and stress under a given moisture field and material parameters, which significantly improves prediction efficiency. Using this surrogate model, the engineers can obtain the deformation law just only modifying key parameters. Moreover, the surrogate model can be packaged into a mobile app to provide support for rapid decision-making on the project site. [ABSTRACT FROM AUTHOR]

The domain shift is a very common phenomenon in Zero-shot Learning (ZSL), because the data distribution between the source and target domain might not match well in real scenarios. This paper focuses on dealing with the domain drift problem in ZSL based on the semantic knowledge graph propagation. Our method consists of visual feature extraction (VFE) module, semantic feature extraction module (SFE) and feature mapping module (FM). In VFE module, the high-level visual features are extracted under Convolutional Neural Network (CNN) framework, where unseen samples participate in model training and stay away from seen ones under the premise of distinguishability. SFE module relies on the Graph Convolutional Network (GCN) framework and focuses on the transmission of semantic embeddings. A modified message aggregation and transformation strategy is proposed to effectively relieve the information smoothing phenomenon caused by the increase of GCN layers. In the recognition stage, AutoEncoder framework achieves a two-way visual and semantic interaction to further minimize the domain drift phenomenon. Simulation results on ImageNet dataset have demonstrated the performance of the proposed method. [ABSTRACT FROM AUTHOR]

This paper introduces a quantum convolutional neural network model that is implementable on real quantum circuits. Three basic blocks, including the quantum encoding block, the model design block and the parameter tuning block, are designed to formulate the quantum convolutional neural network framework. Simulation results on the MNIST dataset are presented to demonstrate the representation, learning, and optimization capacity of the proposed model. [ABSTRACT FROM AUTHOR]

Purpose: Bitcoin has indeed been universally acknowledged as an investment asset in recent decades, after the boom-and-bust of cryptocurrency values. Because of its extreme volatility, it requires accurate forecasts to build economic decisions. Although prior research has utilized machine learning to improve Bitcoin price prediction accuracy, few have looked into the plausibility of using multiple modeling approaches on datasets containing varying data types and volumetric attributes. Thus, this paper aims to propose a bitcoin price prediction model. Design/methodology/approach: In this research work, a bitcoin price prediction model is introduced by following three major phases: Data collection, feature extraction and price prediction. Initially, the collected Bitcoin time-series data will be preprocessed and the original features will be extracted. To make this work good-fit with a high level of accuracy, we have been extracting the second order technical indicator based features like average true range (ATR), modified-exponential moving average (M-EMA), relative strength index and rate of change and proposed decomposed inter-day difference. Subsequently, these extracted features along with the original features will be subjected to prediction phase, where the prediction of bitcoin price value is attained precisely from the constructed two-level ensemble classifier. The two-level ensemble classifier will be the amalgamation of two fabulous classifiers: optimized convolutional neural network (CNN) and bidirectional long/short-term memory (BiLSTM). To cope up with the volatility characteristics of bitcoin prices, it is planned to fine-tune the weight parameter of CNN by a new hybrid optimization model. The proposed hybrid optimization model referred as black widow updated rain optimization (BWURO) model will be conceptual blended of rain optimization algorithm and black widow optimization algorithm. Findings: The proposed work is compared over the existing models in terms of convergence, MAE, MAPE, MARE, MSE, MSPE, MRSE, Root Mean Square Error (RMSE), RMSPE and RMSRE, respectively. These evaluations have been conducted for both algorithmic performance as well as classifier performance. At LP = 50, the MAE of the proposed work is 0.023372, which is 59.8%, 72.2%, 62.14% and 64.08% better than BWURO + Bi-LSTM, CNN + BWURO, NN + BWURO and SVM + BWURO, respectively. Originality/value: In this research work, a new modified EMA feature is extracted, which makes the bitcoin price prediction more efficient. In this research work, a two-level ensemble classifier is constructed in the price prediction phase by blending the Bi-LSTM and optimized CNN, respectively. To deal with the volatility of bitcoin values, a novel hybrid optimization model is used to fine-tune the weight parameter of CNN. [ABSTRACT FROM AUTHOR]

The distributed estimation has important research significance in unmanned systems. This paper investigates the distributed estimation of unmanned surface vessel (USV) via multi-sensor collaboration and 3D object recognition, in which a Knowledge Graph (KG) is constructed to store and represent the estimation results. Kalman-consensus Filter (KCF) and convolutional neural network (CNN) are used to estimate the optimal states of objects, and recognise multiple classes of objects without designing detectors for each class of objects, respectively. The recognition efficiency is improved by dividing the data into pixel blocks whose value is the number of detection points, and a point cloud dataset in different locations and rotations is also provided. Experiments are proposed to show that our method can help the USV accurately perceive entities in the environment, which validates the effectiveness of the proposed algorithm. [ABSTRACT FROM AUTHOR]

Purpose: The density method is one of the powerful topology optimization methods of magnetic devices. The density method has the advantage that it has a high degree of freedom of shape expression which results in a high-performance design. On the other hand, it has also the drawback that unsuitable shapes for actually manufacturing are likely to be generated, e.g. checkerboards or grayscale. The purpose of this paper is to develop a method that enables topology optimization suitable for fabrication while taking advantage of the density method. Design/methodology/approach: This study proposes a novel topology optimization method that combines convolutional neural network (CNN) as an effective smoothing filter with the density method and apply the method to the shield design with magnetic nonlinearity. Findings: This study demonstrated some numerical examples verifying that the proposed method enables to efficiently obtain a smooth and easy-to-manufacture shield shape with high shielding ability. A network architecture suitable as smoothing filter was also exemplified. Originality/value: In the field of magnetic field analysis, very few studies have verified the usefulness of smoothing by using CNN in the topology optimization of magnetic devices. This paper develops a novel topology optimization method that skillfully combines CNN with the nonlinear magnetic field analysis and also clarifies a suitable network architecture that makes it possible to obtain a target device shape that is easy to manufacture while minimizing the objective function value. [ABSTRACT FROM AUTHOR]

Convolution neural network (CNN) with pruning techniques has shown remarkable prospects in electrocardiogram (ECG) classification. However, efficiently deploying the existing pruned neural network to wearable devices for ECG classification is a great challenge due to the limited hardware resource and randomly distributed sparse weights. To address this issue, an efficient unstructured sparse CNN accelerator is proposed in this paper. A tile-first dataflow with compressed data storage format is presented to skip zero weight multiplications and increase the computing efficiency during inference of small-scale model with large sparsity. The two-level weight index matching structure in the dataflow exploits shifting operation to select valid data pairs and maintain the fully-pipelined calculation process. A configurable processing element (PE) array with 32-bit instruction control is proposed to increase the flexibility of the accelerator. Verified in FPGA and post-synthesis simulations in SMIC 40nm process, the proposed sparse CNN accelerator consumes $3.93~\mu $ J/classification at 2MHz clock frequency and it achieves an averaged ECG classification accuracy of 98.99%. A computing efficiency of 118.75% is realized which is improved by 48% compared to the dense baseline. In brief, the proposed efficient CNN accelerator is especially suitable for wearable ECG classification device. [ABSTRACT FROM AUTHOR]

