Deep Neural Network Research Articles

Physical Adversarial Attack Meets Computer Vision: A Decade Survey.

Despite the impressive achievements of Deep Neural Networks (DNNs) in computer vision, their vulnerability to adversarial attacks remains a critical concern. Extensive research has demonstrated that incorporating sophisticated perturbations into input images can lead to a catastrophic degradation in DNNs' performance. This perplexing phenomenon not only exists in the digital space but also in the physical world. Consequently, it becomes imperative to evaluate the security of DNNs-based systems to ensure their safe deployment in real-world scenarios, particularly in security-sensitive applications. To facilitate a profound understanding of this topic, this paper presents a comprehensive overview of physical adversarial attacks. First, we distill four general steps for launching physical adversarial attacks. Building upon this foundation, we uncover the pervasive role of artifacts carrying adversarial perturbations in the physical world. These artifacts influence each step. To denote them, we introduce a new term: adversarial medium. Then, we take the first step to systematically evaluate the performance of physical adversarial attacks, taking the adversarial medium as a first attempt. Our proposed evaluation metric, hiPAA, comprises six perspectives: Effectiveness, Stealthiness, Robustness, Practicability, Aesthetics, and Economics. We also provide comparative results across task categories, together with insightful observations and suggestions for future research directions.

Insight mixed deep neural network architectures for molecular representation

Learning molecular representation is a crucial task in the field of drug discovery, particularly for various specific applications such as predicting molecular properties. Current methods are mainly based on deep neural network models, such as convolutional neural networks (CNN), recurrent neural networks (RNN), graph neural networks (GNN) and their mixed models. However, these neural network models mentioned above do not provide detailed explanations for the ability to learn molecular representations and why such experimental results occurred. In this paper, we aim to compare the performance of these models in predicting molecular properties based on molecular representation, and give more insight into deep neural network architectures for specific molecular tasks. Our experimental results demonstrate that the graph neural network can obtain superior performance on the regression tasks, while the mixed deep neural network models show better performance on the classification tasks. Ablation study also gives more explanation and analysis to the experimental results.

DeepCORE: An interpretable multi-view deep neural network model to detect co-operative regulatory elements

Gene transcription is an essential process involved in all aspects of cellular functions with significant impact on biological traits and diseases. This process is tightly regulated by multiple elements that co-operate to jointly modulate the transcription levels of target genes. To decipher the complicated regulatory network, we present a novel multi-view attention-based deep neural network that models the relationship between genetic, epigenetic, and transcriptional patterns and identifies co-operative regulatory elements (COREs). We applied this new method, named DeepCORE, to predict transcriptomes in various tissues and cell lines, which outperformed the state-of-the-art algorithms. Furthermore, DeepCORE contains an interpreter that extracts the attention values embedded in the deep neural network, maps the attended regions to putative regulatory elements, and infers COREs based on correlated attentions. The identified COREs are significantly enriched with known promoters and enhancers. Novel regulatory elements discovered by DeepCORE showed epigenetic signatures consistent with the status of histone modification marks.

Design and implementation of deep neural network hardware chip and its performance analysis

The artificial neural network (ANN) with a single layer has a limited capacity to process data. Multiple neurons are connected in the human brain, and the actual capacity of the brain lies in the interconnectedness of multiple neurons. As a specified generalization of ANN deep learning makes use of two or more hidden layers, which implies that a greater number of neurons are required to construct the model. A network that has more than one hidden layer, also known as two or more hidden layers, is referred to as a deep neural network, and the process of training such networks is referred to as deep learning. The research article focuses on the design of a multilayer or deep neural network presented for the target field programmable gate array (FPGA) device spartan-6 (xc6stx4-2t9g144) FPGA. The simulation is carried out using Xilinx ISE and ModelSim software. There are two hidden layers in which (2×1) multiplexer blocks are utilized for processing twenty neurons into the output of ten neurons in the first hidden layer and demultiplexers (1×2) and vice versa. The hardware utilization is estimated on FPGA to compute the performance of the deep neural hardware chip based on memory, flip flops, delay, and frequency parameters. The design is scalable and applicable to various FPGA devices, which makes the work novel. FPGA-based neuromorphic hardware acceleration platform with high speed and low power for discrete spike processing on hardware with great real-time performance.

