Application Of Deep Neural Networks Research Articles

Survey on Unsupervised Domain Adaptation for Semantic Segmentation for Visual Perception in Automated Driving

Deep neural networks (DNNs) have proven their capabilities in the past years and play a significant role in environment perception for the challenging application of automated driving. They are employed for tasks such as detection, semantic segmentation, and sensor fusion. Despite tremendous research efforts, several issues still need to be addressed that limit the applicability of DNNs in automated driving. The bad generalization of DNNs to unseen domains is a major problem on the way to a safe, large-scale application, because manual annotation of new domains is costly, particularly for semantic segmentation. For this reason, methods are required to adapt DNNs to new domains without labeling effort. This task is termed unsupervised domain adaptation (UDA). While several different domain shifts challenge DNNs, the shift between synthetic and real data is of particular importance for automated driving, as it allows the use of simulation environments for DNN training. We present an overview of the current state of the art in this research field. We categorize and explain the different approaches for UDA. The number of considered publications is larger than any other survey on this topic. We also go far beyond the description of the UDA state-of-the-art, as we present a quantitative comparison of approaches and point out the latest trends in this field. We conduct a critical analysis of the state-of-the-art and highlight promising future research directions. With this survey, we aim to facilitate UDA research further and encourage scientists to exploit novel research directions.

Deep Neural Networks for Enhancing Conversational AI in Multilingual India

This paper investigates the application of Deep Neural Networks (DNNs) in developing conversational AI systems tailored for India's multilingual population. The study evaluates various neural network architectures, including Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN), to enhance speech recognition and natural language understanding in Indian languages. Challenges like language diversity, dialect variation, and low-resource languages are addressed by introducing new training datasets and multilingual models. The paper also examines how conversational AI systems can be integrated into customer service, education, and healthcare sectors in India, with a focus on inclusivity and accessibility.

Black-box error diagnosis in Deep Neural Networks for computer vision: a survey of tools

The application of Deep Neural Networks (DNNs) to a broad variety of tasks demands methods for coping with the complex and opaque nature of these architectures. When a gold standard is available, performance assessment treats the DNN as a black box and computes standard metrics based on the comparison of the predictions with the ground truth. A deeper understanding of performances requires going beyond such evaluation metrics to diagnose the model behavior and the prediction errors. This goal can be pursued in two complementary ways. On one side, model interpretation techniques "open the box" and assess the relationship between the input, the inner layers and the output, so as to identify the architecture modules most likely to cause the performance loss. On the other hand, black-box error diagnosis techniques study the correlation between the model response and some properties of the input not used for training, so as to identify the features of the inputs that make the model fail. Both approaches give hints on how to improve the architecture and/or the training process. This paper focuses on the application of DNNs to Computer Vision (CV) tasks and presents a survey of the tools that support the black-box performance diagnosis paradigm. It illustrates the features and gaps of the current proposals, discusses the relevant research directions and provides a brief overview of the diagnosis tools in sectors other than CV.

Application of deep learning neural networks for the analysis of fluid-particle dynamics in fibrous filters

A novel hybrid data-driven framework was introduced in this study for the prediction of fluid-particle dynamics of submicron particles under hydrodynamic and Brownian forces in fibrous filters. The distribution of flow parameters was predicted by a high-resolution flow surrogate model based on a deep convolutional neural network (CNN) coupled with a Lagrangian particle tracking method based on the discrete phase model (DPM). The CNN-DPM model was applied to a group of several nonuniformly distributed fibers to calculate filtration efficiency. For CNN training, ground-truth data were obtained using a two-dimensional computational fluid dynamics (CFD) model for laminar incompressible flow. A simulation speedup of three orders of magnitude was obtained by CNN for predicting the velocity components and static pressure distribution within an error range of ±10%. The generalizability of CNN model was also confirmed for overlapped irregular fiber shapes and porosities out of the range of the training dataset with deviations in a range of ±20% in comparison with CFD results. The collection efficiency by CNN-DPM for small particle diameters with dp≤0.3μm was overestimated by approximately 3 % while agreeing well with CFD-DPM for larger particles. The errors of flow velocity predictions by CNN near the fibers added more randomness to the particle motion which was combined with the random Brownian motion of particles, consequently increasing the probability of the small particle hit number and filtration efficiency. The overall acceptable accuracy of the CNN-DPM was confirmed, which is suitable for fibrous filter design and optimization purposes that require thousands of simulations.

