Neural Network Architecture Research Articles

Baby Cry Sound Detection: A Comparison of Mel Spectrogram Image on Convolutional Neural Network Models

Baby cries contain patterns that indicate their needs, such as pain, hunger, discomfort, colic, or fatigue. This study explores the use of Convolutional Neural Network (CNN) architectures for classifying baby cries using Mel Spectrogram images. The primary objective of this research is to compare the effectiveness of various CNN architectures such as VGG-16, VGG-19, LeNet-5, AlexNet, ResNet-50, and ResNet-152 in detecting baby needs based on their cries. The datasets used include the Donate-a-Cry Corpus and Dunstan Baby Language. The results show that AlexNet achieved the best performance with an accuracy of 84.78% on the Donate-a-Cry Corpus dataset and 72.73% on the Dunstan Baby Language dataset. Other models like ResNet-50 and LeNet-5 also demonstrated good performance although their computational efficiency varied, while VGG-16 and VGG-19 exhibited lower performance. This research provides significant contributions to the understanding and application of CNN models for baby cry classification. Practical implications include the development of baby cry detection applications that can assist parents and healthcare provide.

Lightweight deep learning model for underwater waste segmentation based on sonar images

In recent years, the rapid accumulation of marine waste not only endangers the ecological environment but also causes seawater pollution. Traditional manual salvage methods often have low efficiency and pose safety risks to human operators, making automatic underwater waste recycling a mainstream approach. In this paper, we propose a lightweight multi-scale cross-level network for underwater waste segmentation based on sonar images that provides pixel-level location information and waste categories for autonomous underwater robots. In particular, we introduce hybrid perception and multi-scale attention modules to capture multi-scale contextual features and enhance high-level critical information, respectively. At the same time, we use sampling attention modules and cross-level interaction modules to achieve feature down-sampling and fuse detailed features and semantic features, respectively. Relevant experimental results indicate that our method outperforms other semantic segmentation models and achieves 74.66 % mIoU with only 0.68 M parameters. In particular, compared with the representative PIDNet Small model based on the convolutional neural network architecture, our method can improve the mIoU metric by 1.15 percentage points and can reduce model parameters by approximately 91 %. Compared with the representative SeaFormer T model based on the transformer architecture, our approach can improve the mIoU metric by 2.07 percentage points and can reduce model parameters by approximately 59 %. Our approach maintains a satisfactory balance between model parameters and segmentation performance. Our solution provides new insights into intelligent underwater waste recycling, which helps in promoting sustainable marine development.

Automatic segmentation of echocardiographic images using a shifted windows vision transformer architecture

Echocardiography is one the most commonly used imaging modalities for the diagnosis of congenital heart disease. Echocardiographic image analysis is crucial to obtaining accurate cardiac anatomy information. Semantic segmentation models can be used to precisely delimit the borders of the left ventricle, and allow an accurate and automatic identification of the region of interest, which can be extremely useful for cardiologists. In the field of computer vision, convolutional neural network (CNN) architectures remain dominant. Existing CNN approaches have proved highly efficient for the segmentation of various medical images over the past decade. However, these solutions usually struggle to capture long-range dependencies, especially when it comes to images with objects of different scales and complex structures. In this study, we present an efficient method for semantic segmentation of echocardiographic images that overcomes these challenges by leveraging the self-attention mechanism of the Transformer architecture. The proposed solution extracts long-range dependencies and efficiently processes objects at different scales, improving performance in a variety of tasks. We introduce Shifted Windows Transformer models (Swin Transformers), which encode both the content of anatomical structures and the relationship between them. Our solution combines the Swin Transformer and U-Net architectures, producing a U-shaped variant. The validation of the proposed method is performed with the EchoNet-Dynamic dataset used to train our model. The results show an accuracy of 0.97, a Dice coefficient of 0.87, and an Intersection over union (IoU) of 0.78. Swin Transformer models are promising for semantically segmenting echocardiographic images and may help assist cardiologists in automatically analyzing and measuring complex echocardiographic images.

