Article
Computer Science, Information Systems
Zongqing Lu, Swati Rallapalli, Kevin Chan, Shiliang Pu, Thomas La Porta
Summary: This paper focuses on the resource requirements of Convolutional Neural Networks on mobile devices, measuring and analyzing performance and resource usage for different mobile CPUs and GPUs. The findings provide insights on optimizing CNN pipelines on mobile devices.
IEEE TRANSACTIONS ON MOBILE COMPUTING
(2021)
Article
Physics, Fluids & Plasmas
Elham Kiyani, Steven Silber, Mahdi Kooshkbaghi, Mikko Karttunen
Summary: This paper presents data-driven architectures based on machine learning algorithms for discovering nonlinear equations of motion for phase-field models. The experimental results show that we can effectively learn the time derivatives of the field and use the data-driven partial differential equations (PDEs) to propagate the field in time, achieving results in good agreement with the original PDEs.
Article
Chemistry, Multidisciplinary
Luca Mocerino, Andrea Calimera
Summary: The reduction in energy consumption is essential for the usability and reliability of deep neural networks, especially when deployed on devices with limited resources. By leveraging the redundancy in ConvNets, the proposed solution maximizes the reuse of arithmetic results during inference, leading to substantial energy savings with minimal loss in accuracy. The custom processing element integrating associative memory with floating-point unit achieves up to 77% in energy savings, showcasing the effectiveness of the approach.
APPLIED SCIENCES-BASEL
(2021)
Article
Computer Science, Artificial Intelligence
Nicola Rares Franco, Stefania Fresca, Andrea Manzoni, Paolo Zunino
Summary: Recently, deep Convolutional Neural Networks (CNNs) have achieved success in areas such as reduced order modeling of parametrized PDEs. However, there is still a lack of rigorous mathematical foundations in the available approaches. This paper derives rigorous error bounds for the approximation of nonlinear operators using CNN models, specifically addressing the case of mapping finite dimensional input to functional output. The resulting error estimates provide interpretation of the neural network architecture's hyperparameters, and the proofs reveal a deep connection between CNNs and the Fourier transform. Numerical experiments are also conducted to illustrate the application of the derived error bounds.
Article
Computer Science, Artificial Intelligence
Hojjat Rakhshani, Lhassane Idoumghar, Soheila Ghambari, Julien Lepagnot, Mathieu Brevilliers
Summary: The study focuses on the performance of deep learning models in dealing with global optimization problems, and analyzes it through empirical experiments on CEC 2017 benchmark suite and protein structure prediction (PSP) problems. The results reveal that the generated learning models can achieve competitive results given enough computational budget.
APPLIED SOFT COMPUTING
(2021)
Article
Ecology
Loris Nanni, Alessandro Manfe, Gianluca Maguolo, Alessandra Lumini, Sheryl Brahnam
Summary: This study utilized ensemble models of CNNs with different topologies and optimized Adam variants for pest identification. The effectiveness of the new proposed Adam algorithms was validated through experiments, achieving excellent performance on multiple datasets.
ECOLOGICAL INFORMATICS
(2022)
Article
Mathematics
Francisco Garcia Riesgo, Sergio Luis Suarez Gomez, Enrique Diez Alonso, Carlos Gonzalez-Gutierrez, Jesus Daniel Santos
Summary: This study utilized fully convolutional neural networks to address the complexities of solar Shack-Hartmann wavefront sensor correlations, comparing networks that use sensor images and correlations images as inputs. The results showed an improvement in phase recovery performance with the image-to-phase approach, achieving up to 93% similarity in recovering turbulence from high-altitude layers.
Article
Computer Science, Interdisciplinary Applications
Bao-Luo Li, Yu Qi, Jian-Sheng Fan, Yu-Fei Liu, Cheng Liu
Summary: Crack identification is crucial for preventive maintenance of asphalt pavement. This paper describes a fusion model based on the YOLO v5 that combines grid-based classification and box-based detection, achieving high accuracy and efficiency. The proposed NMS-ARS algorithm improves crack topology detection through postprocessing. Experimental results demonstrate the effective automatic crack identification for asphalt pavement.
