Single Network Model Research Articles

A Deep Learning Method for Denoising Based on a Fast and Flexible Convolutional Neural Network

Seismic data denoising has always been an indispensable step in the seismic exploration workflow. The quality of the results directly affects the results of subsequent inversion and migration imaging. In this article, we proposed a fast and flexible convolutional neural network (FFCNN) based on DnCNN. In contrast to the existing DnCNN and other artificial intelligence (AI)-based denoisers, FFCNN enjoys several desirable properties: 1) downsampling and upscaling operations, which can sensibly reduce runtimes and memory requirements while maintaining the denoising performance, and 2) we introduced noised level maps, which can make that a single convolutional neural network (CNN) model is expected to inherit the flexibility of handling noise models with different parameters, even spatially variant noises by noting <inline-formula xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"> <tex-math notation="LaTeX">$M$ </tex-math></inline-formula> can be nonuniform. Another benefit of increasing the noise-level map is that we can preserve more useful seismic data information by controlling the tradeoff of noise removal effect and seismic data detail preservation. For real seismic data denoised work, the main work and advantages of this article are concentrated on the following two aspects: 1) we introduced a data augmentation strategy to overcome the lack of well-labeled samples and 2) transfer learning has been introduced to the training processing, which used the well-trained synthetic seismic data denoising network as a pretrained model. In this way, we can greatly accelerate and optimize the learning efficiency of the training network. Ultimately, we can greatly improve the computational efficiency and denoising performance based on this intelligent denoised network FFCNN. Finally, numerical experiments prove the effectiveness of our method in synthetic and real seismic data

Reconstruction and Localization of Tumors in Breast Optical Imaging via Convolution Neural Network Based on Batch Normalization Layers

The Near-Infrared (NIR) Diffuse Optical Tomography (DOT) aims to reconstruct optical-property images of tissue and identify and localize breast tumors. This study has developed an efficient and fast DOT reconstruction method based on Deep Learning (DL) algorithm. The DL has already been applied in DOT application with limitations such as a limited dataset and the reconstruction of absorption coefficients only. We solve the problem of a limited dataset by generating a large dataset with multiple phantoms and inclusions at various positions. Moreover, a single Deep Neural Network (DNN) model has been designed to reconstruct absorption coefficients and scattering coefficients. For evaluation of the proposed DNN models, the phantom experimental dataset has been used where the results of the proposed DNN models outperform the results of the Tikhonov Regularization (TR) method and other Artificial Neural Networks (ANN). Moreover, it is shown that the DNN model with batch normalization layer results in improved spatial resolution, based on Contrast-and-Size Detail (CSD) analysis, as compared to DNN models without batch normalization layers.

DR-IIXRN : Detection Algorithm of Diabetic Retinopathy Based on Deep Ensemble Learning and Attention Mechanism.

Diabetic retinopathy (DR) is one of the common chronic complications of diabetes and the most common blinding eye disease. If not treated in time, it might lead to visual impairment and even blindness in severe cases. Therefore, this article proposes an algorithm for detecting diabetic retinopathy based on deep ensemble learning and attention mechanism. First, image samples were preprocessed and enhanced to obtain high quality image data. Second, in order to improve the adaptability and accuracy of the detection algorithm, we constructed a holistic detection model DR-IIXRN, which consists of Inception V3, InceptionResNet V2, Xception, ResNeXt101, and NASNetLarge. For each base classifier, we modified the network model using transfer learning, fine-tuning, and attention mechanisms to improve its ability to detect DR. Finally, a weighted voting algorithm was used to determine which category (normal, mild, moderate, severe, or proliferative DR) the images belonged to. We also tuned the trained network model on the hospital data, and the real test samples in the hospital also confirmed the advantages of the algorithm in the detection of the diabetic retina. Experiments show that compared with the traditional single network model detection algorithm, the auc, accuracy, and recall rate of the proposed method are improved to 95, 92, and 92%, respectively, which proves the adaptability and correctness of the proposed method.

