Attention Weights Research Articles

A bidirectional interpretable compound-protein interaction prediction framework based on cross attention

The identification of compound-protein interactions (CPIs) plays a vital role in drug discovery. However, the huge cost and labor-intensive nature in vitro and vivo experiments make it urgent for researchers to develop novel CPI prediction methods. Despite emerging deep learning methods have achieved promising performance in CPI prediction, they also face ongoing challenges: (i) providing bidirectional interpretability from both the chemical and biological perspective for the prediction results; (ii) comprehensively evaluating model generalization performance; (iii) demonstrating the practical applicability of these models. To overcome the challenges posed by current deep learning methods, we propose a cross multi-head attention oriented bidirectional interpretable CPI prediction model (CmhAttCPI). First, CmhAttCPI takes molecular graphs and protein sequences as inputs, utilizing the GCW module to learn atom features and the CNN module to learn residue features, respectively. Second, the model applies cross multi-head attention module to compute attention weights for atoms and residues. Finally, CmhAttCPI employs a fully connected neural network to predict scores for CPIs. We evaluated the performance of CmhAttCPI on balanced datasets and imbalanced datasets. The results consistently show that CmhAttCPI outperforms multiple state-of-the-art methods. We constructed three scenarios based on compound and protein clustering and comprehensively evaluated the model generalization ability within these scenarios. The results demonstrate that the generalization ability of CmhAttCPI surpasses that of other models. Besides, the visualizations of attention weights reveal that CmhAttCPI provides chemical and biological interpretation for CPI prediction. Moreover, case studies confirm the practical applicability of CmhAttCPI in discovering anticancer candidates.

3-D LiDAR Localization Based on Novel Nonlinear Optimization Method for Autonomous Ground Robot

3D Light Detection and Ranging (LiDAR)-based localization for autonomous ground robot in unknown environment is a critical ability, which has been extensively studied. In this paper, we propose a novel nonlinear optimization method to promote the 3D LiDAR localization performance. Firstly, due to that the objective function is the summation of point correspondence matching error items, a novel item quality evaluation criterion is proposed. The quality of each item is directly proportional to the degree of its matching error distance descent during the optimization process. Then, an attention mechanism based on the criterion is proposed and the objective function is iteratively refined by putting different attention weights on the matching error items. Thirdly, in the Gauss-Newton optimization framework for LiDAR localization, we point out that the Hessian matrix is essentially the sum of the Hessian matrices of high-quality and low-quality point correspondence subsets. To increase the dominance of the Hessian matrix derived from the high-quality point correspondences, the Hessian matrix of the estimated high-quality point correspondence subset in the last updating step is added to the current Hessian matrix. The proposed LiDAR localization is extensively evaluated on public and custom datasets. The experimental results demonstrate that the proposed mechanisms can effectively promote the LiDAR localization performance. We will make our source code open to serve as a new baseline for the robotic community.

Anomaly detection using deep learning based model with feature attention

<span lang="EN-US">Anomaly detection is a difficult problem with numerous industrial applications, such as analyzing the quality of objects using images. Anomaly detection is the process of identifying outliers in a given dataset. Recently, machine learning approaches to computer vision problems have outperformed classical state-of-the-art approaches. Anomaly detection problems can be solved using supervised approaches. However, labelled datasets are hard to obtain. Thus, many researchers have taken an unsupervised approach to solving the problem of anomaly detection. In this study, we use an adversarial auto encoder model as a base model and create a custom model to detect anomalies in images and videos. The model was trained exclusively on normal data. The modified national institute of standards and technology database (MNIST) dataset achieved an area under curve (AUC) score of 0.872 for anomaly detection, while the University of California San Diego (UCSD) anomaly dataset (Video dataset) achieved an AUC score of 0.74 for Ped1 and 0.87 for Ped2. To calculate the anomaly score, the concept of attention weights is combined with the reconstruction loss, and the proposed method outperformed other similar methods designed for the same problem. However, the usefulness of the proposed model was demonstrated through the detection of anomalies, and the model is still being improved for use in real-world situations.</span>

Interpretable Multi-Scale Deep Learning for RNA Methylation Analysis across Multiple Species.

