Article
Computer Science, Artificial Intelligence
Kaiyang Zhou, Yongxin Yang, Yu Qiao, Tao Xiang
Summary: The researchers propose MixStyle, a module that addresses the problem of neural networks struggling to generalize to unseen data with domain shifts. MixStyle achieves this by mixing feature statistics of random instances during training, synthesizing new domains in the feature space and improving domain generalization performance.
INTERNATIONAL JOURNAL OF COMPUTER VISION
(2023)
Article
Neurosciences
Yufei Tan, Valerie Chanoine, Eddy Cavalli, Jean-Luc Anton, Johannes C. Ziegler
Summary: The noisy computation hypothesis suggests that developmental dyslexia may be caused by noise in the computational process, leading to less stable word representations. An fMRI experiment tested this hypothesis and found no evidence to support the idea that dyslexic readers have noisier neural representations compared to typical readers.
FRONTIERS IN HUMAN NEUROSCIENCE
(2022)
Article
Mathematics
Mingi Jeon, Taewook Kang, Jae-Jin Lee, Woojoo Lee
Summary: This paper focuses on spike-frequency adaptation and proposes a new method with more biological characteristics. The proposed method is shown to significantly reduce the number of spikes while maintaining performance through simulation experiments. Additionally, the paper demonstrates the close relationship between embedding biological meaning in SNNs and the low-power driving characteristics through in-depth analysis.
Article
Computer Science, Information Systems
Mingxin Zhang, Zipei Fan, Ryosuke Shibasaki, Xuan Song
Summary: In recent years, the use of WiFi fingerprints for indoor positioning has become popular due to the availability of WiFi and mobile communication devices. However, current methods for constructing fingerprint data sets are time-consuming and often focus on ideal laboratory environments. To tackle these issues, a WiFi domain adversarial graph convolutional network model is proposed, which can be trained using a small amount of labeled data and unlabeled crowdsensed WiFi fingerprints. The model constructs heterogeneous graphs based on received signal strength indicators (RSSIs) to effectively capture the topological structure of the data, and utilizes graph convolutional networks (GCNs) to extract graph-level embeddings for improved localization accuracy in large buildings.
IEEE INTERNET OF THINGS JOURNAL
(2023)
Article
Engineering, Civil
Huijun Liu, Chunhua Yang, Ao Li, Sheng Huang, Xin Feng, Zhimin Ruan, Yongxin Ge
Summary: Deep Domain Adaptation-based Crack Detection Network (DDACDN) is proposed in this paper to predict multi-category crack location information in target domain by leveraging source domain knowledge. The network extracts crack features, performs cross-domain adaptation, and recognizes and localizes pavement cracks. Experimental results demonstrate that DDACDN outperforms state-of-the-art methods in crack detection.
IEEE TRANSACTIONS ON INTELLIGENT TRANSPORTATION SYSTEMS
(2023)
Article
Neurosciences
Gavin Mischler, Menoua Keshishian, Stephan Bickel, Ashesh D. Mehta, Nima Mesgarani
Summary: The human auditory system has the ability to adapt to changes in background noise, allowing for continuous speech comprehension. However, the computations underlying this process are not well understood. In this study, a deep neural network (DNN) was used to model neural adaptation to noise and effectively reproduce the complex dynamics observed in the auditory system. The results provide new insights into the mechanisms of noise adaptation and speech perception in dynamic environments.
Article
Quantum Science & Technology
Yihua Wu, Chunhui Wu, Anqi Zhang, Shengmei Zhao
Summary: In this paper, a domain adaptation scheme based on hybrid classical-quantum neural network (hybrid DA) is proposed. It consists of four parts: classical convolutional neural network part, label predictor part based on quantum neural network (QNN), domain classifier part based on QNN, and gradient reversal layer part. The proposed hybrid DA can effectively predict features that cannot be discriminated between the source and target domain, and demonstrates higher classification accuracy with fewer parameters compared to classical DA. The feasibility of hybrid DA is verified on DIGIT-5 dataset, showing promising potential in the era of noisy intermediate-scale quantum devices.