An optimal representation of Random Maximum kSatisfiability on a Hopfield Neural Network for High order logic(k 3)

This paper proposed a new logical rule by incorporating Random maximum kSatsifiability (RANMAXkSAT) in the Hopfield neural network (HNN) as a single network model (HNN-RANMAXkSAT). The purpose is to combine the optimization capacity of the Hopfield neural network (HNN) for optimal representation to random maximum kSatsifiability (MAXRANkSAT). The energy function of a Hopfield neural network has been considered as a programming language for dynamics minimization mechanism. Several optimization and search problems associated with machine learning (ML), decision Science (DS) and artificial intelligence (AI) have been expressed on the Hopfield neural network(HNN) optimally by modelling the problem into variables to minimize the objective function that corresponds to Lyapunov energy function. The computer simulation has been developed based on RANMAXkSAT to explore the feasibility of a Hopfield neural network as a neuro-symbolic integration model in carrying out RANMAXkSAT logic programming optimally. The proposed model has been compared with the existing models published in the literature in term of Global minimum ratio (zM), Fitness energy landscapes (FEL), Root Means square error (RMSE), Mean absolute errors and computation time (CPU). Hence, based on the experimental simulation results, it revealed that the RANMAXkSAT can optimally and effectively represented in the Hopfield neural network (HNN) with 85.1 % classification accuracy.

Stock Price Movement Prediction Using Sentiment Analysis and CandleStick Chart Representation.

Determining the price movement of stocks is a challenging problem to solve because of factors such as industry performance, economic variables, investor sentiment, company news, company performance, and social media sentiment. People can predict the price movement of stocks by applying machine learning algorithms on information contained in historical data, stock candlestick-chart data, and social-media data. However, it is hard to predict stock movement based on a single classifier. In this study, we proposed a multichannel collaborative network by incorporating candlestick-chart and social-media data for stock trend predictions. We first extracted the social media sentiment features using the Natural Language Toolkit and sentiment analysis data from Twitter. We then transformed the stock’s historical time series data into a candlestick chart to elucidate patterns in the stock’s movement. Finally, we integrated the stock’s sentiment features and its candlestick chart to predict the stock price movement over 4-, 6-, 8-, and 10-day time periods. Our collaborative network consisted of two branches: the first branch contained a one-dimensional convolutional neural network (CNN) performing sentiment classification. The second branch included a two-dimensional (2D) CNN performing image classifications based on 2D candlestick chart data. We evaluated our model for five high-demand stocks (Apple, Tesla, IBM, Amazon, and Google) and determined that our collaborative network achieved promising results and compared favorably against single-network models using either sentiment data or candlestick charts alone. The proposed method obtained the most favorable performance with 75.38% accuracy for Apple stock. We also found that the stock price prediction achieved more favorable performance over longer periods of time compared with shorter periods of time.

Crackle Detection In Lung Sounds Using Transfer Learning And Multi-Input Convolutional Neural Networks.

Large annotated lung sound databases are publicly available and might be used to train algorithms for diagnosis systems. However, it might be a challenge to develop a well-performing algorithm for small non-public data, which have only a few subjects and show differences in recording devices and setup. In this paper, we use transfer learning to tackle the mismatch of the recording setup. This allows us to transfer knowledge from one dataset to another dataset for crackle detection in lung sounds. In particular, a single input convolutional neural network (CNN) model is pre-trained on a source domain using ICBHI 2017, the largest publicly available database of lung sounds. We use log-mel spectrogram features of respiratory cycles of lung sounds. The pre-trained network is used to build a multi-input CNN model, which shares the same network architecture for respiratory cycles and their corresponding respiratory phases. The multi-input model is then fine-tuned on the target domain of our self-collected lung sound database for classifying crackles and normal lung sounds. Our experimental results show significant performance improvements of 9.84% (absolute) in F-score on the target domain using the multi-input CNN model and transfer learning for crackle detection.Clinical relevance- Crackle detection in lung sounds, multi-input convolutional neural networks, transfer learning.

Network models of posttraumatic stress disorder: A meta-analysis.