• Local and global self-attention mechanisms are used for character embedding. • CNN with multi-size filters are used to extract character information for NER. • A cross-attention method that fuses character and word embedding for NER is proposed • A modified Mogrifier LSTM is presented to improve the performance of NER. • Proposed methods integrated with a transformer-based model achieve good performance. Clinical and biomedical concept extraction is critical in medical analysis using clinical and biomedical documents from professional literature, EHRs and PHRs. Named entity recognition (NER) accurately marks essential information in the literature based on the characteristics of the target entity, providing a method for extracting clinical and biomedical concepts. The performance of NER is heavily embedding-dependent, so recent studies have proposed the method of generating word embedding from character-level information, which can strengthen the representation ability for word embedding. In this paper, we present a novel neural network model including an attention mechanism network and a convolutional neural network (CNN) to further improve character-level embedding. First, an attention mechanism is applied simultaneously to the local and global character embedding. Then, a CNN with multi-size filters is used to extract more information from the character level, which can capture more meaningful features from words with various lengths. In addition, a cross-attention method is used to leverage the interaction between word embedding and character embedding to generate the final word representation. Finally, we modified Mogrifier LSTM to make it suitable for NER tasks and integrated it into our model. Experimental results show that our method is effective and that the model performs better than the baseline models. We also apply our methods proposed in this paper to the transformer-based model and obtain a 90.36 F1-score on NCBI-Disease. [ABSTRACT FROM AUTHOR]

Designing a Convolutional Neural Networks (CNN) is a complex task and requires expert knowledge to optimize the performance and network architecture. In this paper, a novel data-driven approach is proposed to determine the architecture of CNN models. The proposed Autonomous Convolutional Neural Networks (AutoCNNThe executable code and original numerical results can be downloaded from (https://tinyurl.com/AutoCNN)) algorithm introduces data driven strategies for addition of new convolutional layers, pruning of redundant filters and training cycle optimization. AutoCNN is evaluated using MNIST, MNIST-rot-back-image, Fashion MNIST and the ADHD200 datasets to measure the performance on small datasets with varied feature distributions. The results indicate that AutoCNN optimizes the CNN network architecture and helps maximise the classification performance. The data-driven network determination approach introduced in this paper was found to not only provides competitive performance similar to existing evolutionary computation based network determination algorithms in literature, but was found to be an effective optimization tool to improve the performance of existing CNN architectures. Further, the AutoCNN was found to highly immune to noise in the dataset and has proven to be effective method to transfer knowledge between related datasets. Therefore, the AutoCNN is a highly versatile CNN architecture determination tool that has a wide range of applications in the field of autonomous driving, medical image analysis, image enhancement, camera based security monitoring and image based fault detection. [ABSTRACT FROM AUTHOR]

This paper presents an advanced offline scheduling scheme to improve the accelerator efficiency, especially for the highly-pruned convolutional neural networks (HP-CNNs). Based on the existing outlier-aware accelerator design, we demonstrate the efficiency drop of HP-CNN processing for the first time, and element-wise channel merging is proposed to make a dense processing sequence even for the highly-pruned model. The dedicated hardware architecture is also presented to process the merged channels with the minimum hardware-level overheads, improving the energy efficiency for handling HP-CNNs by preserving the hardware utilization. We further investigate the optimal size of accumulator and multiplexer, in addition to the number of merged channels, exploiting the attractive energy-performance trade-offs. As a result, unlike the practical HP-CNNs for the on-device solutions, the proposed method enhances the overall efficiency by up to 33% compared to the state-of-the-art schemes. [ABSTRACT FROM AUTHOR]

• A novel physics-informed shallow convolutional neural network is proposed. • The solving of the nonlinear physical operator equation is implemented. • The physical information is reconstructed from some noisy observations. • The effectiveness of current development is illustrated through extensive cases. • The speed acceleration of current development is significantly improved. We introduce a novel methodology for solving nonlinear partial differential equation (PDE) on regular or irregular domains using physics-informed ConvNet, which we call the PICN. The network structure consists of three parts: 1) a convolutional neural network for physical field generation, 2) a pre-trained convolutional layer corresponding to the finite-difference filters to estimate differential fields of the generated physical field, and 3) an interpolation network for loss analysis in irregular geometry domains. From a CNN perspective, the physical field is generated by a deconvolution layer and a convolution layer. Unlike the standard Physics-informed Neural Network (PINN) approach, the convolutions corresponding to the finite-difference filters estimate the spatial gradients forming the physical operator and then construct the PDE residual in a PINN-like loss function. The total loss function involving boundary conditions and the physical constraints in irregular geometry domains can be calculated from an efficient linear interpolation network. The theoretical analysis of PICN convergence is performed on a simplified case for solving a one-dimensional physical field, and several examples of nonlinear PDE of solutions with multifrequency characteristics are executed. The theory and examples confirm the effective learning capability of PICN for the physical field solution with high-frequency components, compared to the standard PINN. A series of numerical cases are performed to validate the current PICN, including the solving (and estimation) of nonlinear physical operator equations and recovering physical information from noisy observations. First, the ability of PICN to solve nonlinear PDE has been verified by executing three nonlinear problems including ODE with sine nonlinearity, PDE involving nonlinear sine-square operators, and Schrödinger equation. The proposed PICN has been assessed by solving some nonlinear PDE on irregular domains such as star-shaped domain, bird-like domain, and starfish domain. Moreover, PICN is applied to identify the thermal diffusivity parameters in an anisotropic heat transfer problem from noisy data, and a denoising display of the temperature field from strong noisy data with standard deviations ranging from 0.1 to 0.4. The numerical results demonstrate the high accuracy approximation and fast convergence performance of PICN. The potential advantage in approximating complex physical field with multi-frequency components indicates that PICN may become an alternative efficient neural network solver in physics-informed machine learning. This paper is adapted from the work originally posted on arXiv.com by the same authors (arXiv:2201.10967, Jan 26, 2022). The data and code accompanying this paper are publicly available at https://github.com/zengzhi2015/PICN. [ABSTRACT FROM AUTHOR]