Convolutional neural networks (CNNs)-based multi-category damage detection and recognition of high-speed rail (HSR) reinforced concrete (RC) bridges using test images

The fast networking of high-speed rail (HSR) may cause in-service fatigue and ultimate load damage to bridges. This paper investigates the application of deep convolutional neural networks (CNNs) for multi-category damage image classification recognition of HSR-reinforced concrete (RC) bridges. The present study establishes a deep learning (DL) system based on a large amount of HSR bridge test data. When to begin, the damage done to HSR bridge piers may be broken down into three primary categories: concrete cracks, concrete spalling, and reinforcement exposure. These categories are determined by the statistics of HSR bridge pier testing. Secondly, in order to develop an automated recognition model for the damage of HSR piers, AlexNet CNNs were taught a transfer learning approach and then used to train themselves. The correct recognition rate of the three damaged pictures in the actual application of the model is 86% for cracks, 82% for reinforcement exposure, and 70% for concrete spalling, all of which have good recognition rates. The study's accuracy and precision enhance detection efficiency and may be utilized to identify HSR pier deterioration quickly. The research comprises a random selection from the training and validation sets. It also assesses the training model's generalization to out-of-sample pictures for engineering applications. This aspect of the work sets it apart from previous research in the same study area.

Methodology for the Application of Deep Neural Networks in Searches for New Physics at Colliders and Statistical Interpretation of Expected Results

Methodology for the Application of Deep Neural Networks in Searches for New Physics at Colliders and Statistical Interpretation of Expected Results

Functional Criticality Analysis of Structural Faults in AI Accelerators

The ubiquitous application of deep neural networks (DNNs) has led to a rise in demand for artificial intelligence (AI) accelerators. For example, the tensor processing unit from Google–based on a systolic array–and its variants are of considerable interest for DNN inferencing using AI accelerators. This article studies the problem of classifying structural faults in such an accelerator based on their functional criticality. We first analyze pin-level faults in the processing elements (PEs) of a systolic array. Simulation results for the LeNet network with 8-bit fixed-point, 16-bit floating-point (FP), and 32-bit FP data paths applied to the MNIST dataset show that over 93% of the pin-level structural faults in a PE are functionally benign. We present a greedy iterative framework for determining the criticality of stuck-at faults in a PE netlist and analyze the limitations of criticality analysis methods based on repeated fault simulations. We next present a scalable two-tier machine-learning (ML)-based method to assess the functional criticality of stuck-at faults in a computationally efficient manner. We address the problem of minimizing misclassification by utilizing generative adversarial networks (GANs). Two-tier ML/GAN-based criticality assessment leads to less than 1% test escapes during functional criticality evaluation of structural faults.

Prediction of causative genes in inherited retinal disorders from spectral‐domain optical coherence tomography utilizing deep learning techniques