Posttraumatic stress disorder (PTSD) researchers have increasingly used psychological network models to investigate PTSD symptom interactions, as well as to identify central driver symptoms. It is unclear, however, how generalizable such results are. We have developed a meta-analytic framework for aggregating network studies while taking between-study heterogeneity into account and applied this framework in the first-ever meta-analytic study of PTSD symptom networks. We analyzed the correlational structures of 52 different samples with a total sample size of n = 29,561 and estimated a single pooled network model underlying the data sets, investigated the scope of between-study heterogeneity, and assessed the performance of network models estimated from single studies. Our main findings are that: (a) We identified large between-study heterogeneity, indicating that it should be expected for networks of single studies to not perfectly align with one-another, and meta-analytic approaches are vital for the study of PTSD networks. (b) While several clear symptom-links, interpretable clusters, and significant differences between strength of edges and centrality of nodes can be identified in the network, no single or small set of nodes that clearly played a more central role than other nodes could be pinpointed, except for the symptom "amnesia" that was clearly the least central symptom. (c) Despite large between-study heterogeneity, we found that network models estimated from single samples can lead to similar network structures as the pooled network model. We discuss the implications of these findings for both the PTSD literature as well as methodological literature on network psychometrics. (PsycInfo Database Record (c) 2021 APA, all rights reserved).

Conditional Generative Adversarial Networks for low-dose CT image denoising aiming at preservation of critical image content.

X-ray Computed Tomography (CT) is an imaging modality where patients are exposed to potentially harmful ionizing radiation. To limit patient risk, reduced-dose protocols are desirable, which inherently lead to an increased noise level in the reconstructed CT scans. Consequently, noise reduction algorithms are indispensable in the reconstruction processing chain. In this paper, we propose to leverage a conditional Generative Adversarial Networks (cGAN) model, to translate CT images from low-to-routine dose. However, when aiming to produce realistic images, such generative models may alter critical image content. Therefore, we propose to employ a frequency-based separation of the input prior to applying the cGAN model, in order to limit the cGAN to high-frequency bands, while leaving low-frequency bands untouched. The results of the proposed method are compared to a state-of-the-art model within the cGAN model as well as in a single-network setting. The proposed method generates visually superior results compared to the single-network model and the cGAN model in terms of quality of texture and preservation of fine structural details. It also appeared that the PSNR, SSIM and TV metrics are less important than a careful visual evaluation of the results. The obtained results demonstrate the relevance of defining and separating the input image into desired and undesired content, rather than blindly denoising entire images. This study shows promising results for further investigation of generative models towards finding a reliable deep learning-based noise reduction algorithm for low-dose CT acquisition.

Face mask detection algorithm for multi task recognition in complex environment

The right wearing masks can effectively prevent the spread of viruses. There are usually complex distractions in densely populated environments, which will affect detection tasks of face masks. However, the current detecting algorithm based on Faster R-CNN can not meet mask wearing detection of small targets in complex environments, so this paper puts forward a mask wearing detection algorithm based on the improved Faster R-CNN. This algorithm changes the traditional single RPN network model into one multi-task enhanced RPN model, which can improve the detection and recognition accuracy, the improved Soft-NMS algorithm is used to delete the redundant candidate boxes of regional candidate network output. At the same time, CIoU loss function with better performance, which fully considers the center point distance between the target and the detection frame and improves the accuracy of detection, replaced the original IoU loss function. 3000 sample images are used to do training and experimental verification of the model according to 1:3ratio of positive and negative samples. Experimental results show that compared with the Faster R-CNN algorithm, the algorithm proposed in this paper improves the face detection accuracy and face mask wearing detection by 7.15% and 15.99%respectively. Besides, the average detection accuracy improved by 11.57% and the FPS by 4.6.

Deep learning-based design of broadband GHz complex and random metasurfaces

We are interested in exploring the limit in using deep learning (DL) to study the electromagnetic (EM) response for complex and random metasurfaces, without any specific applications in mind. For simplicity, we focus on a simple pure reflection problem of a broadband EM plane wave incident normally on such complex metasurfaces in the frequency regime of 2–12 GHz. In doing so, we create a DL-based framework called the metasurface design deep convolutional neural network (MSDCNN) for both forward and inverse designs of three different classes of complex metasurfaces: (a) arbitrary connecting polygons, (b) basic pattern combination, and (c) fully random binary patterns. The performance of each metasurface is evaluated and cross-benchmarked. Dependent on the type of complex metasurfaces, sample size, and DL algorithms used, the MSDCNN is able to provide good agreement and can be a faster design tool for complex metasurfaces than the traditional full-wave EM simulation methods. However, no single universal deep convolutional neural network model can work well for all metasurface classes based on detailed statistical analysis (such as mean, variance, kurtosis, and mean-squared error). Our findings report important information on the advantages and limitations of current DL models in designing these ultimately complex metasurfaces.