Currently, the majority of solutions for human interactive action recognition typically rely on machine vision methods, thus leading to limited application scenarios for accurate recognition. In this paper, we proposed a novel two-level multi-head attentional human interaction action recognition model based on inertial measurement units. Considering that individual actions may contribute differently to interaction actions, the overall model of weighting individual actions to identify interaction actions is divided into individual scenes and interaction scenes. A multi-head attention mechanism is added to the model to obtain action weighting information in both scenes. We combine a bi-directional gated recurrent unit (BiGRU) and an improved convolutional neural network constructs to fully capture spatiotemporal features. The experimental results show that the established model can accurately recognize seven interactive actions with an average recognition accuracy of 98.73%, which verifies the excellent performance of the proposed method. • A new inertial sensor-based action recognition model for human interaction. • A dataset of inertial sensor-based human interaction movements was collected. • Human interaction scenes are divided into individual scenes and interaction scenes. • Multi-head attention mechanism captures local and global information. [ABSTRACT FROM AUTHOR]

Graph convolution neural network (GCN) shows strong performance in non-Euclidean structure data. In recent years, many researchers have applied GCN to Euclidean structure data, such as images and languages, and obtained excellent results, which expanded the application range of GCN. Despite a large number of related methods, few studies deal with information systems from the perspective of graphs. In fact, the information system is also a non-Euclidean data structure, and there is serious information loss when using classical measurement methods to construct its data structure. Thus, the promotion of GCN is hindered in information systems. This paper applies GCN to the classification of information systems. Firstly, an intuitionistic fuzzy relation, called positive-negative relation, is established in the information system, and the physical meaning of this relationship is explained. Secondly, based on the positive-negative relation, a positive-negative relation graph convolution neural network model is constructed called PN-GCN. Finally, the effectiveness and robustness of PN-GCN are validated through experiments. [ABSTRACT FROM AUTHOR]

In recent years, Graph Convolutional Networks (GCNs), which extend convolutional neural networks to graph structure, have achieved great success on many graph learning tasks by fusing structure and feature information, such as node classification. However, the graph structure is constructed from real-world data and usually contains noise or redundancy. In addition, this structural information is based on manually defined relations and is not potentially optimal for downstream tasks. In this paper, we utilize the knowledge from node features to enhance the expressive power of GCN models in a plug-and-play fashion. Specifically, we build a node-feature bipartite graph and exploit the bipartite graph convolutional network to model node-feature relations. By aligning results from the original graph structure and node-feature relations, we can make a more accurate prediction for each node in an end-to-end manner. Extensive experiments demonstrate that the proposed model can extract knowledge from two branches and improve the performance of various GCN models on typical graph data sets and 3D point cloud data. [ABSTRACT FROM AUTHOR]

• A novel method is proposed for traffic flow forecasting. • A novel module is used for extracting spatial–temporal similarity. • Dependence and similarity features are modeled to improve forecasting accuracy. Traffic flow forecasting is the cornerstone of the development of intelligent transportation systems. Accurate forecasting is conducive to the control and management of urban traffic. However, it is still a challenge to extract the complex spatial–temporal relationships in the traffic flow. The existing methods based on Graph Convolutional Network (GCN) can effectively extract spatial–temporal dependence, but it is difficult to extract spatial–temporal similarity because of the limitation of the graph structure. In this paper, we propose a spatial–temporal dependence and similarity-aware method called Dependence and Similarity aware Temporal Graph Convolutional Network (DS-TGCN), which combines two modules to extract the complex spatial–temporal relationships and make traffic flow forecasting. One module is our Spatial–Temporal Similarity Feature module (STSF) newly designed for extracting the spatial–temporal similarity directly. Another module is the spatial–temporal convolution module with an attention mechanism, which can extract spatial–temporal dependence dynamically. The experiments on two types of datasets demonstrate that our proposed method outperforms the existing methods in terms of the effectiveness of traffic flow forecasting. [ABSTRACT FROM AUTHOR]

This paper presents a block-based embedded convolutional neural network (CNN) for gesture classification on field-programmable gate array (FPGA) in real time. Gesture recognition is an important tool to spontaneous interact with human machine interface. Many CNN architectures using RGB images have been proposed for gesture classification. RGB based gesture classification may cause incorrect results under insufficient light or similar gestures. In addition, most of the CNN architectures cannot run in real time on edge devices due to their large number of parameters and DRAM data access. In this paper, a block-based CNN using RGB-D data is proposed for gesture classification. Adding depth images to RGB images boots the classification accuracy. A CNN architecture with block-based feature maps is built for embedded FPGA implementations. The total number of parameters of the proposed RGB-D embedded CNN (eCNN) model is only 0.17M and it achieves 99.96% and 99.88% accuracy with 32-bit floating point and 8-bit fixed point implementation for America Sign Language (ASL) data set. The RTL simulation of the proposed eCNN model has the average inference speed of 0.171 milliseconds at frequency of 250MHz for a single pair RGB-D image. Implemented on a FPGA integrated with Microsoft Kinect v2 achieve an inference time in 19.42 ms which achieves high accuracy and real-time performance. [ABSTRACT FROM AUTHOR]

In recent years we have seen a growing number of edge devices adopted by consumers, in their homes (e.g., smart cameras and doorbells), in their cars (e.g., driver assisted systems), and even on their persons (e.g., smart watches and rings). Similar growth is reported in industries including aerospace, agriculture, healthcare, transport, and manufacturing. At the same time that devices are getting smaller, Deep Neural Networks (DNN) that power most forms of artificial intelligence are getting larger, requiring more compute power, memory, and bandwidth. This creates a growing disconnect between advances in artificial intelligence and the ability to develop smart devices at the edge. In this paper, we present a novel approach to running state-of-the-art AI algorithms at the edge. We propose two efficient approximations to standard convolutional neural networks: Binary-Weight-Networks (BWN) and XNOR-Networks. In BWN, the filters are approximated with binary values resulting in 32× memory saving. In XNOR-Networks, both the filters and the input to convolutional layers are binary. XNOR-Networks approximate convolutions using primarily binary operations. This results in 58× faster convolutional operations (in terms of number of the high precision operations) and 32× memory savings. XNORNets offer the possibility of running state-of-the-art networks on CPUs (rather than GPUs) in real-time. Our binary networks are simple, accurate, efficient, and work on challenging visual tasks. We evaluate our approach on the ImageNet classification task. The classification accuracy with a BWN version of AlexNet is the same as the full-precision AlexNet. [ABSTRACT FROM AUTHOR]

