We correct an error in the proof of the main result of our original paper.
The advanced brain–computer interface (BCI) uses electroencephalography (EEG) to decode motor imagery (MI) tasks with applications in stroke rehabilitation. Convolutional neural networks (CNNs) have enhanced this technology using end-to-end algorithms in recent years. However, the MI-BCI paradigms display subject (humans) variability (SV) owing to the variable EEG characteristics. This paper proposes a contrastive attention network (CANet) to address SV and limited training data issues for subject-dependent and subject-independent (SI) evaluations. Specifically, CANet leverages the invariant characteristics of pooling layers to modify the spatial attention module by adding intraclass invariance (ICI) to interclass discrimination (ICD). Moreover, the channel attention module is modified by leveraging the equivariant characteristics of convolutional layers to add ICD to ICI. These modifications were considered because of the EEG’s intensive SV characteristics. The novel hetero-epitaxial attention provides robust refined features for SI evaluation. These mechanisms embed contrastive characteristics to mitigate SV along multiple dimensions using limited training data. The effectiveness of CANet was evaluated using two public MI-BCI datasets through quantitative performance comparisons and qualitative visualization of learned feature representations. The results demonstrate competitive performance and improved robustness under reduced training-data conditions, while t-SNE visualization indicates complementary contributions of the proposed spatial and channel attention branches to feature compactness and class separation.
In this paper, we study the dynamics of the limiting equation (i.e., non-autonomous ODE) of non-autonomous parabolic linear PDEs with dynamic boundary conditions. For this, we first show the well posedness of the PDEs and investigate how the solutions of given PDEs converge to those of the limiting equation. We show that, under certain standard conditions, bounded solutions exhibit additional smoothness - specifically Holder continuity - which allows us to obtain the uniform convergence of solutions. Consequently, the large-time behavior can be described by analyzing the limiting problem. Our approach involves non-autonomous singular perturbations of a reaction-diffusion equation with large diffusion in all domain and its boundary.
Background: Noise sensitivity is a key individual-level predictor of noise response, yet the Weinstein Noise Sensitivity Scale (WNSS), one of the most widely used instruments, has been subject to arbitrary classification approaches-including median splits and Z-score-based groupings-that may fail to reflect the inherent group structure of the data. This study applied Latent Class Analysis (LCA) to derive an empirically grounded classification scheme for the Korean WNSS. Methods: A self-report survey using the 21-item Korean WNSS was administered to 1,929 Korean adults. Reliability and validity of the scale were examined prior to LCA. LCA was conducted using Mplus 8, and classification stability was evaluated through 1,000 repeated random sampling iterations. The derived LCA-based classification was compared with conventional median- and Z-score-based classification methods, and data-driven cut-off scores were determined using kernel density estimation. Results: A four-class solution was identified as optimal (Entropy = 0.909, BIC = 125,318.9), with a replication rate of 43.2 % across repeated sampling iterations (n = 1,000). The four classes differed not only in mean noise sensitivity scores (Class 1: 2.92, Class 2: 3.57, Class 3: 4.38, Class 4: 4.77) but also in the item-level response patterns, particularly on items measuring emotional reactions to noise and everyday discomfort. Compared with conventional classification methods, median-based classification was found to systematically underestimate the size of the low group and overestimate the size of the high group, with 54.1 % of the conventional low-sensitivity group reclassified into LCA Class 2. Data-driven cut-off scores of 3.27, 3.96, and 4.76 were identified as class boundaries, showing high agreement with LCA-based classification (Class 4 specificity = 0.92, Class 2 sensitivity = 0.76). Conclusions: The four-class classification scheme and data-driven cut-off scores proposed in this study serve as empirically grounded criteria that complement conventional arbitrary classification, and may contribute to enhancing the reproducibility and cross-study comparability of noise sensitivity research. The proposed classification framework can provide a basis for formulating differentiated intervention strategies according to noise sensitivity level in environmental impact assessment, occupational health screening, and noise-related policy development.