COMPUTER-AIDED CIVIL AND INFRASTRUCTURE ENGINEERING
(2023)
Article
Radiology, Nuclear Medicine & Medical Imaging
Gian Marco Conte, Alexander D. Weston, David C. Vogelsang, Kenneth A. Philbrick, Jason C. Cai, Maurizio Barbera, Francesco Sanvito, Daniel H. Lachance, Robert B. Jenkins, W. Oliver Tobin, Jeanette E. Eckel-Passow, Bradley J. Erickson
Summary: Generative adversarial networks can synthesize brain MRI scans to replace missing sequences, enabling their use as inputs for deep learning models in brain lesion segmentation. The quality of the generated images was evaluated based on MSE and SSI, while the segmentation results were compared using DSC, confirming their effectiveness.
Article
Computer Science, Artificial Intelligence
Chao Li, Handing Wang, Jun Zhang, Wen Yao, Tingsong Jiang
Summary: This article discusses the adversarial attack problem faced by deep neural networks and the limitations of existing solutions. An approximated gradient sign method using differential evolution is proposed to solve the black-box adversarial attack problem. By transforming the pixel-based decision space into a dimension-reduced decision space and introducing different neighbor selection and optimization search strategies, multiple variants of the proposed method are developed. Experimental results demonstrate the superior performance of the method in solving black-box adversarial attack problems.
IEEE TRANSACTIONS ON EVOLUTIONARY COMPUTATION
(2022)
Article
Computer Science, Artificial Intelligence
Vidyanand Mishra, Lalit Kane
Summary: The paper proposes a genetic algorithm-based method for selecting a convolutional neural network architecture. By optimizing the encoding scheme, the initialization of the population, the generation of offspring, and the fitness function, the algorithm's efficiency and performance are improved. Experimental results on the MNIST, Fashion_MNIST, and CIFAR-10 datasets demonstrate that the method achieves comparable accuracy, convergence rate, and computation resources consumption to the best manual and automatic approaches.
EXPERT SYSTEMS WITH APPLICATIONS
(2023)
Article
Multidisciplinary Sciences
Pichatorn Suppakitjanusant, Somnuek Sungkanuparph, Thananya Wongsinin, Sirapong Virapongsiri, Nittaya Kasemkosin, Laor Chailurkit, Boonsong Ongphiphadhanakul
Summary: Recent breakthroughs in deep learning have allowed for the detection of subtle changes in voice features of COVID-19 patients post-recovery, with the model using polysyllabic sentences achieving the highest accuracy of 85%.
SCIENTIFIC REPORTS
(2021)
Article
Computer Science, Artificial Intelligence
Philipp Christian Petersen, Anna Sepliarskaia
Summary: This study investigates the generalization capacity of group convolutional neural networks and provides precise estimates of VC dimensions for certain simple sets of such networks. It is discovered that even with infinite groups and suitable convolutional kernels, two-parameter families of convolutional neural networks can have infinite VC dimensions, while remaining invariant to the action of an infinite group.
Article
Computer Science, Artificial Intelligence
Zi-Ming Wang, Meng-Han Li, Gui-Song Xia
Summary: The article introduces a new texture synthesis model called conditional generative ConvNet (cgCNN) model, which combines deep statistics and the probabilistic framework. The model learns the weights of ConvNets for each input exemplar instead of relying on pre-trained models, and can synthesize high quality dynamic, sound and image textures in a unified manner.
IEEE TRANSACTIONS ON IMAGE PROCESSING
(2021)
Article
Multidisciplinary Sciences
So Hyun Park, Young Jae Kim, Kwang Gi Kim, Jun-Won Chung, Hyun Cheol Kim, In Young Choi, Myung-Won You, Gi Pyo Lee, Jung Han Hwang
Summary: This study aimed to develop a CNN using the EfficientNet algorithm for automated classification of acute appendicitis, acute diverticulitis, and normal appendix on CT images, and evaluate its diagnostic performance. The results showed that the RGB serial image method had slightly higher sensitivity, accuracy, and specificity for classifying normal appendix and acute diverticulitis compared to the single image method. Additionally, the mean areas under the ROC curve were significantly higher with the RGB serial image method for all three conditions. Therefore, the model accurately distinguished these conditions on CT images, particularly when using the RGB serial image method.