Versatile Video Coding-Post Processing Feature Fusion: A Post-Processing Convolutional Neural Network with Progressive Feature Fusion for Efficient Video Enhancement

Advanced video codecs such as High Efficiency Video Coding/H.265 (HEVC) and Versatile Video Coding/H.266 (VVC) are vital for streaming high-quality online video content, as they compress and transmit data efficiently. However, these codecs can occasionally degrade video quality by adding undesirable artifacts such as blockiness, blurriness, and ringing, which can detract from the viewer’s experience. To ensure a seamless and engaging video experience, it is essential to remove these artifacts, which improves viewer comfort and engagement. In this paper, we propose a deep feature fusion based convolutional neural network (CNN) architecture (VVC-PPFF) for post-processing approach to further enhance the performance of VVC. The proposed network, VVC-PPFF, harnesses the power of CNNs to enhance decoded frames, significantly improving the coding efficiency of the state-of-the-art VVC video coding standard. By combining deep features from early and later convolution layers, the network learns to extract both low-level and high-level features, resulting in more generalized outputs that adapt to different quantization parameter (QP) values. The proposed VVC-PPFF network achieves outstanding performance, with Bjøntegaard Delta Rate (BD-Rate) improvements of 5.81% and 6.98% for luma components in random access (RA) and low-delay (LD) configurations, respectively, while also boosting peak signal-to-noise ratio (PSNR).

A Performance Evaluation of Convolutional Neural Network Architectures for Pterygium Detection in Anterior Segment Eye Images.

In this article, various convolutional neural network (CNN) architectures for the detection of pterygium in the anterior segment of the eye are explored and compared. Five CNN architectures (ResNet101, ResNext101, Se-ResNext50, ResNext50, and MobileNet V2) are evaluated with the objective of identifying one that surpasses the precision and diagnostic efficacy of the current existing solutions. The results show that the Se-ResNext50 architecture offers the best overall performance in terms of precision, recall, and accuracy, with values of 93%, 92%, and 92%, respectively, for these metrics. These results demonstrate its potential to enhance diagnostic tools in ophthalmology.

A study on improving drug–drug interactions prediction using convolutional neural networks

Appropriate studies on drug–drug interactions (DDIs) can evade possible adverse side effects due to the ingestion of multiple drugs. This paper proposes a novel framework called Similarity Network Fusion and Hybrid Convolutional Neural Network (SNF–HCNN) to predict the DDIs better. The proposed framework leverages data from DrugBank, PubChem, and SIDER. Seven critical drug features are extracted: Target, Transporter, Enzymes, Chemical substructure, Carrier, Offside, and Side effects. The Jaccard Similarity measure evaluates the similarity of drug features to construct a comprehensive similarity matrix that effectively captures potential drug relationships and patterns. The similarity selection process identifies the most relevant features, reduces redundancy, and enhances identifying potential drug interactions. Integrating Similarity Network Fusion (SNF) with the selected similarity matrix ensures a comprehensive representation of drug features and leads to superior accuracy compared to conventional methods. Our experimental results demonstrate the effectiveness of the proposed hybrid convolutional neural network (HCNN) architectures, such as CNN+LR (CNN+Logistic Regression), CNN+RF (CNN+Random Forest), and CNN+SVM (CNN+Support Vector Machine), showing impressive accuracies of 95.19%, 94.45%, and 93.65%, respectively. Moreover, CNN+LR outperforms other approaches regarding precision, sensitivity, F1-score, and AUC score, which implicate better outcomes for ensuring medication safety aspects in clinical settings in the future.

Balancing Accuracy and Training Time in Federated Learning for Violence Detection in Surveillance Videos: A Study of Neural Network Architectures

Balancing Accuracy and Training Time in Federated Learning for Violence Detection in Surveillance Videos: A Study of Neural Network Architectures

Research on Optimization of Image Recognition Algorithm Based on Deep Learning

With the rapid development of information technology, image recognition has become increasingly crucial in numerous fields. Deep learning, as a highly potent machine learning approach, has brought about a revolutionary breakthrough in image recognition. This paper delves into the optimization of image recognition algorithms based on deep learning. Firstly, it analyzes the application status of deep learning in image recognition, covering common neural network architectures such as convolutional neural network (CNN). Then, the optimization strategy of the image recognition algorithm is elaborated from multiple aspects, including data augmentation to increase data diversity, model structure optimization for better feature extraction, and hyperparameter adjustment to enhance performance. Through extensive experiments, the effects of different optimization methods are compared meticulously. It is demonstrated that the optimized algorithm shows significant improvements in accuracy, robustness, and efficiency. Finally, the paper looks ahead to the future development trend of image recognition algorithms based on deep learning, envisioning further advancements and broader applications.