Residual block is a very common component in recent state-of-the art CNNs such as EfficientNet/EfficientDet. Shortcut data accounts for nearly 40% of feature-maps access in ResNet152. Most of the previous DNN compilers/accelerators ignore the shortcut data optimization. This paper presents ShortcutFusion, an optimization tool for FPGA-based accelerator with a reuse-aware static memory allocation for shortcut data, to maximize on-chip data reuse given resource constraints. From TensorFlow DNN models, the proposed design generates instruction sets for a group of nodes which uses an optimized data reuse for each residual block. The accelerator design implemented on the Xilinx KCU1500 FPGA card $2.8\times $ faster and $9.9\times $ more power efficient than NVIDIA RTX 2080 Ti for $256\times 256$ input size. Compared to the result from baseline, in which the weights/inputs/outputs are accessed from the off-chip memory exactly once per each layer, ShortcutFusion reduces the DRAM access by 47.8-84.8% for RetinaNet, Yolov3, ResNet152, and EfficientNet. Given a similar buffer size to ShortcutMining, which also “mine” the shortcut data in hardware, the proposed work reduces off-chip access for feature-maps $5.27\times $ while accessing weight from off-chip memory exactly once. [ABSTRACT FROM AUTHOR]

Purpose: This paper aims to present a deep learning–based surrogate model for fast multi-material topology optimization of an interior permanent magnet (IPM) motor. The multi-material topology optimization based on genetic algorithm needs large computational burden because of execution of finite element (FE) analysis for many times. To overcome this difficulty, a convolutional neural network (CNN) is adopted to predict the motor performance from the cross-sectional motor image and reduce the number of FE analysis. Design/methodology/approach: To predict the average torque of an IPM motor, CNN is used as a surrogate model. From the input cross-sectional motor image, CNN infers dq-inductance and magnet flux to compute the average torque. It is shown that the average torque for any current phase angle can be predicted by this approach, which allows the maximization of the average torque by changing the current phase angle. The individuals in the multi-material topology optimization are evaluated by the trained CNN, and the limited individuals with higher potentials are evaluated by finite element method. Findings: It is shown that the proposed method doubles the computing speed of the multi-material topology optimization without loss of search ability. In addition, the optimized motor obtained by the proposed method followed by simplification for manufacturing is shown to have higher average torque than a reference model. Originality/value: This paper proposes a novel method based on deep learning for fast multi-material topology optimization considering the current phase angle. [ABSTRACT FROM AUTHOR]

In-memory computing (IMC) is an effective solution for energy-efficient artificial intelligence applications. Analog IMC amortizes the power consumption of multiple sensing amplifiers with an analog-to-digital converter (ADC) and simultaneously completes the calculation of multi-line data with a high parallelism degree. Based on a universal one-transistor one-magnetic tunnel junction (MTJ) spin transfer torque magnetic RAM (STT-MRAM) cell, this paper demonstrates a novel tunneling magnetoresistance (TMR) ratio magnifying method to realize analog IMC. Previous concerns including low TMR ratio and analog calculation nonlinearity are addressed using device-circuit interaction. The TMR is magnified $7500\times $ using a latch structure in combination with the device. Peripheral circuits are minimally modified to enable in-memory matrix-vector multiplication. A current mirror with a feedback structure is implemented to enhance analog computing linearity and calculation accuracy. The proposed design maximumly supports 1024 2-bit input and 1-bit weight multiply-and-accumulate (MAC) computations simultaneously. The proposal is simulated using the 28-nm CMOS process and MTJ compact model. The integral nonlinearity is reduced by 57.6% compared with the conventional structure. 9.47-25.4 TOPS/W is realized with 2-bit input, 1-bit weight, and 4-bit output convolution neural network (CNN). [ABSTRACT FROM AUTHOR]

Existing deep convolutional neural networks (CNNs) generate massive interlayer feature data during network inference. To maintain real-time processing in embedded systems, large on-chip memory is required to buffer the interlayer feature maps. In this paper, we propose an efficient hardware accelerator with an interlayer feature compression technique to significantly reduce the required on-chip memory size and off-chip memory access bandwidth. The accelerator compresses interlayer feature maps through transforming the stored data into frequency domain using hardware-implemented $8\times 8$ discrete cosine transform (DCT). The high-frequency components are removed after the DCT through quantization. Sparse matrix compression is utilized to further compress the interlayer feature maps. The on-chip memory allocation scheme is designed to support dynamic configuration of the feature map buffer size and scratch pad size according to different network-layer requirements. The hardware accelerator combines compression, decompression, and CNN acceleration into one computing stream, achieving minimal compressing and processing delay. A prototype accelerator is implemented on an FPGA platform and also synthesized in TSMC 28-nm COMS technology. It achieves 403GOPS peak throughput and $1.4\times \sim 3.3\times $ interlayer feature map reduction by adding light hardware area overhead, making it a promising hardware accelerator for intelligent IoT devices. [ABSTRACT FROM AUTHOR]

Most real-world networks are heterogeneous and consist of different types of nodes and edges. Relationships (edges) between nodes of different types carry different semantics, hence relationship prediction in heterogeneous networks aims to predict the existence of a relation between nodes of different types. In this paper, we propose a relation-aware Heterogeneous Graph Convolutional Network architecture for temporal heterogeneous network relationship prediction. In particular, we first present a continuous-time temporal heterogeneous neighbour generation algorithm to capture the continuous-time interactions in the context of a specific node type. The algorithm collects temporal heterogeneous neighbours with K -hops of a node at a recent temporal distance. Moreover, it adds a decaying exponential to ensure that the fewer hops and the closer temporal distances, the more nodes are sampled. Thus, it can capture the evolutionary pattern of interactions in the recent past and collect strongly correlated heterogeneous neighbours for each node. Then the proposed relation-aware Heterogeneous Graph Convolutional Network is used to learn the most relevant relationship to the target relationship to model the heterogeneous graph neural network for predicting target relationships. Our experimental results on three real-world temporal heterogeneous networks indicate significant advances in prediction accuracy and efficiency compared to the state-of-art approaches. [ABSTRACT FROM AUTHOR]

