Review of deep learning: concepts, CNN architectures, challenges, applications, future directions

Journal Of Big Data, Journal Year: 2021, Volume and Issue: 8(1)

Published: March 31, 2021

In the last few years, deep learning (DL) computing paradigm has been deemed Gold Standard in machine (ML) community. Moreover, it gradually become most widely used computational approach field of ML, thus achieving outstanding results on several complex cognitive tasks, matching or even beating those provided by human performance. One benefits DL is ability to learn massive amounts data. The grown fast years and extensively successfully address a wide range traditional applications. More importantly, outperformed well-known ML techniques many domains, e.g., cybersecurity, natural language processing, bioinformatics, robotics control, medical information among others. Despite contributed works reviewing State-of-the-Art DL, all them only tackled one aspect which leads an overall lack knowledge about it. Therefore, this contribution, we propose using more holistic order provide suitable starting point from develop full understanding DL. Specifically, review attempts comprehensive survey important aspects including enhancements recently added field. particular, paper outlines importance presents types networks. It then convolutional neural networks (CNNs) utilized network type describes development CNNs architectures together with their main features, AlexNet closing High-Resolution (HR.Net). Finally, further present challenges suggested solutions help researchers understand existing research gaps. followed list major Computational tools FPGA, GPU, CPU are summarized along description influence ends evolution matrix, benchmark datasets, summary conclusion.

Language: Английский

---

Scientific Machine Learning Through Physics–Informed Neural Networks: Where we are and What’s Next

Journal of Scientific Computing, Journal Year: 2022, Volume and Issue: 92(3)

Published: July 26, 2022

Physics-Informed Neural Networks (PINN) are neural networks (NNs) that encode model equations, like Partial Differential Equations (PDE), as a component of the network itself. PINNs nowadays used to solve PDEs, fractional integral-differential and stochastic PDEs. This novel methodology has arisen multi-task learning framework in which NN must fit observed data while reducing PDE residual. article provides comprehensive review literature on PINNs: primary goal study was characterize these their related advantages disadvantages. The also attempts incorporate publications broader range collocation-based physics informed networks, stars form vanilla PINN, well many other variants, such physics-constrained (PCNN), variational hp-VPINN, conservative PINN (CPINN). indicates most research focused customizing through different activation functions, gradient optimization techniques, structures, loss function structures. Despite wide applications for have been used, by demonstrating ability be more feasible some contexts than classical numerical techniques Finite Element Method (FEM), advancements still possible, notably theoretical issues remain unresolved.

Language: Английский

---

Monitoring inland water quality using remote sensing: potential and limitations of spectral indices, bio-optical simulations, machine learning, and cloud computing

Earth-Science Reviews, Journal Year: 2020, Volume and Issue: 205, P. 103187 - 103187

Published: April 23, 2020

Language: Английский

---

A survey on deep learning tools dealing with data scarcity: definitions, challenges, solutions, tips, and applications

Journal Of Big Data, Journal Year: 2023, Volume and Issue: 10(1)

Published: April 14, 2023

Abstract Data scarcity is a major challenge when training deep learning (DL) models. DL demands large amount of data to achieve exceptional performance. Unfortunately, many applications have small or inadequate train frameworks. Usually, manual labeling needed provide labeled data, which typically involves human annotators with vast background knowledge. This annotation process costly, time-consuming, and error-prone. every framework fed by significant automatically learn representations. Ultimately, larger would generate better model its performance also application dependent. issue the main barrier for dismissing use DL. Having sufficient first step toward any successful trustworthy application. paper presents holistic survey on state-of-the-art techniques deal models overcome three challenges including small, imbalanced datasets, lack generalization. starts listing techniques. Next, types architectures are introduced. After that, solutions address listed, such as Transfer Learning (TL), Self-Supervised (SSL), Generative Adversarial Networks (GANs), Model Architecture (MA), Physics-Informed Neural Network (PINN), Deep Synthetic Minority Oversampling Technique (DeepSMOTE). Then, these were followed some related tips about acquisition prior purposes, well recommendations ensuring trustworthiness dataset. The ends list that suffer from scarcity, several alternatives proposed in order more each Electromagnetic Imaging (EMI), Civil Structural Health Monitoring, Medical imaging, Meteorology, Wireless Communications, Fluid Mechanics, Microelectromechanical system, Cybersecurity. To best authors’ knowledge, this review offers comprehensive overview strategies tackle

Language: Английский

---

A survey of deep learning techniques for weed detection from images

Computers and Electronics in Agriculture, Journal Year: 2021, Volume and Issue: 184, P. 106067 - 106067