Weighted graph convolutional network with feature mask for low back pain prediction

Low back pain (LBP) is one of the most common physical disabilities worldwide, imposing a heavy medical economic burden. To help medical experts make disease predictions at an early stage, this paper proposes an LBP abnormality diagnosis model based on a weighted graph convolutional network with feature mask. Unlike popular neural network architectures, we represent the biomechanical characteristics of the spine as a weighted feature graph. The potential relationship between patients is mined through edge connections, and the importance of features for different patients is dynamically captured through the internal feature weighting of nodes. Finally, the importance score of each feature is calculated according to its identification and independence to generate the feature importance mask. It realizes feature sparsification and reinforces important features at the same time, so as to improve the accuracy and stability of the model. The proposed model is verified on a public dataset available in the Kaggle repository, and the classification accuracy achieves 90.2%, which is 2.14% higher than the current best result. Furthermore, when compared for stability against several commonly used machine learning algorithms, the standard deviation of accuracy over ten experiments is only about 1/3 of that of other algorithms.

A CNN based model for heart disease detection

Cardiovascular diseases (CVDs) pose a formidable global health challenge, claiming millions of lives annually. Despite advancements in healthcare, heart disease remains a leading cause of mortality, especially in developing nations. Early detection of cardiac abnormalities through predictive models is crucial for effective intervention. This research leverages machine learning (ML) and artificial intelligence (AI), focusing on deep learning, to enhance diagnostic capabilities. Unlike previous studies, this work introduced caffeine as a potential risk factor often overlooked in datasets. The study utilized Magnetic Resonance Imaging (MRI) datasets from Enugu State University Teaching Hospital and Kaggle, comprising 90,500 samples, with specific attention to high caffeine intake cases. Data preprocessing involved resizing, normalization, color adjustments, and augmentation to optimize model training. A Convolutional Neural Network (CNN) architecture with four convolutional layers was employed for classification. The CNN model achieved a remarkable accuracy of 94% and low loss values, and demonstrated proficiency in categorizing heart MRI data. Ten-fold cross-validation reaffirmed the model's high success rate with an average accuracy of 94.13% and minimal loss function. Comparative analysis showcased the effectiveness of the developed CNN model, outperforming several existing models in heart disease classification.

Activation function optimization scheme for image classification

Activation function has a significant impact on the dynamics, convergence, and performance of deep neural networks. The search for a consistent and high-performing activation function has always been a pursuit during deep learning model development. Existing state-of-the-art activation functions are manually designed with human expertise except for Swish. Swish was developed using a reinforcement learning-based search strategy. In this study, we propose an evolutionary approach for optimizing activation functions specifically for image classification tasks, aiming to discover functions that outperform current state-of-the-art options. Through this optimization framework, we obtain a series of high-performing activation functions denoted as Exponential Error Linear Unit (EELU). The developed activation functions are evaluated for image classification tasks from two perspectives: (1) five state-of-the-art neural network architectures, such as ResNet50, AlexNet, VGG16, MobileNet, and Compact Convolutional Transformer, which cover computationally heavy to light neural networks, and (2) eight standard datasets, including CIFAR10, Imagenette, MNIST, Fashion MNIST, Beans, Colorectal Histology, CottonWeedID15, and TinyImageNet which cover from typical machine vision benchmark, agricultural image applications to medical image applications. Finally, we statistically investigate the generalization of the resultant activation functions developed through the optimization scheme. With a Friedman test, we conclude that the optimization scheme is able to generate activation functions that outperform the existing standard ones in 92.8% cases among 28 different cases studied, and −x⋅erf(e−x) is found to be the best activation function for image classification generated by the optimization scheme.