On the turnpike to design of deep neural networks: Explicit depth bounds

It is well-known that the training of Deep Neural Networks (DNN) can be formalized in the language of optimal control. In this context, this paper leverages classical turnpike properties of optimal control problems to attempt a quantifiable answer to the question of how many layers should be considered in a DNN. The underlying assumption is that the number of neurons per layer—i.e., the width of the DNN—is kept constant. Pursuing a different route than the classical analysis of approximation properties of sigmoidal functions, we prove explicit bounds on the required depths of DNNs based on asymptotic reachability assumptions and a dissipativity-inducing choice of the regularization terms in the training problem. Numerical results obtained for the two spiral task data set for classification indicate that the proposed constructive estimates can provide non-conservative depth bounds.

Comparative exploration of deep convolutional neural networks using real-time endoscopy images

Until now various deep convolutional neural networks are designed and trained for the purpose of classifying different medical conditions related to the domain of gastroenterology. Most of the study carried out have considered publicly available datasets to train the classification networks. Nevertheless, the main motive for carrying out different works in the field gastroenterology is to administer the developed models in healthcare centers in real-world set-ups. For doing so, it is important to check the generalizing ability of the designed systems by regulating them so as to classify endoscopy images captured in a specific hospital. In this regard, the foremost work completed is the collection of the endoscopy data from the hospital and then correctly annotating the images taking the help of a senior endoscopist with experience of more than 5 years. Once the data annotation is completed, the images are segregated into the class of normal and abnormal endoscopy images. Four different models are designed in the current work based on deep learning models, transfer learning models and ensemble approaches, and trained to classify the hospital endoscopy data as normal or abnormal. The models are then tested and evaluated based on various performance measures. It is observed from the comparative analysis that the transfer learning-based ensemble model has the best generalizing ability and gives the best specificity of 100 ​%. It is believed that deep learning-based models can assist endoscopists in add-on to human prediction efficiency.

Sentiment-electroencephalogram fusion for efficient product review prediction using correlation-based deep learning neural network

<span lang="EN-US">Various techniques have been proposed and implemented in previous work for sentiment analysis prediction. However, achieving satisfactory quality of description and fault prediction remains a challenging task. To overcome these limitations, this study proposes an efficient prediction technique that utilizes sentiment analysis of product reviews and electroencephalogram (EEG) signals using correlation-based deep learning neural network (CDNN). The study employs two types of datasets: EEG signals and Amazon product reviews. During the pre-processing phase, EEG signals undergo normalization, while Amazon product reviews undergo tokenization, stop word removal, and weighting factor application to convert unstructured data into a structured format. Subsequently, the pre-processed EEG signals and reviews are analyzed to extract features like emotion, demographic information, personality traits, and sentiment. These features are then employed in sentiment analysis via an entropy-based deep-learning neural network. The proposed CDNN utilizes the grasshopper optimization algorithm (EGOA) to optimize hyperparameters for each layer. Comparative performance assessment against established methods like convolutional neural network (CNN), <a name="_Hlk167111148"></a>long short-term memory (LSTM), multiclass <a name="_Hlk175651957"></a>support vector machine (M-SVM), and bidirectional encoder representations from transformers (BERT) is conducted, and the results are evaluated. Experimental result reveal that the proposed system outperforms traditional approaches.</span>

Microstructural material design of pearlitic steel lamella for desired mechanical properties

Designing materials for a targeted mechanical property is a complex problem due to its structure and property relationship nature, traditionally relying on experimental methods. Computational approaches offer a promising alternative, but they are often computationally intensive. This limits their applicability for solving the goal-based structure design problem. This study focuses on the design of microstructure to achieve specific mechanical properties by combining generative adversarial networks, deep neural networks, and genetic algorithm optimization techniques. In this work to demonstrate the effectiveness of the proposed approach we select a widely used pearlitic steel lamella microstructure where the phase fraction of ferrite and cementite and the lamella distribution significantly influence the strength and ductility. A method for designing the pearlite microstructure features for the targeted mechanical properties, such as yield strength (σy), ultimate strength (σut), and fracture strain (εf) is developed by applying inverse microstructure design techniques to achieve the targeted mechanical properties, specifically in fully pearlitic steel materials. The findings hold great promise, particularly when coupled with additive manufacturing, in offering an efficient and versatile framework for structural materials design for targeted material applications.