RNA modification plays a crucial role in cellular regulation. However, traditional high-throughput sequencing methods for elucidating their functional mechanisms are time-consuming and labor-intensive, despite extensive research. Moreover, existing methods often limit their focus to specific species, neglecting the simultaneous exploration of RNA modifications across diverse species. Therefore, a versatile computational approach is necessary for interpretable analysis of RNA modifications across species. A multi-scale biological language-based deep learning model is proposed for interpretable, sequential-level prediction of diverse RNA modifications. Benchmark comparisons across species demonstrate the model's superiority in predicting various RNA methylation types over current state-of-the-art methods. The cross-species validation and attention weight visualization also highlight the model's capability to capture sequential and functional semantics from genomic backgrounds. Our analysis of RNA modifications helps us find the potential existence of "biological grammars" in each modification type, which could be effective for mapping methylation-related sequential patterns and understanding the underlying biological mechanisms of RNA modifications.

The line bisection bias as a deficit of proportional reasoning − evidence from number line estimation in neglect

This study aimed to investigate whether neurological patients presenting with a bias in line bisection show specific problems in bisecting a line into two equal parts or their line bisection bias rather reflects a special case of a deficit in proportional reasoning more generally. In the latter case, the bias should also be observed for segmentations into thirds or quarters. To address this question, six neglect patients with a line bisection bias were administered additional tasks involving horizontal lines (e.g., segmentation into thirds and quarters, number line estimation, etc.). Their performance was compared to five neglect patients without a line bisection bias, 10 patients with right hemispheric lesions without neglect, and 32 healthy controls. Most interestingly, results indicated that neglect patients with a line bisection bias also overestimated segments on the left of the line (e.g., one third, one quarter) when dissecting lines into parts smaller than halves. In contrast, such segmentation biases were more nuanced when the required line segmentation was framed as a number line estimation task with either fractions or whole numbers. Taken together, this suggests a generalization of line bisection bias towards a segmentation or proportional processing bias, which is congruent with attentional weighting accounts of line bisection/neglect. As such, patients with a line bisection bias do not seem to have specific problems bisecting a line, but seem to suffer from a more general deficit processing proportions.

A collaborative siege method of multiple unmanned vehicles based on reinforcement learning

A method based on multi-agent reinforcement learning is proposed to tackle the challenges to capture escaping Target by Unmanned Ground Vehicles (UGVs). Initially, this study introduces environment and motion models tailored for cooperative UGV capture, along with clearly defined success criteria for direct capture. An attention mechanism integrated into the Soft Actor-Critic (SAC) is leveraged, directing focus towards pivotal state features pertinent to the task while effectively managing less relevant aspects. This allows capturing agents to concentrate on the whereabouts and activities of the target agent, thereby enhancing coordination and collaboration during pursuit. This focus on the target agent aids in refining the capture process and ensures precise estimation of value functions. The reduction in superfluous activities and unproductive scenarios amplifies efficiency and robustness. Furthermore, the attention weights dynamically adapt to environmental shifts. To address constrained incentives arising in scenarios with multiple vehicles capturing targets, the study introduces a revamped reward system. It divides the reward function into individual and cooperative components, thereby optimizing both global and localized incentives. By facilitating cooperative collaboration among capturing UGVs, this approach curtails the action space of the target UGV, leading to successful capture outcomes. The proposed technique demonstrates enhanced capture success compared to previous SAC algorithms. Simulation trials and comparisons with alternative learning methodologies validate the effectiveness of the algorithm and the design approach of the reward function.

Attention-based speech feature transfer between speakers.

In this study, we propose a simple yet effective method for incorporating the source speaker's characteristics in the target speaker's speech. This allows our model to generate the speech of the target speaker with the style of the source speaker. To achieve this, we focus on the attention model within the speech synthesis model, which learns various speaker features such as spectrogram, pitch, intensity, formant, pulse, and voice breaks. The model is trained separately using datasets specific to the source and target speakers. Subsequently, we replace the attention weights learned from the source speaker's dataset with the attention weights from the target speaker's model. Finally, by providing new input texts to the target model, we generate the speech of the target speaker with the styles of the source speaker. We validate the effectiveness of our model through similarity analysis utilizing five evaluation metrics and showcase real-world examples.