G[formula omitted]BFNN: Generalized geodesic basis function neural network

Real-world data is typically distributed on low-dimensional manifolds embedded in high-dimensional Euclidean spaces. Accurately extracting spatial distribution features on general manifolds that reflect the intrinsic characteristics of data is crucial for effective feature representation. Therefore, we propose a generalized geodesic basis function neural network (G2BFNN) architecture. The generalized geodesic basis functions (G2BF) are defined based on generalized geodesic distances. The generalized geodesic distance metric (G2DM) is obtained by learning the manifold structure. To implement this architecture, a specific G2BFNN, named discriminative local preserving projection-based G2BFNN (DLPP-G2BFNN) is proposed. DLPP-G2BFNN mainly contains two modules, namely the manifold structure learning module (MSLM) and the network mapping module (NMM). In the MSLM module, a supervised adjacency graph matrix is constructed to constrain the learning of the manifold structure. This enables the learned features in the embedding subspace to maintain the manifold structure while enhancing the discriminability. The features and G2DM learned in the MSLM are fed into the NMM. Through the G2BF in the NMM, the spatial distribution features on manifold are obtained. Finally, the output of the network is obtained through the fully connected layer. Compared with the local response neural network based on Euclidean distance, the proposed network can reveal more essential spatial structure characteristics of the data. Meanwhile, the proposed G2BFNN is a generalized network architecture that can be combined with any manifold learning method, showcasing high scalability. The experimental results demonstrate that the proposed DLPP-G2BFNN outperforms existing methods by utilizing fewer kernels while achieving higher recognition performance.

Modelling of Joule-Thomson cooling effect using a modified shift-DeepONet method for predicting hydrate onset during CO2 sequestration

Carbon Storage in underground depleted oil and gas reservoirs play a key role in reducing anthropogenic Carbon Dioxide (CO2) emissions around the world. Injection of CO2 at high pressure into a low pressure depleted gas formations containing methane (CH4) leads to Joule-Thompson (J-T) cooling, possibly resulting in injection blocking precipitation of CO2 or CH4 hydrates. Here, we present a deep neural operator-based approach to model the J-T effects during CO2 injection in depleted reservoirs to demonstrate the effectiveness of Deep Operator Networks (DeepONets) in modelling piece-wise functions with sharp changes in gradient for cases of varying permeabilities. We propose a modified version of the shift-DeepONet model using non-linear constructors in the neural network architecture to accurately model temperature and pressure profiles in space and time. The modified shift-DeepONet model achieves a higher accuracy in modelling temperature for permeability in the range of 1.01mD to 101mD with three to up to ten times improvement in reducing Relative L2 (RL2) norm and with a testing RL2 norm of 3.8 × 10−4 compared to 2.97 × 10−3 for the Vanilla DeepONet. Using the temperature and pressure profiles, estimates are made for hydrate onset in space and time. The proposed modified shift-DeepONet method achieves high accuracy in hydrate classification with precision values > 0.99, recall values > 0.86 and F1 scores >0.92 in classification of CO2 or CH4 hydrate onset during CO2 injection into depleted gas reservoirs.

Scalable spatiotemporal prediction with Bayesian neural fields

Spatiotemporal datasets, which consist of spatially-referenced time series, are ubiquitous in diverse applications, such as air pollution monitoring, disease tracking, and cloud-demand forecasting. As the scale of modern datasets increases, there is a growing need for statistical methods that are flexible enough to capture complex spatiotemporal dynamics and scalable enough to handle many observations. This article introduces the Bayesian Neural Field (BayesNF), a domain-general statistical model that infers rich spatiotemporal probability distributions for data-analysis tasks including forecasting, interpolation, and variography. BayesNF integrates a deep neural network architecture for high-capacity function estimation with hierarchical Bayesian inference for robust predictive uncertainty quantification. Evaluations against prominent baselines show that BayesNF delivers improvements on prediction problems from climate and public health data containing tens to hundreds of thousands of measurements. Accompanying the paper is an open-source software package (https://github.com/google/bayesnf) that runs on GPU and TPU accelerators through the Jax machine learning platform.