• Logic AND operation in FNNs did not depict the uncertainty of pixels in images well. • A new fuzzy rule is proposed for the uncertainty using the similarity between pixels. • A novel SR method is proposed by combining fuzzy set theory and deep learning. • Some deep learning modules are designed to extract more effective image features. • The proposed method has better performance in SR medical image reconstruction. The clarity of medical images is crucial for doctors to identify and diagnose different diseases. High-resolution images have more detailed information and clearer content than low-resolution images. It is well known that medical images can frequently have some blurred object boundaries, and that traditional deep learning models cannot adequately describe the uncertainty of these blurred boundaries. This paper proposes a new fuzzy metric to characterize the uncertainty of pixels and designs a fuzzy hierarchical fusion attention neural network based on multiscale guided learning. Specifically, a fuzzy neural information-processing block is proposed, which converts an input image into a fuzzy domain using fuzzy membership functions. The uncertainty of the pixels is processed using the proposed fuzzy rules, and then the output of the fuzzy rule layer is fused with the result of the convolution in the neural network. Simultaneously, a multiscale guided-learning dense residual block and pyramidal hierarchical attention module are designed to extract more effective hierarchical image information. Finally, a recurrent memory module with a residual structure is used to process the output features of the hierarchical attention modules. A recursive sub-pixel reconstruction module is used at the tail of the network to reconstruct the images. Compared with existing super-resolution methods using the public COVID-CT dataset, the proposed method demonstrated superior performance in high-resolution medical image reconstruction and reduced the number of parameters and analysis time of the models. [ABSTRACT FROM AUTHOR]

Unsupervised Domain Adaptation (UDA) is a popular machine learning technique to reduce the distribution discrepancy among domains. In previous UDA methods, only convolutional neural networks (CNNs) or vision transformers (ViTs) are used as the backbone. Therefore, these methods are subjected to inherent characteristics of single-CNNs or single-ViTs. In fact, CNNs are more biased towards local textures, while ViTs are inclined to learn the shape information of images. In this paper, we consider that the model combining the above two bias properties is closer to the human visual neural system and achieves more robust performance. A novel mutual distillation method for UDA without any assistant networks is proposed. Mutual distillation between two backbones with complementary properties (i.e. CNNs & ViTs) can promote each other, leading to better domain knowledge transfer. Additionally, traditional domain-mixup approaches can only mix limited cross-domain information through linear interpolation. To encourage more cross-domain information interaction between the two backbones and bridge the domain gap, we propose a patch-mixup method, building a train of mixed intermediate domains composed of augmented patches. Meanwhile, a cross-domain semantic alignment loss is proposed to align the semantic information of these domains. Extensive experiments show that our method achieves state-of-the-art results on several standard UDA benchmarks, such as Office-31, Office-Home, ImageCLEF-DA and VisDA-2017. [ABSTRACT FROM AUTHOR]

• Logic AND operation in FNNs did not depict the uncertainty of pixels in images well. • A new fuzzy rule is proposed for the uncertainty using the similarity between pixels. • A novel SR method is proposed by combining fuzzy set theory and deep learning. • Some deep learning modules are designed to extract more effective image features. • The proposed method has better performance in SR medical image reconstruction. The clarity of medical images is crucial for doctors to identify and diagnose different diseases. High-resolution images have more detailed information and clearer content than low-resolution images. It is well known that medical images can frequently have some blurred object boundaries, and that traditional deep learning models cannot adequately describe the uncertainty of these blurred boundaries. This paper proposes a new fuzzy metric to characterize the uncertainty of pixels and designs a fuzzy hierarchical fusion attention neural network based on multiscale guided learning. Specifically, a fuzzy neural information-processing block is proposed, which converts an input image into a fuzzy domain using fuzzy membership functions. The uncertainty of the pixels is processed using the proposed fuzzy rules, and then the output of the fuzzy rule layer is fused with the result of the convolution in the neural network. Simultaneously, a multiscale guided-learning dense residual block and pyramidal hierarchical attention module are designed to extract more effective hierarchical image information. Finally, a recurrent memory module with a residual structure is used to process the output features of the hierarchical attention modules. A recursive sub-pixel reconstruction module is used at the tail of the network to reconstruct the images. Compared with existing super-resolution methods using the public COVID-CT dataset, the proposed method demonstrated superior performance in high-resolution medical image reconstruction and reduced the number of parameters and analysis time of the models. [ABSTRACT FROM AUTHOR]

Unsupervised Domain Adaptation (UDA) is a popular machine learning technique to reduce the distribution discrepancy among domains. In previous UDA methods, only convolutional neural networks (CNNs) or vision transformers (ViTs) are used as the backbone. Therefore, these methods are subjected to inherent characteristics of single-CNNs or single-ViTs. In fact, CNNs are more biased towards local textures, while ViTs are inclined to learn the shape information of images. In this paper, we consider that the model combining the above two bias properties is closer to the human visual neural system and achieves more robust performance. A novel mutual distillation method for UDA without any assistant networks is proposed. Mutual distillation between two backbones with complementary properties (i.e. CNNs & ViTs) can promote each other, leading to better domain knowledge transfer. Additionally, traditional domain-mixup approaches can only mix limited cross-domain information through linear interpolation. To encourage more cross-domain information interaction between the two backbones and bridge the domain gap, we propose a patch-mixup method, building a train of mixed intermediate domains composed of augmented patches. Meanwhile, a cross-domain semantic alignment loss is proposed to align the semantic information of these domains. Extensive experiments show that our method achieves state-of-the-art results on several standard UDA benchmarks, such as Office-31, Office-Home, ImageCLEF-DA and VisDA-2017. [ABSTRACT FROM AUTHOR]

With the growing applications of Graph Convolutional Networks (GCN), there is also an increasing demand for its efficient hardware acceleration. Compared with CNN tasks, GCN tasks have new challenges such as randomness, sparsity, and nonuniformity, which will lead to poor performance of previous AI accelerators. In this paper, we propose DyGA, a hardware-efficient GCN accelerator, which is featured by strategies of graph partitioning, customized storage policy, traffic-aware dynamic scheduling, and out-of-order execution. Synthesized and evaluated under TSMC 28-nm, the accelerator achieves an average throughput of over 95% of its peak performance with full utilization of hardware on representative graph data sets. Having a high area-efficiency with 0.217 GOPS/K-logic-gates and 8.06 GOPS/KB-PE-buffer, and thus an energy-efficiency of 384GOPS/W, the proposed accelerator outperforms previous state-of-the-art works in the sparse data processing. [ABSTRACT FROM AUTHOR]

It remains challenging for a Convolutional Neural Network (CNN) accelerator to maintain high hardware utilization and low processing latency with restricted on-chip memory. This paper presents an In-Execution Configuration Accelerator (IECA) that realizes an efficient control scheme, exploring architectural data reuse, unified in-execution controlling, and pipelined latency hiding to minimize configuration overhead out of the computation scope. The proposed IECA achieves row-wise convolution with tiny distributed buffers and reduces the size of total on-chip memory by removing 40% of redundant memory storage with shared delay chains. By exploiting a reconfigurable Sequence Mapping Table (SMT) and Finite State Machine (FSM) control, the chip realizes cycle-accurate Processing Element (PE) control, automatic loop tiling and latency hiding without extra time slots for pre-configuration. Evaluated on AlexNet and VGG-16, the IECA retains over 97.3% PE utilization and over 95.6% memory access time hiding on average. The chip is designed and fabricated in a UMC 55-nm process running at a frequency of 250 MHz and achieves an area efficiency of 30.55 GOPS/mm2 and 0.244 GOPS/KGE (kilo-gate-equivalent), which makes an over $2.0\times $ and $2.1\times $ improvement, respectively, compared with that of previous related works. Implementation of the IEC control scheme uses only a 0.55% area of the 2.75 mm2 core. [ABSTRACT FROM AUTHOR]