Optimized deep neural network based vulnerability detection enabled secured testing for cloud SaaS

Based on the information technology service model, an on-demand services towards user becomes cost effective, which is provided with cloud computing. The network attack is detected with research community that pays huge interest. The novel proposed framework is intended with the combination of mitigation and detection of attack. While enormous traffic is obtainable, extract the relevant fields decide with Software-as-a-service (SaaS) provider. According to the network vulnerability and mitigation procedure, perform deep learning-based attack detection model. The golf optimization algorithm (GOA) done the selection of features followed by deep neural network (DNN) detect the attacks from the selected features. The correntropy variational features validates the level of risk and performs vulnerability assessment. Perform the process of bait-oriented mitigation during the phase of attack mitigation. The proposed approach demonstrates 0.97kbps throughput with 0.2% packet loss ratio than traditional methods.

Micromechanics and microstructure based machine learning approach: Unveiling the role of porosity and hydrated phases on the tensile behaviour of cement pastes

The fracture process and the mechanical performance of the cement paste rely significantly on its tensile behaviour. Machine learning methods have been recently developed to predict this behaviour based on numerous experimental measurements. Due to multi-phase heterogeneous nature of cement paste microstructure, evaluation of the role of each chemical phase on the cracking of cement paste is necessary to develop optimized binder systems. Numerous experimental tests would be necessary to determine the best cement pastes in a considered environment. Therefore, in this study, a hybrid approach based on an ensemble machine learning model and a microstructure informed micromechanical model is proposed for predicting the tensile behaviour of hydrated cement pastes. The model can be used as a quick prediction tool for tensile strength of a given microstructure or for deep analysis on the optimization of binder system. A Deep Neural Networks (DNN) model was developed by training and validating a large dataset detailing the chemical phase composition, phase volume fractions and the tensile strength of the cement pastes. This dataset covers an extensive range of Portland cements, hydration ages and water-to-cement ratios. It was developed by simulating cement thermodynamics and tensile behaviour using a damage micromechanical model; which is otherwise almost impossible to obtain through experimental means. Additionally, two generative algorithms, tabular generative adversial networks (TGAN) and conditional generative adversial networks (CTGAN), were employed for dataset augmentation as a sampling method for the sensitivity analysis. This machine learning study provides an in-depth investigation to identify the role of volume fraction of solid phases (C-S-H, ettringite, portlandite and hydrogarnet) and micro porosity (dry and filled) in affecting the tensile behaviour of hydrated cement paste. Optimization assessment of the microstructure was then performed using micromechanical analysis and damage propagation in hydrated cement paste phases and porous network.

WISE: Efficient WSI Selection for Active Learning in Histopathology

Deep neural network (DNN) models have been applied to a wide variety of medical image analysis tasks, often with the successful performance outcomes that match those of medical doctors. However, given that even minor errors in a model can impact patients’ life, it is critical that these models are continuously improved. Hence, active learning (AL) has garnered attention as an effective and sustainable strategy for enhancing DNN models for the medical domain. Extant AL research in histopathology has primarily focused on patch datasets derived from whole-slide images (WSIs), a standard form of cancer diagnostic images obtained from a high-resolution scanner. However, this approach has failed to address the selection of WSIs, which can impede the performance improvement of deep learning models and increase the number of WSIs needed to achieve the target performance.This study introduces a WSI-level AL method, termed WSI-informative selection (WISE). WISE is designed to select informative WSIs using a newly formulated WSI-level class distance metric. This method aims to identify diverse and uncertain cases of WSIs, thereby contributing to model performance enhancement. WISE demonstrates state-of-the-art performance across the Colon and Stomach datasets, collected in the real world, as well as the public DigestPath dataset, significantly reducing the required number of WSIs by more than threefold compared to the one-pool dataset setting, which has been dominantly used in the field.