Optimizing graph neural network architectures for schizophrenia spectrum disorder prediction using evolutionary algorithms

Background and Objective:The accurate diagnosis of schizophrenia spectrum disorder plays an important role in improving patient outcomes, enabling timely interventions, and optimizing treatment plans. Functional connectivity analysis, utilizing functional magnetic resonance imaging data, has been demonstrated to offer invaluable biomarkers conducive to clinical diagnosis. However, previous studies mainly focus on traditional machine learning methods or hand-crafted neural networks, which may not fully capture the spatial topological relationship between brain regions. Methods:This paper proposes an evolutionary algorithm (EA) based graph neural architecture search (GNAS) method. EA-GNAS has the ability to search for high-performance graph neural networks for schizophrenia spectrum disorder prediction. Moreover, we adopt GNNExplainer to investigate the explainability of the acquired architectures, ensuring that the model’s predictions are both accurate and comprehensible. Results:The results suggest that the graph neural network model, derived using genetic algorithm search, outperforms under five-fold cross-validation, achieving a fitness of 0.1850. Relative to conventional machine learning and other deep learning approaches, the proposed method yields superior accuracy, F1 score, and AUC values of 0.8246, 0.8438, and 0.8258, respectively. Conclusion:Based on a multi-site dataset from schizophrenia spectrum disorder patients, the findings reveal an enhancement over prior methods, advancing our comprehension of brain function and potentially offering a biomarker for diagnosing schizophrenia spectrum disorder.

A Deep Learning Model for Cervical Optical Coherence Tomography Image Classification.

Objectives: Optical coherence tomography (OCT) has recently been used in gynecology to detect cervical lesions in vivo and proven more effective than colposcopy in clinical trials. However, most gynecologists are unfamiliar with this new imaging technique, requiring intelligent computer-aided diagnosis approaches to help them interpret cervical OCT images efficiently. This study aims to (1) develop a clinically-usable deep learning (DL)-based classification model of 3D OCT volumes from cervical tissue and (2) validate the DL model's effectiveness in detecting high-risk cervical lesions, including high-grade squamous intraepithelial lesions and cervical cancer. Method: The proposed DL model, designed based on the convolutional neural network architecture, combines a feature pyramid network (FPN) with texture encoding and deep supervision. We extracted, represent, and fused four-scale texture features to improve classification performance on high-risk local lesions. We also designed an auxiliary classification mechanism based on deep supervision to adjust the weight of each scale in FPN adaptively, enabling low-cost training of the whole model. Results: In the binary classification task detecting positive subjects with high-risk cervical lesions, our DL model achieved an 81.55% (95% CI, 72.70-88.51%) F1-score with 82.35% (95% CI, 69.13-91.60%) sensitivity and 81.48% (95% CI, 68.57-90.75%) specificity on the Renmin dataset, outperforming five experienced medical experts. It also achieved an 84.34% (95% CI, 74.71-91.39%) F1-score with 87.50% (95% CI, 73.20-95.81%) sensitivity and 90.59% (95% CI, 82.29-95.85%) specificity on the Huaxi dataset, comparable to the overall level of the best investigator. Moreover, our DL model provides visual diagnostic evidence of histomorphological and texture features learned in OCT images to assist gynecologists in making clinical decisions quickly. Conclusions: Our DL model holds great promise to be used in cervical lesion screening with OCT efficiently and effectively.

Rewireable Building Blocks for Enzyme-Powered DNA Computing Networks.