The early and accurate detection of COVID-19 is vital nowadays to avoid the vast and rapid spread of this virus and ease lockdown restrictions. As a result, researchers developed methods to diagnose COVID-19. However, these methods have several limitations. Therefore, presenting new methods is essential to improve the diagnosis of COVID-19. Recently, investigation of the electrocardiogram (ECG) signals becoming an easy way to detect COVID-19 since the ECG process is non-invasive and easy to use. Therefore, we proposed in this paper a novel end-to-end deep learning model (ECG-COVID) based on ECG for COVID-19 detection. We employed several deep Convolutional Neural Networks (CNNs) on a dataset of 1109 ECG images, which is built for screening the perception of COVID-19 and cardiac patients. After that, we selected the most efficient model as our model for evaluation. The proposed model is end-to-end where the input ECG images are fed directly to the model for the final decision without using any additional stages. The proposed method achieved an average accuracy of 98.81%, Precision of 98.8%, Sensitivity of 98.8% and, F1-score of 98.81% for COVID-19 detection. As cases of corona continue to rise and hospitalizations continue again, hospitals may find our study helpful when dealing with these patients who did not get significantly worse. [ABSTRACT FROM AUTHOR]

Accurate and timely crop yield prediction is difficult to achieve due to the nonlinear and dynamic spatial–temporal correlations included during the crop growth process. The latest approaches utilize deep learning techniques to implicitly capture sequential dependencies but fail to consider prior knowledge. In this paper, a novel Knowledge-guided Spatial–Temporal Attention Graph Network (KSTAGE), which incorporates prior knowledge to model dynamic temporal relationships and geospatial topology structures, is proposed. The proposed KSTAGE first utilizes a 3D convolutional neural network (CNN) to embed the initial spectral feature. Following its development, a new Knowledge-guided Temporal Multi-head Attention Algorithm (KTMA), the core of KSTAGE, is applied to automatically generate temporal attention weights from a self-attention mechanism under the guidance of prior knowledge. Furthermore, a new strategy for aligning self-attention scores with the prior distribution is developed. Finally, a location-aware Spatial Attention Graph Network utilizing geospatial knowledge is proposed to aggregate the spatial neighborhood features for the final yield prediction. The experimental results obtained on county-level yield prediction tasks in China and the contiguous United States (CONUS) demonstrate that the proposed KSTAGE achieves obvious improvements over the baselines. [ABSTRACT FROM AUTHOR]

Depthwise separable convolution (DSC) has become one of the essential structures for lightweight convolutional neural networks. Nevertheless, its hardware architecture has not received much attention. Several previous hardware designs incur either high off-chip memory traffic or large on-chip memory usage, and hence have deficiency in terms of hardware efficiency as well as performance. This paper proposes two efficient dynamic design techniques, i.e. adaptive row-based dataflow scheduling and adaptive computation mapping, to achieve a much better trade-off between hardware efficiency and performance for DSC-based lightweight CNN accelerator. The effectiveness and efficiency of the proposed dynamic design techniques have been extensively evaluated using six DSC-based lightweight CNNs. Compared with the reference architectures, the simulation results show the proposed architectural techniques can at least reduce on-chip buffer size by 50.4% and improve the performance of convolution calculation by 1.18 × while maintaining the minimum off-chip memory traffic. MobileNetV2 is implemented on Zynq UltraScale+ ZCU102 SoC FPGA, and the results show the proposed accelerator can achieve 381.7 frames per second (fps), which is 1.43 × of the reference design, and it can save about 36.3% on-chip buffer size compared with the reference design, while maintaining the same off-chip memory traffic. [ABSTRACT FROM AUTHOR]

Convolutional neural networks (CNN) have been widely used in various tasks, largely due to their ability to efficiently extract n-gram features for text analysis and document representation. In this paper, we intend to insight the CNN model regarding its capability on text analysis. Vanilla CNNs do have weaknesses when it comes to the representation and feature extraction. Duplicate filters are inevitable with vanilla CNNs, which reduces the discriminative power of the representations. In addition, the current pooling operations either limit the CNN to the local optimum (i.e., max pooling) or they do not consider the importance of all features (i.e., mean pooling). In this paper, we propose two modules for vanilla CNNs to overcome these shortcomings. The first equips the CNN with discriminative filters (distinct filters with maximised divergence) and the second provides the ability to comprehensively extract all salient features. Specifically, our model increases the discriminative power of the model by maximizing the distance between different filters, and a novel global pooling mechanism for feature extraction. Validation tests against state-of-the-art baselines on five benchmark classification datasets achieve the competitive performance of our proposed model. Furthermore, visualization on upgrade filters and pooling features verify our hypothesis that the proposed model can receive discriminative filters and salient features. [ABSTRACT FROM AUTHOR]

In-memory computing provides unprecedented power and area efficiency for the execution of convolutional neural networks by using memory bitcells to perform dot-product (DP) operations in the analog domain. Yet, these operators suffer from analog non-idealities (ANIs) that degrade the inference accuracy. This paper proposes design guidelines inferred from a holistic simulation-based analysis of the impact of ANIs on the accuracy-efficiency trade-off that affects current-domain DP operators based on conventional 6T-SRAM bitcell arrays. We define a custom SNR metric aware of the DP operand distribution to quantify decision errors associated with various ANIs, over ranges of input/output resolution and hardware design parameters. We find out that non-linearity and local mismatch are the dominant ANIs limiting the design space, while IR drops turn out to be critical only when targeting high parallelism. We then quantify the accuracy-efficiency trade-off related to these dominant ANIs across the design space and propose optimal design choices. We notably identify that using larger operators can either improve or worsen the SNR depending on the target output resolution. Furthermore, we show that hardware calibration techniques which mitigate mismatch help to recover a fraction of the lost SNR, with greater effectiveness when scaling down the supply voltage. [ABSTRACT FROM AUTHOR]

