Despite the recent success of large language models (LLMs) in time-series forecasting, most existing methods still adopt a Deep Synchronous Fusion strategy, where dense interactions between textual and temporal features are enforced at every layer of the network.This design overlooks the inherent granularity mismatch between modalities and leads to what we term semantic perceptual dissonance: highlevel abstract semantics provided by the LLM become inappropriately entangled with the lowlevel, fine-grained numerical dynamics of time series, making it difficult for semantic priors to effectively guide forecasting.To address this issue, we propose TimeSAF, a new framework based on hierarchical asynchronous fusion.Unlike synchronous approaches, TimeSAF explicitly decouples unimodal feature learning from cross-modal interaction.It introduces an independent cross-modal semantic fusion trunk, which uses learnable queries to aggregate global semantics from the temporal and prompt backbones in a bottom-up manner, and a stage-wise semantic refinement decoder that asynchronously injects these high-level signals back into the temporal backbone.This mechanism provides stable and efficient semantic guidance while avoiding interference with lowlevel temporal dynamics.Extensive experiments on standard long-term forecasting benchmarks show that TimeSAF significantly outperforms state-of-the-art baselines, and further exhibits strong generalization in both few-shot and zero-shot transfer settings.

Transformer-based foundation models have achieved remarkable progress in tasks such as time-series forecasting and image segmentation. However, they frequently suffer from error accumulation in multivariate long-sequence prediction and exhibit vulnerability to out-of-distribution samples in image-related tasks. Furthermore, these challenges become particularly pronounced in large-scale Web data analysis tasks, which typically involve complex temporal patterns and multimodal features. This complexity substantially increases optimization difficulty, rendering models prone to stagnation at saddle points within high-dimensional parameter spaces. To address these issues, we propose a lightweight Transformer architecture in conjunction with a novel Escape-Explore Optimizer (EEO). The optimizer enhances both exploration and generalization while effectively avoiding sharp minima and saddle-point traps. Experimental results show that, in representative Web data scenarios, our method achieves performance on par with state-of-the-art models across 11 time-series benchmark datasets and the Synapse medical image segmentation task. Moreover, it demonstrates superior generalization and stability, thereby validating its potential as a versatile cross-task foundation model for Web-scale data mining and analysis.

ABSTRACT In recent years, deep learning has made significant progress in image denoising. However, the complexity of advanced methods' systems is also increasing, which will increase the calculation cost and hinder the convenient analysis and comparison of methods. Therefore, a lightweight model based on invertible networks is proposed. The invertible network has great advantages in image denoising. It is lightweight, memory‐saving, and information‐lossless in backpropagation. To effectively remove the noise and restore a clean image, the high‐frequency part of the image is resampled and modeled to remove the impact of noise better. The channel context block is proposed to better focus on useful channels and improve the network's perception of useful information in images while ensuring the complexity and computing cost. At the same time, the residual structure with channel correlation modeling is used to extract the features in the convolutional flow, to effectively retain the details and texture of the image, and learn more details of the spatial features of the image, so as to prevent the blur and distortion of the image in the denoising process. The proposed method allows the model to enjoy lower computational complexity on the premise of ensuring performance.

Recent advances in transformer architectures have significantly improved performance in time‐series forecasting. Despite the excellent performance of attention mechanisms in global modeling, they often overlook local correlations between seasonal cycles. Drawing on the idea of trend‐seasonality decomposition, we design a seasonal cycle‐aware time‐series forecasting model (SCA‐Net). This model uses a dual‐branch extraction architecture to decompose time series into seasonal and trend components, modeling them based on their intrinsic features, thereby improving prediction accuracy and model interpretability. We propose a method combining global modeling and local feature extraction within seasonal cycles to capture the global view and explore latent features. Specifically, we introduce a frequency‐domain attention mechanism for global modeling and use multiscale dilated convolution to capture local correlations within each cycle, ensuring more comprehensive and accurate feature extraction. For simpler trend components, we apply a regression method and merge the output with the seasonal components via residual connections. To improve seasonal cycle identification, we design an adaptive decomposition method that extracts trend components layer by layer, enabling better decomposition and more useful information extraction. Extensive experiments on eight classic datasets show that SCA‐Net achieves a performance improvement of 12.1% in multivariate forecasting and 15.6% in univariate forecasting compared to the baseline.

