Article
Computer Science, Theory & Methods
Phillip Allen Lane, Joshua Dennis Booth
Summary: This paper introduces a heterogeneous format called CSR-k based on CSR, which achieves high-performance SpMV execution on different devices by reordering and grouping rows into hierarchical structures. It outperforms Intel MKL, NVIDIA cuSPARSE, and Sandia National Laboratories' KokkosKernels for regular sparse matrices.
PARALLEL COMPUTING
(2023)
Article
Computer Science, Theory & Methods
Shruti Shivakumar, Jiajia Li, Ramakrishnan Kannan, Srinivas Aluru
Summary: Tensor-based methods have gained attention due to their prevalence in real-world applications. There is substantial literature on tensor representations and algorithms for decompositions. Many applications result in sparse and symmetric tensors, making them important for further study.
IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS
(2023)
Article
Computer Science, Software Engineering
Longyue Xing, Zhaoshun Wang, Zhezhao Ding, Genshen Chu, Lingyu Dong, Nan Xiao
Summary: The performance of sparse stiffness matrix-vector multiplication is crucial for large-scale structural mechanics numerical simulation. This article introduces a new CSR-vector row algorithm that achieves fine-grained computing optimization for sparse stiffness matrices on AMD GPUs, demonstrating efficient reduce operations and deep memory access optimization, resulting in improved computing performance.
CONCURRENCY AND COMPUTATION-PRACTICE & EXPERIENCE
(2022)
Article
Mathematics, Applied
Jared Tanner, Simon Vary
Summary: Expressing a matrix as the sum of a low-rank matrix plus a sparse matrix is a flexible model for capturing global and local features in data. This model is widely used in robust principle component analysis and dynamic-foreground/static-background separation. Compressed sensing, matrix completion, and their variants have shown that data satisfying low complexity models can be efficiently recovered from a number of measurements proportional to the model complexity. This manuscript presents guarantees that matrices expressed as the sum of a rank-r matrix and a s-sparse matrix can be recovered by computationally tractable methods from a small number of linear measurements. Numerical experiments are provided to support the results. Evaluation: 7/10
APPLIED AND COMPUTATIONAL HARMONIC ANALYSIS
(2023)
Article
Physics, Multidisciplinary
Yue Wang, Linlin Xue, Yuqian Yan, Zhongpeng Wang
Summary: This paper investigates the influence of measurement matrices on the performance of image reconstruction in compressed sensing and proposes an improvement method. Experimental results show that the complex-valued Gaussian matrix after orthogonalization has better image reconstruction performance, and the sparse measurement matrix can effectively reduce the amount of calculation.
Article
Chemistry, Multidisciplinary
Shizhao Chen, Jianbin Fang, Chuanfu Xu, Zheng Wang
Summary: Optimizing sparse matrix-vector multiplication is challenging due to the non-uniform distribution of non-zero elements. This paper presents a new hybrid storage format and employs machine learning to automatically select the appropriate storage format based on the target matrix and hardware. Experimental results show significant performance improvement on different multi-core CPU platforms.
APPLIED SCIENCES-BASEL
(2022)
Article
Mathematics, Applied
Siham Boukhris, Artem Napov, Yvan Notay
Summary: We propose a new sparse matrix format for discretized partial differential equations with piecewise-constant coefficients. This format saves memory and is suitable for parallel computing on GPUs. It is well suited for algebraic multigrid methods and has been shown to outperform other solvers in terms of both run time and memory usage.
SIAM JOURNAL ON SCIENTIFIC COMPUTING
(2023)
Article
Automation & Control Systems
Qianru Jiang, Sheng Li, Liping Chang, Xiongxiong He, Rodrigo C. de Lamare
Summary: In this work, compressed sensing techniques based on prior knowledge are investigated for supporting telemedicine. The prior knowledge obtained by computing the probability of appearance of non-zero elements in each row of a sparse matrix is utilized in sensing matrix design and recovery algorithms. A robust sensing matrix is designed by jointly reducing the average mutual coherence and the projection of the sparse representation error. A Probability-Driven Normalized Iterative Hard Thresholding algorithm is developed as the recovery method, which exploits the prior knowledge and provides performance benefits. Simulations for synthetic data and endoscopy images of different organs demonstrate the superior performance of the proposed methods compared to previous algorithms.