MSNSegNet: attention-based multi-shape nuclei instance segmentation in histopathology images.

In clinical research, the segmentation of irregularly shaped nuclei, particularly in mesenchymal areas like fibroblasts, is crucial yet often neglected. These irregular nuclei are significant for assessing tissue repair in immunotherapy, a process involving neovascularization and fibroblast proliferation. Proper segmentation of these nuclei is vital for evaluating immunotherapy's efficacy, as it provides insights into pathological features. However, the challenge lies in the pronounced curvature variations of these non-convex nuclei, making their segmentation more difficult than that of regular nuclei. In this work, we introduce an undefined task to segment nuclei with both regular and irregular morphology, namely multi-shape nuclei segmentation. We propose a proposal-based method to perform multi-shape nuclei segmentation. By leveraging the two-stage structure of the proposal-based method, a powerful refinement module with high computational costs can be selectively deployed only in local regions, improving segmentation accuracy without compromising computational efficiency. We introduce a novel self-attention module to refine features in proposals for the sake of effectiveness and efficiency in the second stage. The self-attention module improves segmentation performance by capturing long-range dependencies to assist in distinguishing the foreground from the background. In this process, similar features get high attention weights while dissimilar ones get low attention weights. In the first stage, we introduce a residual attention module and a semantic-aware module to accurately predict candidate proposals. The two modules capture more interpretable features and introduce additional supervision through semantic-aware loss. In addition, we construct a dataset with a proportion of non-convex nuclei compared with existing nuclei datasets, namely the multi-shape nuclei (MsN) dataset. Our MSNSegNet method demonstrates notable improvements across various metrics compared to the second-highest-scoring methods. For all nuclei, the score improved by approximately 1.66 , by about 2.15 , and by roughly 0.65 . For non-convex nuclei, which are crucial in clinical applications, our method's improved significantly by approximately 3.86 and by around 2.54 . These enhancements underscore the effectiveness of our approach on multi-shape nuclei segmentation, particularly in challenging scenarios involving irregularly shaped nuclei.

OII: An Orientation Information Integrating Network for Oriented Object Detection in Remote Sensing Images

Oriented object detection for remote sensing images poses formidable challenges due to arbitrary orientation, diverse scales, and densely distributed targets (e.g., across terrain). Current investigations in remote sensing object detection have primarily focused on improving the representation of oriented bounding boxes yet have neglected the significant orientation information of targets in remote sensing contexts. Recent investigations point out that the inclusion and fusion of orientation information yields substantial benefits in training an accurate oriented object system. In this paper, we propose a simple but effective orientation information integrating (OII) network comprising two main parts: the orientation information highlighting (OIH) module and orientation feature fusion (OFF) module. The OIH module extracts orientation features from those produced by the backbone by modeling the frequency information of spatial features. Given that low-frequency components in an image capture its primary content, and high-frequency components contribute to its intricate details and edges, the transformation from the spatial domain to the frequency domain can effectively emphasize the orientation information of images. Subsequently, our OFF module employs a combination of a CNN attention mechanism and self-attention to derive weights for orientation features and original features. These derived weights are adopted to adaptively enhance the original features, resulting in integrated features that contain enriched orientation information. Given the inherent limitation of the original spatial attention weights in explicitly capturing orientation nuances, the incorporation of the introduced orientation weights serves as a pivotal tool to accentuate and delineate orientation information related to targets. Without unnecessary embellishments, our OII network achieves competitive detection accuracy on two prevalent remote sensing-oriented object detection datasets: DOTA (80.82 mAP) and HRSC2016 (98.32 mAP).