Decomposing time series into odd and even component sequences is an effective method in time series analysis. However, this data partitioning sometimes leads to the weakening or even disappearance of local features in the original sequence within the odd and even component sequences, thereby reducing the accuracy of the model. To address this issue, we propose a novel neural network with an interactive attention mechanism in this paper. In order to allow the odd and even component sequences obtained after decomposition to capture more global information from the time series and compensate for the lost local features, we introduce odd-even fusion components. Through an interactive attention mechanism, the information of the odd component sequence, even component sequence, and odd-even fusion component sequence complement each other, yielding feature sub-sequences with different temporal relationship weights. Furthermore, we introduce an improved spatial attention submodule with C1 activation functions to better preserve local feature mappings. The segmented polynomial curve C1 activation function PP(x) not only incurs low computational overhead but also effectively alleviates the vanishing gradient problem, resulting in improved feature recognition capabilities. The C1 functional characteristics ensure continuity during backpropagation, guaranteeing stability during the model training process. Experimental results on multiple real-world datasets demonstrate the superior predictive and generalization capabilities of our model for time series forecasting tasks.

Recently, facial-expression recognition (FER) has primarily focused on images in the wild, including factors such as face occlusion and image blurring, rather than laboratory images. Complex field environments have introduced new challenges to FER. To address these challenges, this study proposes a cross-fusion dual-attention network. The network comprises three parts: (1) a cross-fusion grouped dual-attention mechanism to refine local features and obtain global information; (2) a proposed C2 activation function construction method, which is a piecewise cubic polynomial with three degrees of freedom, requiring less computation with improved flexibility and recognition abilities, which can better address slow running speeds and neuron inactivation problems; and (3) a closed-loop operation between the self-attention distillation process and residual connections to suppress redundant information and improve the generalization ability of the model. The recognition accuracies on the RAF-DB, FERPlus, and AffectNet datasets were 92.78%, 92.02%, and 63.58%, respectively. Experiments show that this model can provide more effective solutions for FER tasks.

Abstract This article introduces customized screening ensemble with shape‐adaptive quantile regression (CseAQR), a novel probabilistic interval forecasting method built upon the quantile regression model. CseAQR utilizes ensemble learning to perform adaptive quantile regression prediction, which can handle the heteroscedasticity feature in time series data by using a weighted adaptive allocation loss function to enhance the adaptability of the basic quantile regression model on the dataset. The model performance predictor is used to select the optimal ensemble learner combination, assign reasonable adaptive weights to it, and obtain a preliminary prediction interval through weighted aggregation. Combining ensemble learners not only improves the accuracy and robustness of prediction intervals but also ensures the commutativity required for conformal prediction. Finally, the conformal prediction method is applied to locally adjust the prediction interval, constructing a more consistently aligned prediction interval with the actual data on a narrower basis.

In recent years, time series forecasting models based on the Transformer framework have shown great potential, but they suffer from the inherent drawback of high computational complexity and only focus on global modeling. Inspired by trend-seasonality decomposition, we propose a method that combines global modeling with local feature extraction within the seasonal cycle. It aims at capturing the global view while fully exploring the potential features within each seasonal cycle and better expressing the long-term and periodic characteristics of time series. We introduce a frequency domain parity correction block to compute global attention and utilize multi-scale dilated convolution to extract local correlations within each cycle. Additionally, we adopt a dual-branch structure to separately model the seasonality and trend based on their intrinsic features, improving prediction performance and enhancing model interpretability. This model is implemented on a completely single-layer decoder architecture, breaking through the traditional encoder-decoder architecture paradigm and reducing computational complexity to a certain extent. We conducted sufficient experimental validation on eight benchmark datasets, and the results demonstrate its superior performance compared to existing methods in both univariate and multivariate forecasting.