QUANTUM INFORMATION PROCESSING
(2023)
Article
Computer Science, Artificial Intelligence
Zhiyong Huang, Kekai Sheng, Ke Li, Jian Liang, Taiping Yao, Weiming Dong, Dengwen Zhou, Xing Sun
Summary: Batch normalization is widely used in deep neural networks, but it is ineffective for cross-domain tasks. This paper proposes a novel normalization method called Reciprocal Normalization (RN), which utilizes cross-domain relation to improve adaptability. Compared to batch normalization, RN is more suitable for unsupervised domain adaptation and can be easily integrated into popular domain adaptation methods.
PATTERN RECOGNITION
(2023)
Article
Computer Science, Artificial Intelligence
Luyang Li, Tai Ma, Yue Lu, Qingli Li, Lianghua He, Ying Wen
Summary: In this paper, a multi-grained unsupervised domain adaptation approach (Muda) is proposed for semantic segmentation. Muda aims to enforce multi-grained semantic consistency between domains by aligning domains at both global and category level. Experimental results show that our model outperforms the state-of-the-art methods on two synthetic-to-real benchmarks.
PATTERN RECOGNITION
(2023)
Article
Engineering, Mechanical
Yumeng Liu, Xudong Li, Yang Hu
Summary: In recent years, deep transfer learning has been extensively used in fault diagnosis. Research has focused on utilizing domain adaptation methods in unsupervised transfer learning. However, there has been a lack of emphasis on the design of network structures. To address this, a novel approach that uses differentiable neural architecture search for automatic exploration of network structures suitable for fault diagnosis is proposed. The method explores key components within the Inception search space, resulting in the construction of comprehensive network architectures that selectively incorporate domain-relevant features. The effectiveness of the proposed method is demonstrated through experiments using public datasets, showing its remarkable capacity in generating networks with satisfactory performance.
MECHANICAL SYSTEMS AND SIGNAL PROCESSING
(2023)
Article
Behavioral Sciences
Eva-Maria Reuter, Arthur Booms, Li-Ann Leow
Summary: Sensorimotor adaptation is crucial for our ability to adapt to changes in a dynamic world. Recent studies combining EEG with behavioral and computational methods have provided insight into the neural mechanisms of adaptation.
NEUROSCIENCE AND BIOBEHAVIORAL REVIEWS
(2022)
Article
Computer Science, Artificial Intelligence
Xia Xue, Xia Sun, Hongyu Wang, Hao Zhang, Jun Feng
Summary: In this paper, a novel employee turnover prediction model called FATPNN is proposed. The model learns feature representations of personnel samples using GRU and employs an attention mechanism to model profile information, resulting in an effective prediction of employee turnover.
COMPLEX & INTELLIGENT SYSTEMS
(2023)
Article
Acoustics
Mohammed Alabsi, Larry Pearlstein, Michael Franco-Garcia
Summary: Data-driven fault diagnosis using deep learning algorithms is a popular research topic. However, without proper training, these models often struggle to generalize to different operating conditions. Most research in domain adaptation for machinery fault diagnosis focuses on transferring between similar working conditions. This paper proposes a semi-unsupervised domain adaptation approach that integrates model optimization and Generative Adversarial Networks (GANs) to bridge the gap between different machine domains. Experiments on bearing data sets demonstrate the effectiveness of this method in training a model that generalizes well and a generator that learns the source domain distribution for domain shifts.
JOURNAL OF VIBRATION AND CONTROL
(2023)
Article
Engineering, Biomedical
Robert M. Hinson, Joseph Berman, I-Chieh Lee, William G. Filer, He Huang
Summary: There is controversy over the value of offline evaluation of EMG-based neural-machine interfaces (NMIs) for their real-time application. This study aimed to investigate the relationship between the offline decoding accuracy of EMG-based NMIs and user adaptation, cognitive load, and physical effort in real-time NMI use. Results showed that high-level offline performance decoders yield the fastest adaptation rate and highest posture matching completion rate with the least muscle effort in users during online testing.
IEEE TRANSACTIONS ON NEURAL SYSTEMS AND REHABILITATION ENGINEERING
(2023)
Article
Computer Science, Information Systems
Changchun Zhang, Junguo Zhang
Summary: This paper proposes a method called Transferable Regularization and Normalization (TRN) for unsupervised domain adaptation. TRN adjusts feature norms and improves normalization techniques to avoid negative transfer and facilitate positive transfer. Evaluation results show that TRN achieves state-of-the-art performance on multiple benchmark datasets.
INFORMATION SCIENCES
(2022)