Hierarchical classification has become one of the most popular research topics because the scale of data has increased exponentially. Top-down hierarchical classification is an effective classification method using the hierarchical class structure as important side information. However, inter-level error propagation is a crucial problem in the top-down strategy. In this paper, we propose a hierarchical classifier based on deep branch convolutional neural networks, which achieves hierarchical classification based on coarse- to fine-grained knowledge transfer. Specifically, we use a deep convolutional neural network to extract image rough and detailed features from shallow and deep networks. We then embed the branch network at different depths of the convolutional network for hierarchical classification. We splice features from the previous branch network to the current branch. It alleviates inter-level error propagation by knowledge transfer from coarse- to fine-grained. Experimental results on four datasets show that the proposed method outperforms nine popular classifiers for hierarchical classification. Especially on the CIFAR10 dataset, our method is about 5% better than the second-best method. [ABSTRACT FROM AUTHOR]

Fusion functions and their most important subclass, aggregation functions, have been successfully applied in fuzzy modeling. However, there are practical problems, such as classification via Convolutional Neural Networks (CNNs), where the data to be aggregated are not modeling membership degrees in the unit interval. In this scenario, systems could benefit from the application of operators defined in domains different from [ 0 , 1 ] , although, presenting similar behavior of some aggregation functions whose subclasses are currently defined only in the fuzzy context (e.g., overlap functions and t-norms). So, the main objective of this paper is to present a general framework to characterize classes of fusion functions with floating domains, called (a , b) -fusion functions, defined on any closed real interval [ a , b ] , based on classes of core fusion functions defined on [ 0 , 1 ]. The fundamental aspect of this framework is that the properties of a core fusion function are preserved in the context of the analogous (a , b) -fusion function. Construction methods are presented, and some properties are studied. We also introduce a framework to define fusion functions in which the inputs come from an interval [ a , b ] but the output is mapped on a possibly different interval [ c , d ]. Finally, we present an illustrative example in image classification via CNNs. [ABSTRACT FROM AUTHOR]

Recent works on residual network and its variants help us to advance our understanding of deep neural networks as differential equations. Specifically, when taking convolution operator as the combinations of differential operators, deep convolutional neural network (DCNN) can be regarded as partial differential equation (PDE). Learn such PDEs have become popular in image restoration especially for image denoising. While current methods for learning the denoising PDEs face two critical issues. On the one hand, the learned PDEs are not guaranteed to behave as progressive denoising as traditional denoising PDEs. On the other hand, the low level feature may diminish with time increases in PDE. To resolve these issues, this paper designs to learn a specialized PDE network (PDNet) which depicts the information propagation in the network as a reaction-diffusion–advection process. Specifically, a reaction-term is introduced to prevent the low-level features from diminishing in the PDE network, and a decreasing residual gain strategy with stochastic supervision is adopted to facilitate the progressive denoising property. The progressive denoising property not only reveals the semantic of the network path but also regularizes the network training. Compared to vanilla DCNNs, the PDNet has clear physical meaning. Meanwhile PDNet simplifies the ResNet in that batch normalization is removed from the residual blocks and there is only one nonlinear unit in a residual block (2 for ResNet). Extensive experimental results on benchmark datasets verified that PDNet is not only efficient but also effective. [ABSTRACT FROM AUTHOR]

Fusion functions and their most important subclass, aggregation functions, have been successfully applied in fuzzy modeling. However, there are practical problems, such as classification via Convolutional Neural Networks (CNNs), where the data to be aggregated are not modeling membership degrees in the unit interval. In this scenario, systems could benefit from the application of operators defined in domains different from [ 0 , 1 ] , although, presenting similar behavior of some aggregation functions whose subclasses are currently defined only in the fuzzy context (e.g., overlap functions and t-norms). So, the main objective of this paper is to present a general framework to characterize classes of fusion functions with floating domains, called (a , b) -fusion functions, defined on any closed real interval [ a , b ] , based on classes of core fusion functions defined on [ 0 , 1 ]. The fundamental aspect of this framework is that the properties of a core fusion function are preserved in the context of the analogous (a , b) -fusion function. Construction methods are presented, and some properties are studied. We also introduce a framework to define fusion functions in which the inputs come from an interval [ a , b ] but the output is mapped on a possibly different interval [ c , d ]. Finally, we present an illustrative example in image classification via CNNs. [ABSTRACT FROM AUTHOR]

Recent works on residual network and its variants help us to advance our understanding of deep neural networks as differential equations. Specifically, when taking convolution operator as the combinations of differential operators, deep convolutional neural network (DCNN) can be regarded as partial differential equation (PDE). Learn such PDEs have become popular in image restoration especially for image denoising. While current methods for learning the denoising PDEs face two critical issues. On the one hand, the learned PDEs are not guaranteed to behave as progressive denoising as traditional denoising PDEs. On the other hand, the low level feature may diminish with time increases in PDE. To resolve these issues, this paper designs to learn a specialized PDE network (PDNet) which depicts the information propagation in the network as a reaction-diffusion–advection process. Specifically, a reaction-term is introduced to prevent the low-level features from diminishing in the PDE network, and a decreasing residual gain strategy with stochastic supervision is adopted to facilitate the progressive denoising property. The progressive denoising property not only reveals the semantic of the network path but also regularizes the network training. Compared to vanilla DCNNs, the PDNet has clear physical meaning. Meanwhile PDNet simplifies the ResNet in that batch normalization is removed from the residual blocks and there is only one nonlinear unit in a residual block (2 for ResNet). Extensive experimental results on benchmark datasets verified that PDNet is not only efficient but also effective. [ABSTRACT FROM AUTHOR]

• Novel application of local optima network analysis to neural architecture landscapes. • Findings on a reduced search space were found to still hold on a larger search space. • Neural architecture search landscapes have relatively few local optima. The global structure of the hyperparameter spaces of neural networks is not well understood and it is therefore not clear which hyperparameter search algorithm will be most effective. In this paper we analyze the landscapes of convolutional neural network architecture search spaces to provide insight into appropriate search algorithms for these spaces. Using a classical fitness landscape analysis approach (fitness distance correlation) and a more recent tool (local optima networks) we study the global structure of these spaces. Our analysis on six image classification datasets reveals that the landscapes are multi-modal, but with relatively few local optima from which it is not hard to escape with a simple perturbation operator. This led us to explore the performance of iterated local search, which we found to more effectively search the training landscapes than three evolutionary algorithm variants. Evolutionary algorithms, however, outperformed iterated local search in terms of generalization on problems with larger discrepancies between the training and testing landscapes. [ABSTRACT FROM AUTHOR]