Article
Engineering, Electrical & Electronic
Bruna G. Palm, Fabio M. Bayer, Renato J. Cintra
Summary: This paper introduces a new beta binomial autoregressive moving average model (BBARMA) for modeling quantized amplitude data and bounded count data, which estimates the conditional mean through a dynamic structure and provides parameter estimation, detection tools, forecasting scheme, and diagnostic measures. Simulation results show that the proposed model outperforms traditional detectors in sinusoidal signal detection.
DIGITAL SIGNAL PROCESSING
(2021)
Article
Statistics & Probability
Bruna Gregory Palm, Fabio M. Bayer, Renato J. Cintra
Summary: This article proposes five prediction intervals for the beta autoregressive moving average model and evaluates their performance through Monte Carlo simulations. The bias-corrected and acceleration (BCa) prediction interval shows the best performance among the evaluated intervals.
COMMUNICATIONS IN STATISTICS-SIMULATION AND COMPUTATION
(2023)
Article
Mathematics, Applied
Josimar M. Vasconcelos, Renato J. Cintra, Abraao D. C. Nascimento
Summary: In this paper, new qualitative and quantitative GoF tools for model selection within the beta-G class are provided by combining probability weighted moments (PWMs) and the Mellin transform (MT). PWMs for the Frechet and Kumaraswamy distributions are derived, along with expressions for the MT and log-cumulants (LC) of various beta distributions. LC diagrams are constructed and confidence ellipses for the LCs are derived based on the Hotelling's T-2 statistic. The proposed GoF measures are applied on five real data sets to demonstrate their applicability.
MATHEMATICAL METHODS IN THE APPLIED SCIENCES
(2021)
Article
Computer Science, Interdisciplinary Applications
Bruna G. Palm, Fabio M. Bayer, Renato J. Cintra
Summary: The application of a two-dimensional ARMA model tailored for Rayleigh-distributed data is studied. The model is derived and conditional likelihood inferences are discussed. Monte Carlo simulations and numerical experiments are conducted to evaluate the model's performance and results compared to traditional 2-D ARMA models and competing methods in the literature.
COMPUTATIONAL STATISTICS & DATA ANALYSIS
(2022)
Article
Biology
Vitor A. Coutinho, Renato J. Cintra, Fabio M. Bayer
Summary: The study introduces multiplierless DHT approximations for medical image compression, offering significant computational resource savings and suitable for devices with limited resources. These methods achieve the same level of visual quality while reducing computational effort. By implementing the proposed transforms on an ARM Cortex-M0+ processor, the execution time is significantly reduced.
COMPUTERS IN BIOLOGY AND MEDICINE
(2021)
Article
Geochemistry & Geophysics
Bruna G. Palm, Fabio M. Bayer, Renato J. Cintra
Summary: This letter introduces bias-adjusted estimators tailored for the Rayleigh regression model and presents numerical experiments on synthetic and actual SAR data sets. The results show that the bias-adjusted estimators yield nearly unbiased estimates and accurate modeling results.
IEEE GEOSCIENCE AND REMOTE SENSING LETTERS
(2022)
Article
Geochemistry & Geophysics
Bruna G. Palm, Fabio M. Bayer, Renato Machado, Mats Pettersson, Viet T. Vu, Renato J. Cintra
Summary: This article proposes a Rayleigh regression model based on robust estimation to address outliers in SAR data. Through numerical experiments, the robust estimator is found to outperform traditional estimators in corrupted signals, demonstrating better performance in the presence of anomalies.
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
(2022)
Article
Engineering, Electrical & Electronic
Anabeth P. Radunz, Thiago L. T. da Silveira, Fabio M. Bayer, Renato J. Cintra
Summary: This work proposes low-computational cost approximations for the KLT that are suitable for image and video compression. Extensive computational experiments on blocklengths of 4, 8, 16, and 32 demonstrate the effectiveness of these low-complexity transforms.