Multi-scale Convolutional Feature Fusion Network Based on Attention Mechanism for IoT Traffic Classification

AbstractThe Internet of Things (IoT) has been extensively utilized in domains such as smart homes, healthcare, and other industries. With the exponential growth of Internet of Things (IoT) devices, they have become prime targets for malicious cyber-attacks. Effective classification of IoT traffic is, therefore, imperative to enable robust intrusion detection systems. However, IoT traffic data contain intricate spatial relationships and topological information, which traditional methods for traffic identification lack the capability to fully extract features and capture crucial characteristics. We propose a multi-scale convolutional feature fusion network augmented with a Convolutional Block Attention Module (MCF-CBAM) for accurate IoT traffic classification. The network incorporates three critical innovations: (1) Parallel convolution extracts multi-scale spatial features from traffic data. The 1 × 1 convolution operation reduces the amount of parameters and calculations of the network, thereby improving work efficiency. (2) The attention module suppresses less informative features while highlighting the most discriminative ones, enabling focused learning on decisive features. (3) Cross-scale connections with channel jumps reuse features from prior layers to enhance generalization. We evaluate the method extensively on three widely adopted public datasets. Quantitative results demonstrate MCF-CBAM establishes new state-of-the-art performance benchmarks for IoT traffic classification, surpassing existing methods by a significant margin. Qualitative visualizations of the learned attention weights provide intuitive insights into how the network automatically discovers the most decisive spatial features for identification. With its strong empirical performance and interpretable attention mechanisms, this work presents a promising deep learning solution to augment real-world IoT intrusion detection systems against growing cybersecurity threats.

Multi-knowledge enhanced graph convolution for learning resource recommendation

In recent years, E-learning has gained immense popularity as a prominent mode of education. However, accurately recommending learning resources from a vast amount of data remains a significant challenge. This study addresses two primary challenges impacting recommendation performance. Firstly, an imbalance exists between the abundance of available learning resources and the limited interaction behavior of learners. Secondly, the existing algorithms often overlook dynamic preference information, focusing primarily on learners’ short-term, static preferences only by learning interactive behavior but disregarding the multi-correlation between learning resources and learners. To tackle these challenges, we propose MkEGC (Multi-knowledge Enhanced Graph Convolution), a novel framework for learning resource recommendation. We approach the recommendation process as a Markov decision process. Initially, we construct a dual knowledge graph convolutional network, operating in learning resource-knowledge and learner-knowledge domains. This network facilitates the extraction of vector features from learning resources, enhances the learner vector representation, and captures higher-order preferences. Subsequently, we design hierarchical and attention weighting strategies to effectively extract latent hierarchical information from the knowledge graph. Finally, we integrate the learning resource state, the learning interaction state, and the sequence state to represent a multi-dimensional learner-state within the Markov decision framework, enabling prise learning resource recommendations. To validate the effectiveness of MkEGC, we conduct extensive experiments, comparing multiple sets of metrics with six state-of-the-art recommendation algorithms, utilizing two real-world datasets.

GPCR-BERT: Interpreting Sequential Design of G Protein-Coupled Receptors Using Protein Language Models.

With the rise of transformers and large language models (LLMs) in chemistry and biology, new avenues for the design and understanding of therapeutics have been opened up to the scientific community. Protein sequences can be modeled as language and can take advantage of recent advances in LLMs, specifically with the abundance of our access to the protein sequence data sets. In this letter, we developed the GPCR-BERT model for understanding the sequential design of G protein-coupled receptors (GPCRs). GPCRs are the target of over one-third of Food and Drug Administration-approved pharmaceuticals. However, there is a lack of comprehensive understanding regarding the relationship among amino acid sequence, ligand selectivity, and conformational motifs (such as NPxxY, CWxP, and E/DRY). By utilizing the pretrained protein model (Prot-Bert) and fine-tuning with prediction tasks of variations in the motifs, we were able to shed light on several relationships between residues in the binding pocket and some of the conserved motifs. To achieve this, we took advantage of attention weights and hidden states of the model that are interpreted to extract the extent of contributions of amino acids in dictating the type of masked ones. The fine-tuned models demonstrated high accuracy in predicting hidden residues within the motifs. In addition, the analysis of embedding was performed over 3D structures to elucidate the higher-order interactions within the conformations of the receptors.

Velocity-aware spatial-temporal attention LSTM model for inverse dynamic model learning of manipulators.