JOURNAL OF THE FRANKLIN INSTITUTE-ENGINEERING AND APPLIED MATHEMATICS
(2022)
Article
Computer Science, Software Engineering
Hui Huang, Di Xiao, Jia Liang, Min Li
Summary: Compressed sensing (CS) is a popular signal processing technique, but its performance needs improvement for secure visual applications. In this study, we propose a prior-based measurement matrix design and sparse recovery algorithm for privacy-assured CS scheme in the cloud. The proposed scheme optimizes the measurement matrix, achieves privacy assurance, and improves sparse recovery performance.
Article
Computer Science, Theory & Methods
Jianhua Gao, Weixing Ji, Zhaonian Tan, Yizhuo Wang, Feng Shi
Summary: This article presents a new approach for compressed binary sparse matrix-vector multiplication, which reduces data transfer and improves computational performance through partitioning and encoding.
IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS
(2022)
Article
Computer Science, Information Systems
Ziheng Li, Zhanxuan Hu, Feiping Nie, Rong Wang, Xuelong Li
Summary: The paper presents a robust matrix completion method for incomplete data matrices with column outliers and sparse noise. Through iterative decomposition and re-weighted optimization, the accurate recovery of data is achieved.
INFORMATION SCIENCES
(2021)
Article
Computer Science, Information Systems
Dejian Li, Rongqiang Fang, Jing Wang, Dongyan Zhao, Ting Chong, Zengmin Ren, Jun Ma
Summary: In this study, a sparse neural network data partition and loop scheduling scheme is proposed. The memory access efficiency is improved by using compression algorithms, and the efficiency of neural network processing is enhanced through partition selection and loop scheduling.
Article
Mathematics, Applied
Hong Du, Huixian Lin
Summary: Compressed sensing (CS) technique allows for simultaneous sampling and compression, and the block CS (BCS) technique is an improved version that divides image signals into non-overlapping sub-blocks for separate processing. In this study, we propose an improved BCS (IBCS) method based on the Mallat reconstruction algorithm, which constructs a non-square sparse matrix to retain more original signal information. Our experiments show that IBCS achieves better reconstruction quality and lower implementation cost compared to traditional CBCS at lower sparsity levels.
COMPUTATIONAL & APPLIED MATHEMATICS
(2022)
Article
Computer Science, Theory & Methods
Elmira Karimi, Nicolas Bohm Agostini, Shi Dong, David Kaeli
Summary: This paper introduces a novel memory-aware format called VCSR, which out-performs previous formats on a GPU. VCSR achieves high thread-level parallelism and memory utilization by exploiting knowledge of GPU memory microarchitecture, reducing the number of global memory transactions, and providing a reordering mechanism. Experimental results demonstrate significant performance improvements of VCSR on different GPUs.
IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS
(2022)
Review
Neurosciences
Biao Sun, Wenfeng Zhao
Summary: This article provides a comprehensive survey of literature on compressed sensing of neurophysiology signals, discussing its applications, technical challenges, and prospects in neural signal transmission.
FRONTIERS IN NEUROSCIENCE
(2021)
Article
Biology
Stefan Engblom
BULLETIN OF MATHEMATICAL BIOLOGY
(2019)
Article
Mathematics, Applied
Augustin Chevallier, Stefan Engblom
SIAM JOURNAL ON NUMERICAL ANALYSIS
(2018)
Article
Mathematics, Applied
Doghonay Arjmand, Stefan Engblom, Gunilla Kreiss
JOURNAL OF COMPUTATIONAL AND APPLIED MATHEMATICS
(2019)
Article
Optics
Jing Liu, Gijs van der Schot, Stefan Engblom
Article
Computer Science, Interdisciplinary Applications
Jonatan Linden, Pavol Bauer, Stefan Engblom, Bengt Jonsson
ACM TRANSACTIONS ON MODELING AND COMPUTER SIMULATION
(2019)
Article
Optics
Jing Liu, Stefan Engblom, Carl Nettelblad
JOURNAL OF THE OPTICAL SOCIETY OF AMERICA A-OPTICS IMAGE SCIENCE AND VISION
(2020)
Article
Infectious Diseases
Stefan Engblom, Robin Eriksson, Stefan Widgren
Article
Physics, Mathematical
Jonathan Bull, Stefan Engblom
Summary: The study introduces a distributed implementation of an adaptive fast multipole method in three dimensions, utilizing a balanced type of adaptive space discretization. Complexity analysis shows favorable scaling properties, and numerical experiments confirm the scalability up to 512 cores and 1 billion source points. In-depth experiments on the algorithm parameters suggest that the overall implementation is well-suited for automated tuning.