AbstractTechnological advancement has markedly increased the medical information required for clinical management. Processing medical information automatically and comprehensively through the application of artificial intelligence (AI) is becoming essential to improve the quality of medical data management and reduce medical expenses.Deep learning techniques have been successfully applied in medical fields, especially in analysing clinical images for the predicted diagnosis of common ocular disorders, such as diabetic retinopathy and age‐related macular degeneration. In contrast, due to limited clinical resources, AI‐oriented bioinformatic engineering has not yet been applied in a clinic for orphan diseases, such as inherited retinal disease (IRD), although IRD is the most prevalent cause of blindness.Therefore, the purpose of the current study was to investigate the utility of deep learning in a nationwide cohort of patients with IRD, which enables to explore the typical phenotypes of IRD based on unique findings from clinical images and predict the genetic diagnosis.Clinical and genetic data with IRD were reviewed. Three categories of genetic diagnosis were selected based on the high prevalence of their causative genes: Stargardt disease (ABCA4), retinitis pigmentosa (EYS), and occult macular dystrophy (RP1L1). Optical coherence tomographic (OCT) images were cropped in a standardized manner with a macro algorithm. Algorithms for pipeline analyses, based on TensorFlow Inception V‐3 were determined with learning parameters through evaluation with a randomized 4‐fold cross‐validation method, given the accuracy of concordance (aimed >80%) between the genetic diagnosis and the machine diagnosis (ABCA4, EYS, RP1L1, and normal).A total of 178 images from 75 subjects were examined. The mean training accuracy was 98.5%, ranging from 90.6 to 100.0. The mean overall test accuracy was 90.9% (82.0–97.6).This study highlighted a novel application of deep neural networks in the prediction of the causative gene in IRD retinopathies from SD‐OCT, with a high prediction accuracy. The obtained study results have been registered in the database to support clinical/genetic diagnosis. These achievements will extensively promote the improvement of medical care quality by providing accurate real‐time diagnoses, reducing cost by avoiding unnecessary referrals and testing, enriching the education of non‐specialists, and realizing an application of personalized medicine.

Layerwise Security Protection for Deep Neural Networks in Industrial Cyber Physical Systems

Although deep neural networks (DNNs) have been increasingly applied in industrial cyber physical systems (ICPSs), they are vulnerable to security attacks due to the tight interaction between cyber elements and physical elements. In this article, we aim to protect the core IP of DNNs, i.e., the model weights, against security attacks. Different from conventional approaches, a layerwise protection framework is proposed to ensure the confidentiality of DNN model weights during the inference procedure such that the security quality is maximized, while satisfying the latency constraint of the DNN task. Based on the layerwise execution characteristics of DNN tasks, the encrypted layer-related weights are decrypted and fed to the next layer of DNN in plaintext. CPU-field programmable gate array (FPGA) coscheduling is considered to accelerate the execution of confidentiality protection, where CPU is utilized to conduct the decryption of weights and FPGA is used to perform the layer execution of DNN. Considering to provide optimal confidential protection for each layer, the problem is transformed into a quality of security maximization problem subject to layerwise execution constraint and deadline constraint of the DNN application. Due to the problem being NP-hard, a fast approximation algorithm is proposed to obtain the near-optimal solution under given real-time and security constraints. Extensive experiments and a real-life ICPS application evaluate the efficiency of the proposed techniques.

Application of Deep Neural Network to Discriminating Chalcopyrite Deposits, Ore Types, and Textures

Application of Deep Neural Network to Discriminating Chalcopyrite Deposits, Ore Types, and Textures

A method to measure non-Newtonian fluids viscosity using inertial viscometer with a computer vision system

The theory of rheology of non-Newtonian fluids is based on the generalized Newtonian hypothesis of viscosity. The viscometers for non-Newtonian fluids should implement fluid flows with the known stress and strain state parameters distributions. Ideally, the distributions should be homogeneous in the flow domain. The idea of the proposed method is based on a combination of a capillary and a rotational viscometers implemented in the torus-shaped capillary viscometer. Analysis of the mathematical model of the inertial non-Newtonian fluid flow in the torus allowed to determine the conditions of homogeneity of the mechanical and thermal parameters in the flow domain and to develop method of viscosity measurement. The measured values are the shear rate on the inner surface of the capillary and the flow rate. The measurements are implemented with the computer vision system that processes data obtained from the high speed CMOS camera that records inertial flow in the transparent capillary illuminated with laser. The computer vision system is based on the application of deep convolutional neural network for laser speckle contrast imaging processing. During the experiments, the proposed viscometer was compared with the Brookfield rotational viscometer. The relative error of the proposed viscometer and method is less than 2%. The inertial viscometer is compact, it allows to study the wide range of shear rates per one test in automatic mode, and it has low fluid capacity of approximately 1.87 ml. That makes it possible to use the viscometer as a point on care testing device in medicine to study the rheology of physiological fluids, in particular blood.