Simulation of multi-species flow and heat transfer using physics-informed neural networks

In the present work, single- and segregated-network physics-informed neural network (PINN) architectures are applied to predict momentum, species, and temperature distributions of a dry air humidification problem in a simple two-dimensional (2D) rectangular domain. The created PINN models account for variable fluid properties, species- and heat-diffusion, and convection. Both the mentioned PINN architectures were trained using different hyperparameter settings, such as network width and depth, to find the best-performing configuration. It is shown that the segregated-network PINN approach results in on-average 62% lower losses when compared to the single-network PINN architecture for the given problem. Furthermore, the single-network variant struggled to ensure species mass conservation in different areas of the computational domain, whereas the segregated approach successfully maintained species conservation. The PINN predicted velocity, temperature, and species profiles for a given set of boundary conditions were compared to results generated using OpenFOAM software. Both the single- and segregated-network PINN models produced accurate results for temperature and velocity profiles, with average percentage difference relative to the computational fluid dynamics results of approximately 7.5% for velocity and 8% for temperature. The mean error percentages for the species mass fractions are 9% for the single-network model and 1.5% for the segregated-network approach. To showcase the applicability of PINNs for surrogate modeling of multi-species problems, a parameterized version of the segregated-network PINN is trained that could produce results for different water vapor inlet velocities. The normalized mean absolute percentage errors, relative to the OpenFOAM results, across three predicted cases for velocity and temperature are approximately 7.5% and 2.4% for water vapor mass fraction.

Recurrent Neural Networks for Hierarchically Mapping Human-Robot Poses

To perform many critical manipulation tasks successfully, human-robot mimicking systems should not only accurately copy the position of a human hand, but its orientation as well. Deep learning methods trained from pairs of corresponding human and robot poses offer one promising approach for constructing a human-robot mapping to accomplish this. However, ignoring the spatial and temporal structure of this mapping makes learning it less effective. We propose two different hierarchical architectures that leverage the structural and temporal human-robot mapping. We partially separate the robotic manipulator's end-effector position and orientation while considering the mutual coupling effects between them. This divides the main problem---making the robot match the human's hand position and mimic its orientation accurately along an unknown trajectory---into several sub-problems. We address these using different recurrent neural networks (RNNs) with Long-Short Term Memory (LSTM) that we combine and train hierarchically based on the coupling over the aspects of the robot that each controls. We evaluate our proposed architectures using a virtual reality system to track human table tennis motions and compare with single artificial neural network (ANN) and RNN models. We compare the benefits of using deep learning neural networks with and without our architectures and find smaller errors in position and orientation, along with increased flexibility in wrist movement are obtained by our proposed architectures. Also, we propose a hybrid approach to collect the training dataset. The hybrid training dataset is collected by two approaches when the robot mimics human motions (standard learn from demonstrator LfD) and when the human mimics robot motions (LfDr). We evaluate the hybrid training dataset and show that the performance of the machine learning system trained by the hybrid training dataset is better with less error and faster training time compared to using the collected dataset using standard LfD approach.

Variable-length image compression based on controllable learning network

While neural-network-based lossy image compression methods have shown impressive performance, most of them output a fixed-length coding using a trained-specific network. However, it is essential to support the variable-length compression or meet a target rate with a high-coding performance in practice. This paper steps forward the neural-network-based image compression method, making it possible for a single network model to generate variable compression rates. Our network model combines an auto-encoder (AE) and a generative adversarial network (GAN) for generative compression. We introduce a noise interference mechanism to train the feature representation produced by the encoder, making the feature nodes training controllable and distributed from top to bottom according to their importance in feature expression. Based on this importance distribution, the latent nodes are quantized into bits and the variable-length compression can be achieved by discarding bits of those less-important feature nodes to meet the compression target. We propose several noise interference methods, and the experiments confirm the feasibility of method Random-add and Dropout in controllable learning. Further experiments illustrate that our compression method can not only achieve variable-length compression but also can recover high-quality compressed images at extremely low bit rates, outperforming that with a fixed rate.