COMMUNICATIONS IN COMPUTATIONAL PHYSICS
(2021)
Article
Computer Science, Artificial Intelligence
David Peer, Sebastian Stabinger, Stefan Engl, Antonio Rodriguez-Sanchez
Summary: In this paper, a method called Greedy-layer pruning is introduced to reduce the size of Transformer models. This method aims to outperform traditional layer pruning algorithms in terms of performance and achieve performance close to knowledge distillation, while allowing dynamic adjustment of model size without the need for additional pre-training phases.
PATTERN RECOGNITION LETTERS
(2022)
Article
Multidisciplinary Sciences
Beatrice Kennedy, Hugo Fitipaldi, Ulf Hammar, Marlena Maziarz, Neli Tsereteli, Nikolay Oskolkov, Georgios Varotsis, Camilla A. Franks, Diem Nguyen, Lampros Spiliopoulos, Hans-Olov Adami, Jonas Bjork, Stefan Engblom, Katja Fall, Anna Grimby-Ekman, Jan-Eric Litton, Mats Martinell, Anna Oudin, Torbjorn Sjostrom, Toomas Timpka, Carole H. Sudre, Mark S. Graham, Julien Lavigne du Cadet, Andrew T. Chan, Richard Davies, Sajaysurya Ganesh, Anna May, Sebastien Ourselin, Joan Capdevila Pujol, Somesh Selvachandran, Jonathan Wolf, Tim D. Spector, Claire J. Steves, Maria F. Gomez, Paul W. Franks, Tove Fall
Summary: The app-based COVID Symptom Study in Sweden utilized daily symptom reports to create a model for estimating the probability of symptomatic COVID-19 and predicting hospital admissions. The model showed lower prediction errors compared to a model based on case notifications and demonstrated transferability when applied to an English dataset.
NATURE COMMUNICATIONS
(2022)
Article
Mathematical & Computational Biology
Samuel Bronstein, Stefan Engblom, Robin Marin
Summary: This paper qualitatively investigates the convergence of Bayesian parameter inference in disease modeling. The study focuses on the Bayesian model's convergence with increasing amounts of data under measurement limitations. Depending on the informativeness of the disease measurements, a 'best case' and a 'worst case' analysis are provided. These cases consider direct accessibility to prevalence and binary signals corresponding to prevalence detection threshold, respectively. Numerical experiments are conducted to test the adaptability of the results in more realistic scenarios where analytical results are unavailable.
MATHEMATICAL BIOSCIENCES AND ENGINEERING
(2023)
Article
Infectious Diseases
Robin Marin, Hakan Runvik, Alexander Medvedev, Stefan Engblom
Summary: In this study, we design a data-driven compartment-based model of COVID-19 in Sweden to provide regional decision support for public healthcare. We derive parameter priors from national hospital statistics and develop linear filtering techniques for simulations based on daily healthcare demands. Our computational approach produces a Bayesian model of predictive value, giving important insights into the disease progression, including the effective reproduction number, infection fatality rate, and regional-level immunity. We validate our model against various sources, demonstrating its success. By relying on easily-collectible and non-sensitive data, our approach shows promise as a cost-effective tool for monitoring and decision-making in public health.
Proceedings Paper
Computer Science, Artificial Intelligence
Fredrik Wrede, Robin Eriksson, Richard Jiang, Linda Petzold, Stefan Engblom, Andreas Hellander, Prashant Singh
Summary: State-of-the-art neural network-based methods for learning summary statistics have shown promising results for simulation-based likelihood-free parameter inference. The proposed approach uses Bayesian neural networks to learn summary statistics and generate a proposal posterior density. An adaptive sampling scheme is employed to iteratively refine the predictive proposal posterior of the network, leading to more efficient and robust convergence in large prior spaces.
2022 INTERNATIONAL JOINT CONFERENCE ON NEURAL NETWORKS (IJCNN)
(2022)
Proceedings Paper
Automation & Control Systems
Hakan Runvik, Alexander Medvedev, Robin Eriksson, Stefan Engblom
Article
Computer Science, Interdisciplinary Applications
Stefan Widgren, Pavol Bauer, Robin Eriksson, Stefan Engblom
JOURNAL OF STATISTICAL SOFTWARE
(2019)