Exploring deep learning models for roadside landslide prediction: Insights and implications from comparative analysis

This study undertakes a comparative analysis of four distinct deep learning models, i.e., Convolutional Neural Network (CNN), Deep Neural Network (DNN), Recurrent Neural Network (RNN), and Long Short-Term Memory (LSTM), in the context of roadside landslide prediction, aiming to provide comprehensive insights into their strengths and weaknesses. A geospatial dataset from the Lai Châu province, Vietnam, with thirteen environmental factors (elevation, aspect, slope, curvature, topographic wetness index, stream power index, flow accumulation, geology, normalized difference vegetation index, maximum rainfall, average annual rainfall, and proximity to faults and rivers) and 284 road-along landslides was considered for analysis. Our modeling efforts yielded invaluable insights into the performance of these models during both training and validation phases. The DNN model emerged as the frontrunner in the training phase, boasting the highest area under the curve (AUC) of 0.94, accuracy of 87.47%, kappa of 0.748, and lowest RMSE of 0.125. However, during validation, the CNN model outshone others, exhibiting the highest AUC of 0.88 and overall accuracy of 80.00%. Despite variations in performance metrics across phases, CNN consistently demonstrated robust predictive prowess. The findings of this study underscore the significance of selecting appropriate machine learning models tailored to specific contexts and objectives. Moreover, they contribute valuable insights for decision-makers and researchers alike, ultimately aiming to enhance the safety and resilience of communities inhabiting landslide-prone areas. Moving forward, future research directions may explore ensemble methods, novel architectures, and interpretability techniques to further advance predictive accuracy and applicability in roadside landslide susceptibility modeling.

Small aircraft detection in infrared aerial imagery based on deep neural network

Detection of aerial target is an important part of infrared image processing. Both neural network method and traditional method can be used in infrared object detection. Neural network method has many advantages such as high accuracy and good portability compared with traditional object detection method. Since the features extracted by neural network method can change over detection target, automatic feature extraction makes neural network based detection method more effective. In recent years deep learning method has been also found wide use for object detection in images. In this paper, an object detection model based on the deep learning network YOLO is constructed for solving the infrared aircraft detection problem. We construct the dataset used for training and testing with recognized features being iteratively learned. The task of infrared otject detection is sensitive to model size and detection speed. There is a requirement of using quantization method to reduce the storage space and the computation complexity. We propose a quantized model with appropriate accuracy for infrared object detection task. To solve the detection task for multiple extremely small aircrafts, model adjustment and quantization are used in proposed model and it gets a better performance. Experimental results on the constructed dataset show that the storage space for weight after quantization shrinks to a quarter, and there is no precision loss for extremely small aircrafts compared to the original model. The optimized YOLO-based deep learning model is effective to detect the small aircraft target in infrared aerial imagery.

Towards Lightweight Super-Resolution With Dual Regression Learning.

Deep neural networks have exhibited remarkable performance in image super-resolution (SR) tasks by learning a mapping from low-resolution (LR) images to high-resolution (HR) images. However, the SR problem is typically an ill-posed problem and existing methods would come with several limitations. First, the possible mapping space of SR can be extremely large since there may exist many different HR images that can be super-resolved from the same LR image. As a result, it is hard to directly learn a promising SR mapping from such a large space. Second, it is often inevitable to develop very large models with extremely high computational cost to yield promising SR performance. In practice, one can use model compression techniques to obtain compact models by reducing model redundancy. Nevertheless, it is hard for existing model compression methods to accurately identify the redundant components due to the extremely large SR mapping space. To alleviate the first challenge, we propose a dual regression learning scheme to reduce the space of possible SR mappings. Specifically, in addition to the mapping from LR to HR images, we learn an additional dual regression mapping to estimate the downsampling kernel and reconstruct LR images. In this way, the dual mapping acts as a constraint to reduce the space of possible mappings. To address the second challenge, we propose a dual regression compression (DRC) method to reduce model redundancy in both layer-level and channel-level based on channel pruning. Specifically, we first develop a channel number search method that minimizes the dual regression loss to determine the redundancy of each layer. Given the searched channel numbers, we further exploit the dual regression manner to evaluate the importance of channels and prune the redundant ones. Extensive experiments show the effectiveness of our method in obtaining accurate and efficient SR models.