Urban traffic congestion is not only an important cause of traffic accidents, but also a major hinder to urban development. By learning the historical traffic flow data, we can forecast the traffic flow of some regions in the future, which is of great significance to urban road planning, traffic management, traffic control and many more. However, due to the complex topology of traffic network and the diversity of influencing factors to traffic flow, the traffic modes are usually complicated and volatile, which makes traffic flow prediction very difficult. In this paper, we propose a new graph convolution neural network, namely Multi-mode Dynamic Residual Graph Convolution Network (MDRGCN), to capture the dynamic impact of different factors on traffic flow in a road network simultaneously. Firstly, we design a multi-mode dynamic graph convolution module (MDGCN), which is employed to capture the impact of different traffic modes by learning two types of relationship matrices. Then, we design a multi-mode dynamic graph convolution gated recurrent unit (MDGRU) to realize the combination of spatial and temporal dependences. Finally, we use a dynamic residual module (DRM) to integrate the orginal traffic data and the spatio-temporal features extracted by the MDGRU module to forecast the future traffic flow. Experimental reulsts conducted on the NYCTaxi and NYCBike datasets validate that the MDRGCN model performs better than the other eight baselines. [ABSTRACT FROM AUTHOR]

In this paper, we propose a content-oriented no-reference (NR) perceptual video quality assessment (VQA) method for computer graphics (CG) animation videos. First, we extract features in terms of spatiotemporal information and its visual perception from the videos as inputs of our proposed artificial neural network-based VQA model. Second, to facilitate the video quality evaluation, we apply a convolutional neural network (CNN) in the VQA model to generate weight factors for the input features adaptively according to the different types of CG content in videos. Third, we build a subjective CG video quality database for validation of VQA metrics. Experiments demonstrated that our method achieved superior performance in terms of evaluating the quality of CG animation videos. Both the code and proposed database are publicly available at https://github.com/WeizhiXian/CGVQA. The corresponding newly established database is available at https:// pan.baidu.com/s/1_P2ZNrLzJwZfG6xa6tKnDQ (password: cgvq). [ABSTRACT FROM AUTHOR]

A brain tumor is one of the fatal cancer types which causes abnormal growth of brain cells. Earlier diagnosis of a brain tumor can play a vital role in its treatment; however, manual segmentation of the brain tumors from MRI images is a laborious and time-consuming task. Therefore, several automatic segmentation techniques have been proposed, but high inter-and intra-tumor variations in shape and texture make accurate segmentation of enhanced and core tumor regions challenging. The demand for a highly accurate and robust segmentation method persists; therefore, in this paper, we propose a novel fully automatic technique for brain tumor regions segmentation by using multiscale residual attention-UNet (MRA-UNet). MRA-UNet uses three consecutive slices as input to preserve the sequential information and employs multiscale learning in a cascade fashion, enabling it to exploit the adaptive region of interest scheme to segment enhanced and core tumor regions accurately. The study also investigates the application of postprocessing techniques (i.e., conditional random field and test time augmentation), which further helps to improve the overall performance. The proposed method has been rigorously evaluated on the most extensive publicly available datasets, i.e., BraTS2017, BraTS2019, and BraTS2020. Our approach achieved state-of-the-art results on the BraTS2020 dataset with the average dice score of 90.18%, 87.22%, and 86.74% for the whole tumor, tumor core, and enhanced tumor regions, respectively. A significant improvement has been demonstrated for the core and enhancing tumor region segmentation showing the proposed method's evident effectiveness. [ABSTRACT FROM AUTHOR]

Long-tailed classifications make it very challenging to deal with class-imbalanced problems using deep convolutional neural networks (CNNs). Existing solutions based on re-balancing methods perform well and use single-task CNNs to train each fine-grained class independently. However, classification tasks are multiplex and involve a coarse-to-fine hierarchical relation. In this paper, we propose a coarse-to-fine knowledge transfer based multi-task CNN, which utilizes the coarse-to-fine structure to promote long-tailed learning. First, we construct a tail hierarchical structure in a coarse-to-fine way to pay greater attention to tail classes than head classes. Second, a multi-task CNN is adopted to simultaneously train coarse- and fine-grained tasks to extract a more generalized knowledge representation than the single-task CNN. Third, we design a coarse-to-fine knowledge transfer strategy to adaptively adjust the task weights to improve fine-grained performance. Extensive experiments on benchmark datasets show that our model achieves better gains than the re-balancing methods. In particular, the proposed model is 3.25% more accurate than the second-best method on the long-tailed tieredImageNet dataset. [ABSTRACT FROM AUTHOR]

• A deep metric learning method is presented for facial expression recognition. • Chi-squared distance is modified for deep metric learning in histogram space. • A new fully histogram based Convolutional Neural Network is proposed. • A triplet loss based on modified chi-squared distance is proposed to learn the CNN. Facial expression recognition is a challenging problem in machine learning. There is much research has been conducted in this field; however, the accuracy of facial expression recognition, especially in uncontrolled conditions, needs much improvement. In this paper, a deep histogram metric learning in a Convolutional Neural Network (CNN) is presented for facial expression recognition. The proposed CNN utilizes a histogram calculation layer to provide statistical description of feature maps at the output of the convolutional layers. To train the proposed CNN in histogram space, a learnable matrix (equivalent to the fully connected layer) is introduced in chi-squared distance equation. Then, the modified equation is used in the loss function. The recognition rates of the proposed CNN for seven-class facial expression recognition on four well-known databases including CK+, MMI, SFEW, and RAF-DB are 98.47%, 83.41%, 61.01%, and 89.28%, respectively. The results show superiority of the proposed CNN compared to the state-of-the-art methods. [ABSTRACT FROM AUTHOR]

• Point cloud and Mesh models typically use cosine similarity without normalization. • Softmax loss devise a lower-bound and cause a convergence problem. • The softmax loss function is reformulated to let the network converge properly. • Reformulation of softmax loss shows performance improvements over previous methods. Feature normalization has been a crucial step in convolutional neural networks (CNNs) in the past few years. Discriminative feature abstraction is indispensable for boosting the overall performance of learning models. For 3D data, in both point cloud and mesh models, the inner product (cosine similarity) is frequently applied for similarity estimation without normalization. In this paper, we first demonstrate that softmax loss will produce a lower bound and cause a convergence problem if we directly perform feature comparison by cosine similarity. Second, we propose a revised formulation of the softmax loss function that uses an optimized scaled cosine similarity with normalization in the training phase rather than the inner product of two vectors, allowing the network to converge correctly. Experiments show significant improvements over previous methods on several datasets. The SHREC dataset with split16 and split10 achieved 99.16% and 97.66% accuracy, respectively, and the Cube dataset achieved 96.96% accuracy, surpassing previous methods in a mesh classification task. [ABSTRACT FROM AUTHOR]