Article
Automation & Control Systems
Songhua Wu, Tongliang Liu, Bo Han, Jun Yu, Gang Niu, Masashi Sugiyama
Summary: SU classification is a method that uses similar data pairs and unlabeled data to build classifiers, providing an alternative to supervised classifiers that require labeled data points. However, SU classification has limitations due to the possibility of respondents answering questions in a favorable manner instead of truthfully. This paper studies how to learn from noisy similar data pairs and unlabeled data, proposing an algorithm for nSU classification.
JOURNAL OF MACHINE LEARNING RESEARCH
(2022)
Article
Automation & Control Systems
Filippo Cacace, Alfredo Germani
Summary: This article studies the identification problem of linear systems from a set of noisy input-output trajectories. The problem is formulated and solved as a least-square regularized estimate on a suitable function space of finite-bandwidth operators. This abstract setting is well suited to represent a broad class of finite- and infinite-dimensional linear systems. We determine the value of the regularization parameter as a function of the amount of noise on the learning trajectories and we show how to obtain recursive and causal estimates for the case of linear dynamical systems.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL
(2023)
Article
Engineering, Multidisciplinary
Yasen Wang, Ye Yuan, Huazhen Fang, Han Ding
Summary: In modern science and engineering disciplines, data-driven methods are essential for system modeling. However, noisy data remains a significant obstacle. This study presents a method that can uncover linear dynamical systems from noisy data and successfully applies it to various simulation systems, demonstrating its potential for practical applications.
SCIENCE CHINA-TECHNOLOGICAL SCIENCES
(2023)
Article
Automation & Control Systems
Claudio De Persis, Pietro Tesi
Summary: This paper proposes a method to return a controller without estimating a model of the system, and provides sufficient conditions for returning a stabilizing controller when the data is affected by noise. The method has low complexity, requiring only a finite number of samples and can be efficiently implemented.
Article
Computer Science, Artificial Intelligence
Zhenhuang Cai, Guo-Sen Xie, Xingguo Huang, Dan Huang, Yazhou Yao, Zhenmin Tang
Summary: This paper proposes a simple yet effective method named MS-DeJOR for training robust models in the presence of web noise in deep neural networks. Unlike existing methods, MS-DeJOR decouples sample selection from the training procedure, uses a negative entropy term to prevent false positives from being overemphasized, and leverages accumulated predictions to refurbish noisy labels and re-weight training images for improved performance.
PATTERN RECOGNITION
(2023)
Article
Computer Science, Artificial Intelligence
Long Chen, Fei Wang, Ruijing Yang, Fei Xie, Wenjing Wang, Cai Xu, Wei Zhao, Ziyu Guan
Summary: This paper proposes a novel weakly-supervised anti-noise contrastive learning framework for sentiment classification, which learns robust representations through pre-training and fine-tuning, and demonstrates its superiority on multiple datasets.
INTERNATIONAL JOURNAL OF MACHINE LEARNING AND CYBERNETICS
(2022)
Article
Automation & Control Systems
Amr Alanwar, Anne Koch, Frank Allgoewer, Karl Henrik Johansson
Summary: This paper discusses how to compute reachable sets directly from noisy data without a given system model. Several reachability algorithms are presented for different types of systems generating the data. For linear systems, an algorithm based on matrix zonotopes is proposed, which computes over-approximated reachable sets. Constrained matrix zonotopes are introduced to provide less conservative reachable sets at the cost of increased computational expenses and incorporate prior knowledge about the unknown system model. The approach is also extended to polynomial and nonlinear systems with theoretical guarantees of proper over-approximation.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL
(2023)
Article
Computer Science, Interdisciplinary Applications
Pratyush Kumar, James B. Rawlings
Summary: This paper proposes a novel model-free Q-learning approach to estimate linear feedback controllers from noisy process data. The approach is modified to handle unknown noise covariances and is applied to estimate feedback controllers for linear systems with both process and measurement noise. A model-based approach is also presented for comparison.
COMPUTERS & CHEMICAL ENGINEERING
(2023)
Article
Mathematics, Applied
Bin Dong, Zuowei Shen, Jianbin Yang
Summary: This paper explores the approximation of functions from noisy and nonsmooth observed data, with a focus on sparse noise removal schemes. Theoretical analysis is presented, highlighting the importance of sparsity-based denoising for effective approximation. A new approximation scheme is proposed for large datasets to significantly reduce noise level and ensure asymptotic convergence.
SIAM JOURNAL ON NUMERICAL ANALYSIS
(2021)
Article
Computer Science, Artificial Intelligence
Jia Li, Yafei Song, Jianfeng Zhu, Lele Cheng, Ying Su, Lin Ye, Pengcheng Yuan, Shumin Han
Summary: This paper proposes a Ubiquitous Reweighting Network (URNet) for learning image classification models from noisy web data. By addressing five key challenges observed in web data, the approach outperforms state-of-the-art models in the WebVision 2018 challenge with 16 million noisy training images from 5000 classes.
IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE
(2021)
Article
Environmental Sciences
Yuyang Liu, Jiacheng Liu, Yubo Zhao, Xueji Wang, Shuyao Song, Hong Liu, Tao Yu
Summary: This study focuses on the inversion of water quality from unmanned airborne hyperspectral image and proposes a method to address the issue of noisy label data in flowing water bodies. The experimental results demonstrate that the noisy-label instance selection method greatly improves retrieval performance, especially on turbidity and chroma data.
Article
Geochemistry & Geophysics
Feng Wang, Bo Yang, Yuqing Wang, Ming Wang
Summary: This paper proposes an unsupervised denoising method based on model-based deep learning, which combines domain knowledge and a data-driven approach. This method reduces the dependency on labeled data and explores insights into the denoising system. Experimental results demonstrate that the proposed method achieves competitive performance compared to conventional methods.
IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING
(2022)
Article
Automation & Control Systems
Andrea Bisoffi, Claudio De Persis, Pietro Tesi
Summary: This work addresses the problem of designing a stabilizing controller for an unknown linear system using noisy data, considering an alternative disturbance model. Unlike existing methods, a simple argument based on the S-procedure can lead to a more effective controller design without the need for conversion steps, resulting in a larger feasible set and a smaller set of system matrices consistent with the data.
SYSTEMS & CONTROL LETTERS
(2021)
Article
Computer Science, Artificial Intelligence
Ning Xu, Jia-Yu Li, Yun-Peng Liu, Xin Geng
Summary: This article proposes a novel LE method named TALEN which recovers and refines label distribution guided by trusted data, effectively dealing with the problem of corrupted labels. Experimental results demonstrate the advantages of TALEN over existing noise-robust learning methods on various datasets.
IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS
(2023)
Article
Multidisciplinary Sciences
Suryanarayana Maddu, Bevan L. Cheeseman, Ivo F. Sbalzarini, Christian L. Mueller
Summary: This study presents a robust statistical learning framework for identifying differential equations from noisy spatio-temporal data. The proposed stability-based model selection approach improves robustness against noise by determining the appropriate level of regularization. The combination of stability selection and sparsity-promoting regression methods provides an interpretable criterion and outperforms previous approaches in terms of accuracy, data requirements, and robustness.
PROCEEDINGS OF THE ROYAL SOCIETY A-MATHEMATICAL PHYSICAL AND ENGINEERING SCIENCES
(2022)