IMU Based Deep Stride Length Estimation With Self-Supervised Learning

Stride length estimation using inertial measurement unit (IMU) sensors is getting popular recently as one representative gait parameter for health care and sports training. The traditional estimation method requires some explicit calibrations and design assumptions. Current deep learning methods suffer from few labeled data problem. To solve above problems, this paper proposes a single convolutional neural network (CNN) model to predict stride length of running and walking and classify the running or walking type per stride. The model trains its pretext task with self-supervised learning on a large unlabeled dataset for feature learning, and its downstream task on the stride length estimation and classification tasks with supervised learning with a small labeled dataset. The proposed model can achieve better average percent error, 4.78\%, on running and walking stride length regression and 99.83\% accuracy on running and walking classification, when compared to the previous approach, 7.44\% on the stride length estimation.

Multi-model fusion framework based on multi-input cross-language emotional speech recognition

Aiming at the limitations of characteristic performance and network performance in cross-language emotional speech recognition, this paper proposes a multi-model fusion framework based on multi-input cross-language emotional speech recognition. First of all, four kinds of emotional speech shared by four languages are selected as the experimental samples. Secondly, the affective features of three different modes of multi-lingual emotional speech signals are combined with SVM and two deep neural networks (MobileNet26 and ResNet38) to form the basic framework of multi-input corresponding multi-model fusion, in which the feature map in the deep neural network model carries out global maximum pooling and global average pooling so as to capture different features to double the diversity of the model. Finally, through the comparative experimental results, it is found that the multi-model fusion framework can distinguish the emotional differences of multiple languages more effectively than the single network model. At the same time, through the learning of large languages, we can also achieve transfer learning in small language emotional speech recognition to effectively increase the learning ability of the model.

Multi-model fusion framework based on multi-input cross-language emotional speech recognition

Aiming at the limitations of characteristic performance and network performance in cross-language emotional speech recognition, this paper proposes a multi-model fusion framework based on multi-input cross-language emotional speech recognition. First of all, four kinds of emotional speech shared by four languages are selected as the experimental samples. Secondly, the affective features of three different modes of multi-lingual emotional speech signals are combined with SVM and two deep neural networks (MobileNet26 and ResNet38) to form the basic framework of multi-input corresponding multi-model fusion, in which the feature map in the deep neural network model carries out global maximum pooling and global average pooling so as to capture different features to double the diversity of the model. Finally, through the comparative experimental results, it is found that the multi-model fusion framework can distinguish the emotional differences of multiple languages more effectively than the single network model. At the same time, through the learning of large languages, we can also achieve transfer learning in small language emotional speech recognition to effectively increase the learning ability of the model.

Forecast of LSTM-XGBoost in Stock Price Based on Bayesian Optimization

The prediction of the “ups and downs” of stock market prices is one of the important undertakings of the financial market. Since accurate prediction helps foster considerable economic benefits, stock market prediction has attracted significant interest by both investors and researchers. Efforts into building an accurate, stable and effective model to predict stock prices’ movements have been proliferating at a fast pace, to meet such a challenge. Firstly, this paper uses a correlation analysis to analyze the attributes of a stock dataset, processing missing values, determining the data attributes to be retained data, then divide it in a training set and a testing set. Then, the LSTM model is subsequently used to predict the retained attributes after the analysis and retention of prediction results to construct a new testing set. At the same time, the Bo-XGBoost model based on XGBoost is used to train the original training set. It uses Bayesian calculation to optimize the parameters which are proved to be difficult to find the best solution in the XGBoost model. Finally, the LSTM-BO-XGBoost model proposed in this paper is applied to “ES=F”, “YM=F”, “CL=F”, “^TNX”, “^N225”, “NQ=F”, “AAPL”, “GC= F”, “JPY=X” and “SI=F” rates with 10 stocks in the forecast and evaluated by four evaluation indexes: root mean square error (RMSE), average absolute error (MAE), accuracy rate, and F1-score. It is found that the LSTM-BO-XGBoost model proposed in this paper performs better than LSTM in stock price prediction. In order to further evaluate the performance of the algorithm, the LSTM-BO-XGBoost mode is compared with the single LSTM network model and RNN network model, the LSTM-BO-XGBoost hybrid model. The results show that the LSTM-BO-XGBoost mode has high performance, stability and feasibility than the others.