Training Compact DNNs with [formula omitted] Regularization

Deep neural network(DNN) has achieved unprecedented success in many fields. However, its large model parameters which bring a great burden on storage and calculation hinder the development and application of DNNs. It is worthy of compressing the model to reduce the complexity of the DNN. Sparsity-inducing regularizer is one of the most common tools for compression. In this paper, we propose utilizing the ℓ1/2 quasi-norm to zero out weights of neural networks and compressing the networks automatically during the learning process. To our knowledge, it is the first work applying the non-Lipschitz continuous regularizer for the compression of DNNs. The resulting sparse optimization problem is solved by stochastic proximal gradient algorithm. For further convenience of calculation, an approximation of the threshold-form solution to the proximal operator with ℓ1/2 is given at the same time. Extensive experiments with various datasets and baselines demonstrate the advantages of our new method.

Design and Analysis of Multipliers for DNN application using approximate 4:2 Compressors

The demand for Deep Learning applications in resource constrained devices is booming in recent years. Theuse of Deep Neural Network (DNN) is the leading method in these applications which has error resilient nature.This allows the use of Approximate Computing for efficient computation to leverage efficiency accuracy tradeoff by replacing Approximate Multiplier in place of exact multipliers. In this paper we proposed ApproximateCompressors and compared them in different cases of 8 bit integer dadda multipliers in terms of the Error metricsand accuracy in real life object classification applications. The approximate multipliers are designed using differentcompressors and used to perform multiplication in ResNet. We have proposed two approximate compressorsdesigns Design 1 and Design 2.The proposed 4:2 compressors design shows the more correct outputs and lessWorst Case Relative Error (WCRE) in the range of 2-16. Our proposed 4:2 compressor Design1 is utilized in themodified Reduction circuitry of dadda multiplier and shows the accuracy of 81.6 % for DNN application

Image-Based Approach to Intrusion Detection in Cyber-Physical Objects

Recently, approaches based on the transformation of tabular data into images have gained a lot of scientific attention. This is explained by the fact that convolutional neural networks (CNNs) have shown good results in computer vision and other image-based classification tasks. Transformation of features without spatial relations to images allows the application of deep neural networks to a wide range of analysis tasks. This paper analyzes existing approaches to feature transformation based on the conversion of the features of network traffic into images and discusses their advantages and disadvantages. The authors also propose an approach to the transformation of raw network packets into images and analyze its efficiency in the task of network attack detection in a cyber-physical object, including its robustness to novel and unseen attacks.

Modelling of Flow-Induced Vibration of Bluff Bodies: A Comprehensive Survey and Future Prospects

A comprehensive review of modelling techniques for the flow-induced vibration (FIV) of bluff bodies is presented. This phenomenology involves bidirectional fluid–structure interaction (FSI) coupled with non-linear dynamics. In addition to experimental investigations of this phenomenon in wind tunnels and water channels, a number of modelling methodologies have become important in the study of various aspects of the FIV response of bluff bodies. This paper reviews three different approaches for the modelling of FIV phenomenology. Firstly, we consider the mathematical (semi-analytical) modelling of various types of FIV responses: namely, vortex-induced vibration (VIV), galloping, and combined VIV-galloping. Secondly, the conventional numerical modelling of FIV phenomenology involving various computational fluid dynamics (CFD) methodologies is described, namely: direct numerical simulation (DNS), large-eddy simulation (LES), detached-eddy simulation (DES), and Reynolds-averaged Navier–Stokes (RANS) modelling. Emergent machine learning (ML) approaches based on the data-driven methods to model FIV phenomenology are also reviewed (e.g., reduced-order modelling and application of deep neural networks). Following on from this survey of different modelling approaches to address the FIV problem, the application of these approaches to a fluid energy harvesting problem is described in order to highlight these various modelling techniques for the prediction of FIV phenomenon for this problem. Finally, the critical challenges and future directions for conventional and data-driven approaches are discussed. So, in summary, we review the key prevailing trends in the modelling and prediction of the full spectrum of FIV phenomena (e.g., VIV, galloping, VIV-galloping), provide a discussion of the current state of the field, present the current capabilities and limitations and recommend future work to address these limitations (knowledge gaps).

