Article
Food Science & Technology
Chao-Hui Feng, Hirofumi Arai, Francisco J. Rodriguez-Pulido
Summary: This study evaluated the pH values of sausages stuffed in natural hog casings with different modifications using hyperspectral imaging and response surface methodology. The results showed that soy lecithin and orange extracts interactively affected the pH of sausages. The first derivative was found to be the best spectral pretreatment method, and 12 feature wavelengths were selected for prediction.
Article
Statistics & Probability
Jie Zhou, Will Wei Sun, Jingfei Zhang, Lexin Li
Summary: In this article, we propose a regression model for partially observed dynamic tensors, characterized by low-rankness, sparsity, and fusion structures on the regression coefficient tensor. We develop an efficient nonconvex alternating updating algorithm and derive the error bounds of the estimators. Our approach differs significantly from existing tensor completion or tensor response regression solutions. It is illustrated using simulations and real applications in neuroimaging dementia study and digital advertising study.
JOURNAL OF THE AMERICAN STATISTICAL ASSOCIATION
(2023)
Article
Mathematical & Computational Biology
Yeonhee Park, Zhihua Su, Dongjun Chung
Summary: Partial least squares (PLS) regression is a superior alternative to ordinary least squares regression with demonstrated prediction performance. However, treating categorical variables as continuous in PLS regression may result in biased estimates and invalid inferences. This study proposes an envelope-based partial PLS estimator that considers the conditional distributions of the response and continuous predictors on categorical predictors, achieving more efficient estimation and better predictions. The method is applied to identify cytokine-based biomarkers for COVID-19 patients, revealing associations with clinical information.
STATISTICS IN MEDICINE
(2022)
Article
Neurosciences
Suprateek Kundu, Alec Reinhardt, Serena Song, Joo Han, M. Lawson Meadows, Bruce Crosson, Venkatagiri Krishnamurthy
Summary: This study proposes a novel Bayesian tensor response regression approach for longitudinal neuroimaging data. The method utilizes low-rank decomposition and joint credible regions for feature selection and more accurate inference. The advantages of the proposed approach over traditional voxel-wise regression are highlighted through extensive simulation studies and application to real longitudinal aphasia data.
HUMAN BRAIN MAPPING
(2023)
Article
Computer Science, Artificial Intelligence
Jiani Liu, Ce Zhu, Zhen Long, Huyan Huang, Yipeng Liu
Summary: This paper introduces a tensor ring ridge regression (TRRR) model with a low-rank tensor ring structure in the coefficient array, and develops two optimization models to solve the problem. Experimental results demonstrate the enhanced performance of the algorithm in tasks such as spatio-temporal forecasting and 3D reconstruction of human motion capture data, especially in terms of training time.
PATTERN RECOGNITION
(2021)
Article
Biology
Bo Wei, Limin Peng, Ying Guo, Amita Manatunga, Jennifer Stevens
Summary: Collecting neuroimaging data in the form of tensors has become common in mental health studies. The study introduces a tensor response quantile regression framework to analyze the association between neuroimaging phenotypes and clinical predictors, specifically focusing on post-traumatic stress disorder (PTSD) research.
Article
Computer Science, Artificial Intelligence
Zongwen Bai, Ying Li, Marcin Wozniak, Meili Zhou, Di Li
Summary: The study proposed a novel comprehensive solution to compress and accelerate Visual Question Answering systems. By applying various decomposition methods and regression strategies, the Fully Connected layers in Convolutional Neural Network and Long Short Term Memory were successfully compressed, achieving high compression ratios with minimal accuracy drop.
PATTERN RECOGNITION
(2021)
Article
Mathematics, Applied
Taoran Fu, Bo Jiang, Zhening Li
Summary: Hermitian matrices have been important in matrix theory and complex quadratic optimization. The high-order generalization, CPS tensors, have shown growing interest in tensor theory and computation, especially in application-driven complex polynomial optimization problems. This paper studies CPS tensors with a focus on ranks, computing rank-one decompositions and approximations, as well as their applications.