Model Parameter Prediction Method for Accelerating Distributed DNN Training

As the size of deep neural network (DNN) models and datasets increases, distributed training becomes popular to reduce the training time. However, a severe communication bottleneck in distributed training limits its scalability. Many methods aim to solve this communication bottleneck by reducing communication traffic, such as gradient sparsification and quantification. However, these methods either are at the expense of losing model accuracy or introducing lots of computing overhead. We observe that the data distribution between layers of neural network models is similar. Thus, we propose a model parameter prediction method (MP2) to accelerate distributed DNN training under parameter server (PS) framework, where workers push only a subset of model parameters to PS, and residual model parameters are locally predicted by an already-trained deep neural network model on PS. Some key challenges are addressed. First, we build a hierarchical parameters dataset by randomly sampling a subset of model from normal distributed trainings. Second, we build a neural network model with the structure of “convolution + channel attention + Max pooling” for predicting model parameters by using a prediction result-based evaluation method. For VGGNet, ResNet, and AlexNet models on CIFAR10 and CIFAR100 datasets, compared with Baseline, Top-k, deep gradient compression (DGC), and weight nowcaster network (WNN), MP2 can reduce traffic by up to 88.98%; and accelerates the training by up to 47.32% while not losing the model accuracy. MP2 has shown good generalization.

Transfer learning based leaf disease detection model using convolution neural network

The plants are attacked from various micro-organisms, bacterial illnesses, and pests. The signs are normally identified via leaves, stem, or fruit inspection. Illnesses that generally appeared on vegetation are from leaves and causes big harm if not managed in the early ranges. To stop this huge harm and manipulate the unfold of disorder this work implements a software system. This research work customs deep neural network to gain knowledge of probable illnesses on leaves within the early phases so it can be stopped early. Deep neural network (DNN) used for image classification. This work mainly focuses a neural network model of leaves ailment detection. The commonly available plant leaves dataset is undertaken with a dataset having special training of disease detection. In this work VGG16, ResNet50, Inception V3 and Inception ResNetV2 architectural techniques are implemented to generate and compare the results. Results are compared on the factors like precision, accuracy, recall and F1-Score. The results lead to the conclusion, that the convolution neural network (CNN) is more impactful technique to perceive and predict plant diseases.

G2LCPS: End-to-end semi-supervised landmark prediction with global-to-local cross pseudo supervision for airway difficulty assessment

Difficult tracheal intubation is a major cause of anesthesia-related injuries, including brain damage and death. While deep neural networks have improved difficult airways (DA) predictions over traditional assessment methods, existing models are often black boxes, making them difficult to trust in critical medical settings. Traditional DA assessment relies on facial and neck features, but detecting neck landmarks is particularly challenging. This paper introduces a novel semi-supervised method for landmark prediction, namely G2LCPS, which leverages hierarchical filters and cross-supervised signals. The novelty lies in ensuring that the networks select good unlabeled samples at the image level and generate high-quality pseudo heatmaps at the pixel level for cross-pseudo supervision. The extended versions of the public AFLW, CFP, CPLFW and CASIA-3D FaceV1 face datasets and show that G2LCPS achieves superior performance compared to other state-of-the-art semi-supervised methods, achieving the lowest normalized mean error (NME) of 3.588 when only 1/8 of data is labelled. Notably, the inclusion of the local filter improved the prediction by at least 0.199 NME, whereas the global filter contributed an additional improvement of at least 0.216 NME. These findings underscore the effectiveness of our approach, particularly in scenarios with limited labeled data, and suggest that G2LCPS can significantly enhance the reliability and accuracy of DA predictions in clinical practice. The results highlight the potential of our method to improve patient safety by providing more trustworthy and precise predictions for difficult airway management.