MR-self Noise2Noise: self-supervised deep learning-based image quality improvement of submillimeter resolution 3D MR images.

The study aimed to develop a deep neural network (DNN)-based noise reduction and image quality improvement by only using routine clinical scans and evaluate its performance in 3D high-resolution MRI. This retrospective study included T1-weighted magnetization-prepared rapid gradient-echo (MP-RAGE) images from 185 clinical scans: 135 for DNN training, 11 for DNN validation, 20 for qualitative evaluation, and 19 for quantitative evaluation. Additionally, 18 vessel wall imaging (VWI) data were included to evaluate generalization. In each scan of the DNN training set, two noise-independent images were generated from the k-space data, resulting in an input-label pair. 2.5D U-net architecture was utilized for the DNN model. Qualitative evaluation between conventional MP-RAGE and DNN-based MP-RAGE was performed by two radiologists in image quality, fine structure delineation, and lesion conspicuity. Quantitative evaluation was performed with full sampled data as a reference by measuring quantitative error metrics and volumetry at 7 different simulated noise levels. DNN application on VWI was evaluated by two radiologists in image quality. Our DNN-based MP-RAGE outperformed conventional MP-RAGE in all image quality parameters (average scores = 3.7 vs. 4.9, p < 0.001). In the quantitative evaluation, DNN showed better error metrics (p < 0.001) and comparable (p > 0.09) or better (p < 0.02) volumetry results than conventional MP-RAGE. DNN application to VWI also revealed improved image quality (3.5 vs. 4.6, p < 0.001). The proposed DNN model successfully denoises 3D MR image and improves its image quality by using routine clinical scans only. • Our deep learning framework successfully improved conventional 3D high-resolution MRI in all image quality parameters, fine structure delineation, and lesion conspicuity. • Compared to conventional MRI, the proposed deep neural network-based MRI revealed better quantitative error metrics and comparable or better volumetry results. • Deep neural network application to 3D MRI whose pulse sequences and parameters were different from the training data showed improvement in image quality, revealing the potential to generalize on various clinical MRI.

Toward the Automated Detection of Light Echoes in Synoptic Surveys: Considerations on the Application of Deep Convolutional Neural Networks

Light echoes (LEs) are the reflections of astrophysical transients off of interstellar dust. They are fascinating astronomical phenomena that enable studies of the scattering dust as well as of the original transients. LEs, however, are rare and extremely difficult to detect as they appear as faint, diffuse, time-evolving features. The detection of LEs still largely relies on human inspection of images, a method unfeasible in the era of large synoptic surveys. The Vera C. Rubin Observatory Legacy Survey of Space and Time (LSST) will generate an unprecedented amount of astronomical imaging data at high spatial resolution, exquisite image quality, and over tens of thousands of square degrees of sky: an ideal survey for LEs. However, the Rubin data processing pipelines are optimized for the detection of point sources and will entirely miss LEs. Over the past several years, artificial intelligence (AI) object-detection frameworks have achieved and surpassed real-time, human-level performance. In this work, we leverage a data set from the Asteroid Terrestrial-impact Last Alert System telescope to test a popular AI object-detection framework, You Only Look Once, or YOLO, developed by the computer-vision community, to demonstrate the potential of AI for the detection of LEs in astronomical images. We find that an AI framework can reach human-level performance even with a size- and quality-limited data set. We explore and highlight challenges, including class imbalance and label incompleteness, and road map the work required to build an end-to-end pipeline for the automated detection and study of LEs in high-throughput astronomical surveys.