Article
Engineering, Electrical & Electronic
Arinbjorn Kolbeinsson, Jean Kossaifi, Yannis Panagakis, Adrian Bulat, Animashree Anandkumar, Ioanna Tzoulaki, Paul M. Matthews
Summary: CNNs achieve high performance with deep, over-parametrized neural architectures, but lack robustness and generalization abilities. Tensor layers provide better inductive biases through multi-linear structure, and tensor dropout is a technique to improve generalization for image classification and phenotypic trait prediction tasks. Superior performance and improved robustness are demonstrated in experiments, validating the theoretical validity and regularizing effect of tensor dropout.
IEEE JOURNAL OF SELECTED TOPICS IN SIGNAL PROCESSING
(2021)
Article
Statistics & Probability
Daiwei Zhang, Lexin Li, Chandra Sripada, Jian Kang
Summary: A novel non-parametric approach is proposed to delineate associations between images and covariates using deep neural networks in the framework of spatially varying coefficient models. The method incorporates spatial smoothness, handles subject heterogeneity, and provides straightforward interpretations. It is also highly flexible and accurate, making it ideal for capturing complex association patterns.
JOURNAL OF THE ROYAL STATISTICAL SOCIETY SERIES B-STATISTICAL METHODOLOGY
(2023)
Article
Astronomy & Astrophysics
Camille Bonvin, Levon Pogosian
Summary: Our current observations cannot distinguish between models with a fifth force acting on dark matter and those with modified laws of gravity. However, future gravitational redshift data may provide a solution. Alternative gravity theories typically include additional degrees of freedom that can mediate forces between matter particles. A fifth force in these theories can be detected through a gravitational slip, while a dark sector interaction does not cause such a slip. By measuring the effective gravitational slip through cosmological surveys, we can determine if dark matter is affected by a fifth force or modified gravity. Future observations of gravitational redshift can provide a direct measurement of time distortion and help resolve this question.
Article
Engineering, Industrial
Huihui Miao, Andi Wang, Bing Li, Jianjun Shi
Summary: The STOTI method proposes a structured tensor regression approach considering interaction effects to alleviate the curse of dimensionality and resolve computational challenges. By describing the specific structure of the main and interaction effect tensors using regularization terms based on prior knowledge, the method outperforms in estimation and prediction accuracy as demonstrated in extensive simulations and a real case study.
JOURNAL OF QUALITY TECHNOLOGY
(2022)
Article
Computer Science, Artificial Intelligence
Ian Convy, K. Birgitta Whaley
Summary: This paper proposes an interaction decomposition method to assess the relative importance of different regressors in tensor network regression models. The study finds that up to 75% of interaction degrees contribute meaningfully to the models. The paper also introduces a new type of tensor network model that is trained on a small subset of interaction degrees, and shows that it can match or outperform the full models using only a fraction of the exponential feature space.
MACHINE LEARNING-SCIENCE AND TECHNOLOGY
(2022)
Article
Computer Science, Artificial Intelligence
Zhe Cheng, Xiangjian Xu, Zihao Song, Weihua Zhao
Summary: This paper discusses the estimation algorithm of tensor response on vector covariate regression model and proposes three new algorithms. Theoretical analysis and empirical results demonstrate the advantages of our proposed algorithms in estimation accuracy and computing speed.
STATISTICAL ANALYSIS AND DATA MINING
(2023)
Article
Computer Science, Interdisciplinary Applications
Giuseppe Brandi, T. Di Matteo
Summary: This paper proposes a parsimonious tensor regression model that retains the intrinsic multidimensional structure of the dataset using Tucker structure and shrinkage penalization to address overfitting and collinearity, as well as developing an Alternating Least Squares algorithm. The model's performance and robustness are validated through simulation exercises and empirical analysis, demonstrating its superiority over benchmark models in forecasting literature.
JOURNAL OF COMPUTATIONAL SCIENCE
(2021)