An accurate inverse dynamics model of manipulators can be effectively learned using neural networks. However, further research is required to investigate the impact of spatiotemporal variations in manipulator motion sequences on network learning. In this work, the Velocity Aware Spatial-Temporal Attention Residual LSTM neural network (VA-STA-ResLSTM) is proposed to learn a more accurate inverse dynamics model, which uses a velocity-aware spatial-temporal attention mechanism to extract dynamic spatiotemporal features selectively from the motion sequence of the serial manipulator. The multi-layer perception (MLP) attention mechanism is adopted to capture the correlation between joint position and velocity in the motion sequence, and the state correlation between hidden units in the LSTM network to reduce the weight of invalid features. A velocity-aware state fusion approach of LSTM network hidden units' states is proposed, which utilizes variation in joint velocity to adapt to the temporal characteristics of the manipulator dynamic motion, improving the generalization and accuracy of the neural network. Comparative experiments have been conducted on two open datasets and a self-built dataset. Specifically, the proposed method achieved an average accuracy improvement of 61.88% and 43.93% on the two different open datasets and 71.13% on the self-built dataset compared to the LSTM network. These results demonstrate a significant advancement in accuracy for the proposed method. Compared with the state-of-the-art inverse dynamics model learning methods of manipulators, the modeling accuracy of the proposed method in this paper is higher by an average of 10%. Finally, by visualizing attention weights to explain the training procedure, it was found that dynamic modeling only relies on partial features, which is meaningful for future optimization of inverse dynamic model learning methods.

In silico prediction of ocular toxicity of compounds using explainable machine learning and deep learning approaches.

The accurate identification of chemicals with ocular toxicity is of paramount importance in health hazard assessment. In contemporary chemical toxicology, there is a growing emphasis on refining, reducing, and replacing animal testing in safety evaluations. Therefore, the development of robust computational tools is crucial for regulatory applications. The performance of predictive models is heavily reliant on the quality and quantity of data. In this investigation, we amalgamated the most extensive dataset (4901 compounds) sourced from governmental GHS-compliant databases and literature to develop binary classification models of chemical ocular toxicity. We employed 12 molecular representations in conjunction with six machine learning algorithms and two deep learning algorithms to create a series of binary classification models. The findings indicated that the deep learning method GCN outperformed the machine learning models in cross-validation, achieving an impressive AUC of 0.915. However, the top-performing machine learning model (RF-Descriptor) demonstrated excellent performance with an AUC of 0.869 on the test set and was therefore selected as the best model. To enhance model interpretability, we conducted the SHAP method and attention weights analysis. The two approaches offered visual depictions of the relevance of key descriptors and substructures in predicting ocular toxicity of chemicals. Thus, we successfully struck a delicate balance between data quality and model interpretability, rendering our model valuable for predicting and comprehending potential ocular-toxic compounds in the early stages of drug discovery.

Knowledge‐based deep learning system for classifying Alzheimer's disease for multi‐task learning

AbstractDeep learning has recently become a viable approach for classifying Alzheimer's disease (AD) in medical imaging. However, existing models struggle to efficiently extract features from medical images and may squander additional information resources for illness classification. To address these issues, a deep three‐dimensional convolutional neural network incorporating multi‐task learning and attention mechanisms is proposed. An upgraded primary C3D network is utilised to create rougher low‐level feature maps. It introduces a new convolution block that focuses on the structural aspects of the magnetic resonance imaging image and another block that extracts attention weights unique to certain pixel positions in the feature map and multiplies them with the feature map output. Then, several fully connected layers are used to achieve multi‐task learning, generating three outputs, including the primary classification task. The other two outputs employ backpropagation during training to improve the primary classification job. Experimental findings show that the authors’ proposed method outperforms current approaches for classifying AD, achieving enhanced classification accuracy and other indicators on the Alzheimer's disease Neuroimaging Initiative dataset. The authors demonstrate promise for future disease classification studies.