Published: March 18, 2021

Language: Английский

---

Current Status, Challenges, and Possible Solutions of EEG-Based Brain-Computer Interface: A Comprehensive Review

Frontiers in Neurorobotics, Journal Year: 2020, Volume and Issue: 14

Published: June 3, 2020

Brain-Computer Interface (BCI), in essence, aims at controlling different assistive devices through the utilization of brain waves. It is worth noting that application BCI not limited to medical applications, and hence, research this field has gained due attention. Moreover, significant number related publications over past two decades further indicates consistent improvements breakthroughs have been made particular field. Nonetheless, it also mentioning with these improvements, new challenges are constantly discovered. This article provides a comprehensive review state-of-the-art complete system. First, brief overview electroencephalogram (EEG)-based systems given. Secondly, considerable popular applications reviewed terms electrophysiological control signals, feature extraction, classification algorithms, performance evaluation metrics. Finally, recent discussed, possible solutions mitigate issues recommended.

Language: Английский

---

Deep learning for motor imagery EEG-based classification: A review

Biomedical Signal Processing and Control, Journal Year: 2020, Volume and Issue: 63, P. 102172 - 102172

Published: Oct. 8, 2020

Language: Английский

---

Rational Design of Semiconductor‐Based Chemiresistors and their Libraries for Next‐Generation Artificial Olfaction

Advanced Materials, Journal Year: 2020, Volume and Issue: 32(51)

Published: Sept. 15, 2020

Abstract Artificial olfaction based on gas sensor arrays aims to substitute for, support, and surpass human olfaction. Like mammalian olfaction, a larger number of sensors more signal processing are crucial for strengthening artificial Due rapid progress in computing capabilities machine‐learning algorithms, on‐demand high‐performance that can eclipse becomes inevitable once diverse versatile sensing materials provided. Here, rational strategies design myriad different semiconductor‐based chemiresistors grow libraries enough identify wide range odors gases reviewed, discussed, suggested. Key approaches include the use p‐type oxide semiconductors, multinary perovskite spinel oxides, carbon‐based materials, metal chalcogenides, their heterostructures, as well heterocomposites distinctive utilization bilayer design, robust high‐throughput screening materials. In addition, state‐of‐the‐art key issues implementation electronic noses discussed. Finally, perspective chemiresistive next‐generation is

Language: Английский

---

Multivariate Temporal Convolutional Network: A Deep Neural Networks Approach for Multivariate Time Series Forecasting

Electronics, Journal Year: 2019, Volume and Issue: 8(8), P. 876 - 876

Published: Aug. 7, 2019

Multivariable time series prediction has been widely studied in power energy, aerology, meteorology, finance, transportation, etc. Traditional modeling methods have complex patterns and are inefficient to capture long-term multivariate dependencies of data for desired forecasting accuracy. To address such concerns, various deep learning models based on Recurrent Neural Network (RNN) Convolutional (CNN) proposed. improve the accuracy minimize dependence aperiodic data, this article, Beijing PM2.5 ISO-NE Dataset analyzed by a novel Multivariate Temporal Convolution (M-TCN) model. In model, multi-variable is constructed as sequence-to-sequence scenario non-periodic datasets. The multichannel residual blocks parallel with asymmetric structure convolution neural network results compared rich competitive algorithms long short term memory (LSTM), convolutional LSTM (ConvLSTM), (TCN) Attention LSTM-FCN (MALSTM-FCN), which indicate significant improvement accuracy, robust generalization our

Language: Английский

---

Machine Learning for Electrocatalyst and Photocatalyst Design and Discovery

Chemical Reviews, Journal Year: 2022, Volume and Issue: 122(16), P. 13478 - 13515

Published: July 21, 2022

Electrocatalysts and photocatalysts are key to a sustainable future, generating clean fuels, reducing the impact of global warming, providing solutions environmental pollution. Improved processes for catalyst design better understanding electro/photocatalytic essential improving effectiveness. Recent advances in data science artificial intelligence have great potential accelerate electrocatalysis photocatalysis research, particularly rapid exploration large materials chemistry spaces through machine learning. Here comprehensive introduction to, critical review of, learning techniques used research provided. Sources electro/photocatalyst current approaches representing these by mathematical features described, most commonly methods summarized, quality utility models evaluated. Illustrations how applied novel discovery elucidate electrocatalytic or photocatalytic reaction mechanisms The offers guide scientists on selection research. application catalysis represents paradigm shift way advanced, next-generation catalysts will be designed synthesized.

Language: Английский

---