3D visualization diagnostics for lung cancer detection

<span>Lung cancer is a leading cause of cancer deaths worldwide with an estimated 2 million new cases and 1·76 million deaths yearly. Early detection can improve survival, and CT scans are a precise imaging technique to diagnose lung cancer. However, analyzing hundreds of 2D CT slices is challenging and can cause false alarms. 3D visualization of lung nodules can aid clinicians in detection and diagnosis. The MobileNet model integrates multi-view and multi-scale nodule features using depthwise separable convolutional layers. These layers split standard convolutions into depthwise and pointwise convolutions to reduce computational cost. Finally, the 3D pulmonary nodular models were created using a ray-casting volume rendering approach. Compared to other state-of-the-art deep neural networks, this factorization enables MobileNet to achieve a much lower computational cost while maintaining a decent degree of accuracy. The proposed approach was tested on an LIDC dataset of 986 nodules. Experiment findings reveal that MobileNet provides exceptional segmentation performance on the LIDC dataset, with an accuracy of 93.3%. The study demonstrates that the MobileNet detects and segments lung nodules somewhat better than other older technologies. As a result, the proposed system proposes an automated 3D lung cancer tumor visualization.</span>

Reliable Prediction of Solar Photovoltaic Power and Module Efficiency using Bayesian Surrogate Assisted Explainable Data-Driven Model

This study proposes a Bayesian surrogate-driven explainable deep neural network model to predict and interpret the module efficiency and maximum output power of three commercially available photovoltaic modules: monocrystalline silicon, polycrystalline silicon, and amorphous silicon during the winter season. In addition, the influence of the photovoltaic material and ambient conditions on the predicted power and module efficiency is investigated. The model inputs include solar irradiance, module material (monocrystalline, polycrystalline, and amorphous silicon), season of the year (months), and time of day (morning to evening). Experiments were conducted on these three modules during the winter using data from Rawalpindi, Pakistan. These data were then fed into the deep neural network model to train and yield predictions. Several preprocessing techniques such as logarithmic transformation, square root, cube root, reciprocal, and exponential transformations are employed to improve the linearity of distributed data. For hyper-parameters optimization, Gaussian process, gradient boost regression trees, and random forest are used. The results show that the optimal deep neural network using random forest surrogate model along with square root and exponential transformation is able to predict the maximum power output and module efficiency of the considered photovoltaic modules for the entire investigated period with a correlation coefficient, R2 = 0.998. The least accurate model (R2=0.991) is a Gaussian process using simple data distribution. These results hold valuable insights for predicting and optimizing the performance of other solar photovoltaic modules in various climates and different seasons of the year.

Deep reinforcement learning for autonomous SideLink radio resource management in platoon-based C-V2X networks: An overview

Dynamic and autonomous SideLink (SL) Radio Resource Management (RRM) is essential for platoon-based cellular vehicular networks. However, this task is challenging due to several factors. These include the limited spectrum below 6 GHz, stringent vehicle-to-everything (V2X) communications requirements, uncertain and dynamic environments, limited vehicle sensing capabilities, and inherent distributed operation. These limitations often lead to resource collisions, data packet loss, and increased latency. Current standardized approaches in Long-Term Evolution-V2X (LTE-V2X) and New Radio-V2X (NR-V2X) rely on random resource selection, limiting their efficiency. Moreover, RRM is inherently a complex combinatorial optimization problem. It may involve conflicting objectives and constraints, making traditional approaches inadequate. Platoon-based communication necessitates careful resource allocation to support a diverse mix of communication types. These include safety-critical control messaging within platoons, less time-sensitive traffic management information between platoons, and even infotainment services like media streaming. Optimizing resource sharing inter- and intra- platoons is crucial to avoid excessive interference and ensure overall network performance. Deep Reinforcement Learning (DRL), combining Deep Learning (DL) and Reinforcement Learning (RL), has recently been investigated for network resource management. It offers a potential solution for these challenges. A DRL agent, represented by deep neural networks, interacts with the environment and learns optimal decision-making through trial and error. This paper overviews proposed DRL-based methods for autonomous SL RRM in single and multi-agent platoon-based C-V2X networks. It considers both intra- and inter-platoon communications with their specific requirements. We discuss the components of Markov Decision Processes (MDP) used to model the sequential decision-making of RRM. We then detail the DRL algorithms, training paradigms, and insights on the achieved results. Finally, we highlight challenges in existing works and suggest strategies for addressing them.