Self-adaptive physics-informed neural networks

Physics-Informed Neural Networks (PINNs) have emerged recently as a promising application of deep neural networks to the numerical solution of nonlinear partial differential equations (PDEs). However, it has been recognized that adaptive procedures are needed to force the neural network to fit accurately the stubborn spots in the solution of “stiff” PDEs. In this paper, we propose a fundamentally new way to train PINNs adaptively, where the adaptation weights are fully trainable and applied to each training point individually, so the neural network learns autonomously which regions of the solution are difficult and is forced to focus on them. The self-adaptation weights specify a soft multiplicative soft attention mask, which is reminiscent of similar mechanisms used in computer vision. The basic idea behind these SA-PINNs is to make the weights increase as the corresponding losses increase, which is accomplished by training the network to simultaneously minimize the losses and maximize the weights. In addition, we show how to build a continuous map of self-adaptive weights using Gaussian Process regression, which allows the use of stochastic gradient descent in problems where conventional gradient descent is not enough to produce accurate solutions. Finally, we derive the Neural Tangent Kernel matrix for SA-PINNs and use it to obtain a heuristic understanding of the effect of the self-adaptive weights on the dynamics of training in the limiting case of infinitely-wide PINNs, which suggests that SA-PINNs work by producing a smooth equalization of the eigenvalues of the NTK matrix corresponding to the different loss terms. In numerical experiments with several linear and nonlinear benchmark problems, the SA-PINN outperformed other state-of-the-art PINN algorithm in L2 error, while using a smaller number of training epochs.

Transition of electrochemical measurement to machine learning in the perspective of two-dimensional materials

Two-dimensional materials (e.g. graphene, and transition metal dichalcogenides) have become ubiquitous in electrochemical contexts including energy storage, electrocatalyst, and ion-selective membranes. This is due to its superior electrochemical properties, specifically “capacitance”, which can be referred to the storage ions at the electrolyte/materials interfaces. Experimental work and computational chemistry were carried out in the past decade for solving and improving the understanding of two-dimensional materials; however, these techniques are relatively expensive, complex, and time-consuming. Therefore, we accentuate the future trend of two-dimensional material study with machine learning as the modest alternative. In this perspective, the intrinsic capacitance properties of the two dimension materials were described from an atomic level, explaining the heteroatom doping to a nanoscopic level, showing (basal vs edge capacitance). The studies also extended to the macroscopic level i.e., the flake size of the two-dimensional materials. We then shed more light on the applicability of machine learning coupled with the “fundamental measurement” for solving electrochemistry of two-dimensional materials. The shallow artificial neural network was demonstrated for the prediction of CV curves using the data from size-dependent graphene. In addition, the application of deep neural networks with complicated architecture has also been explored through the prediction of capacitance for heteroatom-doped graphene. This perspective provides a clear background and creates the connection between fundamental measurement and machine learning for understanding the capacitance properties of two-dimensional materials.

Floating Gate Transistor‐Based Accurate Digital In‐Memory Computing for Deep Neural Networks

To improve the computing speed and energy efficiency of deep neural network (DNN) applications, in‐memory computing with nonvolatile memory (NVM) is proposed to address the time‐consuming and energy‐hungry data shuttling issue. Herein, a digital in‐memory computing method for convolution computing, which holds the key to DNNs, is proposed. Based on the proposed method, a floating gate transistor‐based in‐memory computing chip for accurate convolution computing with high parallelism is created. The proposed digital in‐memory computing method can achieve the central processing unit (CPU)‐equivalent precision with the same neural network architecture and parameters, different from the analogue or digital–analogue‐mixed in‐memory computing techniques. Based on the fabricated floating gate transistor‐based in‐memory computing chip, a hardware LeNet‐5 neural network is built. The chip achieves 96.25% accuracy on the full Modified National Institute of Standards and Technology database, which is the same as the result computed by the CPU with the same neural network architecture and parameters.