SIGNAL PROCESSING-IMAGE COMMUNICATION
(2022)
Article
Computer Science, Artificial Intelligence
Anabeth P. Radunz, Fabio M. Bayer, Renato J. Cintra
Summary: The paper introduces a new class of low-complexity transforms obtained through applying the round function to KLT matrix elements, which perform well in image compression with low implementation cost.
JOURNAL OF REAL-TIME IMAGE PROCESSING
(2022)
Article
Engineering, Electrical & Electronic
Thiago L. T. da Silveira, Diego Ramos Canterle, Diego F. G. Coelho, Vitor A. Coutinho, Fabio M. Bayer, Renato J. Cintra
Summary: The discrete cosine transform (DCT) is a relevant tool in signal processing applications, known for its good decorrelation properties. Recent research has focused on low-complexity approximations of the DCT, which are important for real-time computation and low-power consumption applications. This paper presents a new multiparametric transform class and its associated fast algorithm. By solving an optimization problem, four novel low-complexity DCT approximations are obtained. Experimental results show that these new transforms perform as well as or better than current state-of-the-art DCT approximations.
IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS FOR VIDEO TECHNOLOGY
(2022)
Article
Computer Science, Information Systems
Anabeth P. P. Radunz, Luan Portella, R. S. Oliveira, Fabio M. Bayer, Renato J. J. Cintra
Summary: In this paper, low-complexity approximations of the discrete cosine transform (DCT) with blocklengths of 16, 32, and 64 are introduced by minimizing the angle between the rows of the exact DCT matrix and the matrix induced by the approximate transforms. These proposed transforms outperform existing DCT approximations in terms of performance, as judged by classical figures of merit. Fast algorithms are also developed for these low-complexity transforms, striking a good balance between performance and computational cost. Experimental results in image encoding demonstrate the relevance of these transforms, showing superior performance compared to known approximations for blocklengths of 16, 32, and 64.
JOURNAL OF SIGNAL PROCESSING SYSTEMS FOR SIGNAL IMAGE AND VIDEO TECHNOLOGY
(2023)
Article
Computer Science, Information Systems
Haixiang Zhao, Arjuna Madanayake, Renato J. Cintra, Soumyajit Mandal
Summary: This paper presents a current-mode on-chip analog architecture for multi-beamforming of IF band signals from small- to moderate-sized antenna arrays. The architecture utilizes an approximate discrete Fourier transform (a-DFT) algorithm to achieve the required linear transformation for multi-beamforming. The proposed integrated multi-beamformer demonstrates lower power consumption and supports high-performance beamforming with a wide dynamic range.
Article
Engineering, Electrical & Electronic
Luan Portella, Diego F. G. Coelho, Fabio M. Bayer, Arjuna Madanayake, Renato J. Cintra
Summary: This article presents a variant of transform scaling method to obtain DFT approximations of large blocksize. The proposed method scales up a given transformation successively, leading to different sizes of transformations. The authors fully introduce a 32(4)-point DFT approximation and provide a fast algorithm, arithmetic complexity assessment, and error analysis.
IEEE SIGNAL PROCESSING LETTERS
(2022)
Article
Engineering, Electrical & Electronic
Diego F. G. Coelho, Renato J. Cintra, Arjuna Madanayake, Sirani M. Perera
Summary: This paper introduces a collection of scaling methods for generating 2N-point DCT-II approximations based on N-point low-complexity transformations. The proposed techniques outperform the transforms resulting from the JAM scaling method in terms of total error energy and mean squared error, as demonstrated by extensive error analysis based on statistical simulation. Additionally, a hardware implementation shows the competitiveness of the proposed methods compared to the JAM scaling method.
IEEE TRANSACTIONS ON SIGNAL PROCESSING
(2021)
Article
Mathematics, Applied
Josimar M. Vasconcelos, Renato J. Cintra, Abraao D. C. Nascimento, Leandro C. Rego
JOURNAL OF COMPUTATIONAL AND APPLIED MATHEMATICS
(2020)