COVID-19 Mortality Risk Prediction Using X-Ray Images

The pandemic caused by coronavirus COVID-19 has already had a massive impact in our societies in terms of health, economy, and social distress. One of the most common symptoms caused by COVID-19 are lung problems like pneumonia, which can be detected using X-ray images. On the other hand, the popularity of Machine Learning models has grown exponentially in recent years and Deep Learning techniques have become the state-of-the-art for image classification tasks and is widely used in the healthcare sector nowadays as support for clinical decisions. This research aims to build a prediction model based on Machine Learning, including Deep Learning, techniques to predict the mortality risk of a particular patient given an X-ray and some basic demographic data. Keeping this in mind, this paper has three goals. First, we use Deep Learning models to predict the mortality risk of a patient based on this patient X-ray images. For this purpose, we apply Convolutional Neural Networks as well as Transfer Learning techniques to mitigate the effect of the reduced amount of COVID19 data available. Second, we propose to combine the prediction of this Convolutional Neural Network with other patient data, like gender and age, as input features of a final Machine Learning model, that will act as second and final layer. This second model layer will aim to improve the goodness of fit and prediction power of our first layer. Finally, and in accordance with the principle of reproducible research, the data used for the experiments is publicly available and we make the implementations developed easily accessible via public repositories. Experiments over a real dataset of COVID-19 patients yield high AUROC values and show our two-layer framework to obtain better results than a single Convolutional Neural Network (CNN) model, achieving close to perfect classification.

Asymmetric Mutual Mean-Teaching for Unsupervised Domain Adaptive Person Re-Identification

Unsupervised domain adaptive person re-identification aims to solve the problem of poor performance caused by transferring an unlabeled target domain from the labeled source domain in the re-identification task. The clustering pseudo-labels method in unsupervised learning is widely used in unsupervised adaptive person re-identification tasks, and it maintains state-of-the-art performance. However, pseudo-labels obtained through clustering often have much noise, and the use of a single network model structure and a single clustering algorithm can easily cause model learning to stagnate, making the model not generalizable. To solve this problem, this paper proposes an asymmetric mutual mean-teaching method for unsupervised adaptive person re-identification. In terms of feature extraction, two asymmetric network models with different structures are used for mutual mean-teaching on the target domain, making the features extracted by the network more robust. In terms of feature clustering, two clustering methods are used for mutual teaching to dynamically update the centroid of clusters to improve the confidence of clustering pseudo-labels. Finally, the triplet loss is improved based on the updated cluster centroid to improve the clustering effect. The proposed method is used to perform a large number of verification experiments on three public datasets. The experimental results show that the proposed method has better accuracy than other unsupervised person re-identification based on clustering pseudo-labels.

Prediction of chaotic time series using hybrid neural network and attention mechanism

Chaotic time series forecasting has been widely used in various domains, and the accurate predicting of the chaotic time series plays a critical role in many public events. Recently, various deep learning algorithms have been used to forecast chaotic time series and achieved good prediction performance. In order to improve the prediction accuracy of chaotic time series, a prediction model (Att-CNN-LSTM) is proposed based on hybrid neural network and attention mechanism. In this paper, the convolutional neural network (CNN) and long short-term memory (LSTM) are used to form a hybrid neural network. In addition, a attention model with &lt;i&gt;softmax&lt;/i&gt; activation function is designed to extract the key features. Firstly, phase space reconstruction and data normalization are performed on a chaotic time series, then convolutional neural network (CNN) is used to extract the spatial features of the reconstructed phase space, then the features extracted by CNN are combined with the original chaotic time series, and in the long short-term memory network (LSTM) the combined vector is used to extract the temporal features. And then attention mechanism captures the key spatial-temporal features of chaotic time series. Finally, the prediction results are computed by using spatial-temporal features. To verify the prediction performance of the proposed hybrid model, it is used to predict the Logistic, Lorenz and sunspot chaotic time series. Four kinds of error criteria and model running times are used to evaluate the performance of predictive model. The proposed model is compared with hybrid CNN-LSTM model, the single CNN and LSTM network model and least squares support vector machine(LSSVM), and the experimental results show that the proposed hybrid model has a higher prediction accuracy.