Neural networks enable the processing of large, complex data sets with applications in disease diagnosis, cell profiling, and drug discovery. Beyond electronic computers, neural networks have been implemented using programmable biomolecules such as DNA; this confers unique advantages, such as greater portability, electricity-free operation, and direct analysis of patterns of biomolecules in solution. Analogous to bottlenecks in electronic computers, the computing power of DNA-based neural networks is limited by the ability to add more computing units, i.e., neurons. This limitation exists because current architectures require many nucleic acids to model a single neuron. Each additional neuron compounds existing problems such as long assembly times, high background signal, and cross-talk between components. Here, we test three strategies to solve this limitation and improve the scalability of DNA-based neural networks: (i) enzymatic synthesis for high-purity neurons, (ii) spatial patterning of neuron clusters based on their network position, and (iii) encoding neuron connectivity on a universal single-stranded DNA backbone. We show that neurons implemented via these strategies activate quickly, with a high signal-to-background ratio and process-weighted inputs. We rewired our modular neurons to demonstrate basic neural network motifs such as cascading, fan-in, and fan-out circuits. Finally, we designed a prototype two-layer microfluidic device to automate the operation of our circuits. We envision that our proposed design will help scale DNA-based neural networks due to its modularity, simplicity of synthesis, and compatibility with various neural network architectures. This will enable portable computing power for applications in portable diagnostics, compact data storage, and autonomous decision making for lab-on-a-chips.

AI-Enabled Sensor Fusion of Time-of-Flight Imaging and mmWave for Concealed Metal Detection.

In the field of detection and ranging, multiple complementary sensing modalities may be used to enrich information obtained from a dynamic scene. One application of this sensor fusion is in public security and surveillance, where efficacy and privacy protection measures must be continually evaluated. We present a novel deployment of sensor fusion for the discrete detection of concealed metal objects on persons whilst preserving their privacy. This is achieved by coupling off-the-shelf mmWave radar and depth camera technology with a novel neural network architecture that processes radar signals using convolutional Long Short-Term Memory (LSTM) blocks and depth signals using convolutional operations. The combined latent features are then magnified using deep feature magnification to reveal cross-modality dependencies in the data. We further propose a decoder, based on the feature extraction and embedding block, to learn an efficient upsampling of the latent space to locate the concealed object in the spatial domain through radar feature guidance. We demonstrate the ability to detect the presence and infer the 3D location of concealed metal objects. We achieve accuracies of up to 95% using a technique that is robust to multiple persons. This work provides a demonstration of the potential for cost-effective and portable sensor fusion with strong opportunities for further development.

Exploring the power of KANs: Overcoming MLP limitations in complex data analysis

In the field of artificial intelligence, the requirements of models that can adeptly explore informative features of complex data sets has led to the emergence of advanced neural network architectures. Beyond the perceptron-based architectures that are currently in widespread use, a variety of innovative and cutting-edge designs have been proposed. This paper thoroughly explores the critical role of Kolmogorov-Arnold Networks (KAN) in overcoming the limitations of traditional Multilayer Perceptron (MLP) models, particularly highlighting KANs significant advantages in handling complex nonlinear and high-dimensional data. By analyzing the mathematical foundations and neural network architecture of KAN, and comparing it with MLP, the paper demonstrates KANs outstanding performance in time series analysis and image classification. The research indicates that KAN has distinct advantages in addressing high-dimensional nonlinear data. The paper also summarizes the current research progress on KAN and discusses its enormous potential in future machine learning and real-world applications, pointing out possible future research directions.

Subject-Independent Wearable P300 Brain-Computer Interface Based on Convolutional Neural Network and Metric Learning.

The calibration procedure for a wearable P300 brain-computer interface (BCI) greatly impact the user experience of the system. Each user needs to spend additional time establishing a decoder adapted to their own brainwaves. Therefore, achieving subject independent is an urgent issue for wearable P300 BCI needs to be addressed. A dataset of electroencephalogram (EEG) signals was constructed from 100 individuals by conducting a P300 speller task with a wearable EEG amplifier. A framework is proposed that initially improves cross-subject consistency of EEG features through a common feature extractor. Subsequently, a simple and compact convolutional neural network (CNN) architecture is employed to learn an embedding sub-space, where the mapped EEG features are maximally separated, while pursuing the minimum distance within the same class and the maximum distance between different classes. Finally, the model's generalization capability was further optimized through fine-tuning. The proposed method significantly boosts the average accuracy of wearable P300 BCI to 73.23 ±7.62% without calibration and 78.75±6.37% with fine-tuning. The results demonstrate the feasibility and excellent performance of our dataset and framework. A calibration-free wearable P300 BCI system is feasible, suggesting significant potential for practical applications of the wearable P300 BCI system.