With the continuous development of deep learning, long sequence time-series forecasting (LSTF) has attracted more and more attention in power consumption prediction, traffic prediction and stock prediction. In recent studies, various improved models of Transformer are favored. While these models have made breakthroughs in reducing the time and space complexity of Transformer, there are still some problems, such as the predictive power of the improved model being slightly lower than that of Transformer. And these models ignore the importance of special values in the time series. To solve these problems, we designed a more concise network named Resformer, which has four significant characteristics: (1) The fully sparse self-attention mechanism achieves O⁢(𝐿𝑙𝑜𝑔𝐿) time complexity. (2) The AMS module is used to process the special values of time series and has comparable performance on sequences dependency alignment. (3) Using quadratic linear transformation, a simple LT module is designed to replace the self-attention mechanism. It effectively reduces redundant information. (4) The DistPooling method based on data distribution is proposed to suppress redundant information and noise. A large number of experiments on real data sets show that the Resformer method is superior to the existing improved model and standard Transformer method.

Video super-resolution techniques aim to obtain high-resolution equivalents of existing low-resolution videos through a series of operations. In recent research, transformers have been increasingly popular because of their remarkable abilities in parallel computing and efficient extraction of space-time sequence features from videos. Moreover, combining self-attention and multi-scale methods has yielded excellent results. However, the combination of the two methods has limitations, current up-sampling methods struggle to match the global modeling capacity of self-attention mechanisms. Therefore, this paper proposes three strategies to combine the two methods. Based on the approximation strategy, we first construct a new bilinear up-sampling method for multi-scale acquisition. Convolution and cross-attention techniques are then used to correct and align features at different scales to prevent large deviations in feature extraction at a specific scale, which can affect subsequent feature extraction. Finally, to effectively solve the common computational complexity,$ C^{0} $continuity, and neuron death problems of existing activation functions, a new method to construct the activation function is proposed. The cubic spline function is used to construct a new activation function approximating tanh. The new activation function is$ C^{2} $continuous, which is piecewise defined by cubic polynomial curves. In this study, better results were achieved on three public video super-resolution test sets: REDS4, Vid4, and Vimeo-90K-T. Experiments demonstrated that the proposed method could provide a new solution for video super-resolution tasks.

Image segmentation is a basic problem in medical image analysis and useful for disease diagnosis. However, the complexity of medical images makes image segmentation difficult. In recent decades, fuzzy clustering algorithms have been preferred due to their simplicity and efficiency. However, they are sensitive to noise. To solve this problem, many algorithms using non-local information have been proposed, which perform well but are inefficient. This paper proposes an improved fuzzy clustering algorithm utilizing nonlocal self-similarity and a low-rank prior for image segmentation. Firstly, cluster centers are initialized based on peak detection. Then, a pixel correlation model between corresponding pixels is constructed, and similar pixel sets are retrieved. To improve efficiency and robustness, the proposed algorithm uses a novel objective function combining non-local information and a low-rank prior. Experiments on synthetic images and medical images illustrate that the algorithm can improve efficiency greatly while achieving satisfactory results.

Abstract Self‐similarity, a prior of natural images, has attracted much attention. The attribute means that low‐rank group matrices can be constructed from similar image patches. For low‐rank approximation denoising methods based on singular value decomposition (SVD) the ability to accurately construct group matrices with noise and handle singular values are keys. Here, combining image priors, a two‐stage clustering method to adaptively construct group matrices is designed. The method is anti‐noise, that is, when noise levels are high, these matrices are more accurate than that constructed by other algorithms. Then, according to the significance of singular values and singular vectors, singular vectors of the low‐rank estimations are corrected so that the residual noise in the low‐rank estimations is further suppressed. For back projection , the authors use the original noise level and the residual image to adaptively determine projection parameters and new noise levels . So, authors' back projection can provide a good foundation for authors' two‐stage denoising methods, better remove noise and preserve image details. Experimental results show that compared with the existing state‐of‐the‐art denoising algorithms, the proposed algorithm achieves competitive denoising performances in terms of quantitative metrics and preserving details. Especially with the increase of noise, the competitiveness of authors' algorithms is gradually enhanced.