Transliteration Characteristics in Romanized Assamese Language Social Media Text and Machine Transliteration

This article aims to understand different transliteration behaviors of Romanized Assamese text on social media. Assamese, a language that belongs to the Indo-Aryan language family, is also among the 22 scheduled languages in India. With the increasing popularity of social media in India and also the common use of the English Qwerty keyboard, Indian users on social media express themselves in their native languages, but using the Roman/Latin script. Unlike some other popular South Asian languages (say Pinyin for Chinese), Indian languages do not have a common standard romanization convention for writing on social media platforms. Assamese and English are two very different orthographical languages. Thus, considering both orthographic and phonemic characteristics of the language, this study tries to explain how Assamese vowels, vowel diacritics, and consonants are represented in Roman transliterated form. From a dataset of romanized Assamese social media texts collected from three popular social media sites: (Facebook, YouTube, and X (formerly known as Twitter)), 1 we have manually labeled them with their native Assamese script. A comparison analysis is also carried out between the transliterated Assamese social media texts with six different Assamese romanization schemes that reflect how Assamese users on social media do not adhere to any fixed romanization scheme. We have built three separate character-level transliteration models from our dataset. One using a traditional phrase-based statistical machine transliteration model, (1) PBSMT model and two separate neural transliteration models, (2) BiLSTM neural seq2seq model with attention, and (3) Neural transformer model. A thorough error analysis has been performed on the transliteration result obtained from the three state-of-the-art models mentioned above. This may help to build a more robust machine transliteration system for the Assamese social media domain in the future. Finally, an attention analysis experiment is also carried out with the help of attention weight scores taken from the character-level BiLSTM neural seq2seq transliteration model built from our dataset.

Spatial transfer of object-based statistical learning

A large number of recent studies have demonstrated that efficient attentional selection depends to a large extent on the ability to extract regularities present in the environment. Through statistical learning, attentional selection is facilitated by directing attention to locations in space that were relevant in the past while suppressing locations that previously were distracting. The current study shows that we are not only able to learn to prioritize locations in space but also locations within objects independent of space. Participants learned that within a specific object, particular locations within the object were more likely to contain relevant information than other locations. The current results show that this learned prioritization was bound to the object as the learned bias to prioritize a specific location within the object stayed in place even when the object moved to a completely different location in space. We conclude that in addition to spatial attention prioritization of locations in space, it is also possible to learn to prioritize relevant locations within specific objects. The current findings have implications for the inferred spatial priority map of attentional weights as this map cannot be strictly retinotopically organized.

Spatiotemporal attention aided graph convolution networks for dynamic spectrum prediction

To solve the spectrum scarcity problem, dynamic spectrum access (DSA) technology has emerged as a promising solution. Effectively implementing DSA demands accurate and efficient spectrum prediction. However, complex spatiotemporal correlation and heterogeneity in spectrum observations usually make spectral prediction arduous and even ambiguous. In this letter, we propose a spectrum prediction method based on an attention-aided graph convolutional neural network (AttGCN) to capture features in both spatial and temporal dimensions. By leveraging the attention mechanism, the AttGCN adapts its attention weights at different time steps and spatial positions, thus enabling itself to seize changes in spatiotemporal correlations dynamically. Simulation results show that the proposed spectrum prediction method performs better than baseline algorithms in long-term forecasting tasks.

Adaptive Semantic-Enhanced Transformer for Image Captioning.

In the research on image captioning, rich semantic information is very important for generating critical caption words as guiding information. However, semantic information from offline object detectors involves many semantic objects that do not appear in the caption, thereby bringing noise into the decoding process. To produce more accurate semantic guiding information and further optimize the decoding process, we propose an end-to-end adaptive semantic-enhanced transformer (AS-Transformer) model for image captioning. For semantic enhancement information extraction, we propose a constrained weaklysupervised learning (CWSL) module, which reconstructs the semantic object's probability distribution detected by the multiple instances learning (MIL) through a joint loss function. These strengthened semantic objects from the reconstructed probability distribution can better depict the semantic meaning of images. Also, for semantic enhancement decoding, we propose an adaptive gated mechanism (AGM) module to adjust the attention between visual and semantic information adaptively for the more accurate generation of caption words. Through the joint control of the CWSL module and AGM module, our proposed model constructs a complete adaptive enhancement mechanism from encoding to decoding and obtains visual context that is more suitable for captions. Experiments on the public Microsoft Common Objects in COntext (MSCOCO) and Flickr30K datasets illustrate that our proposed AS-Transformer can adaptively obtain effective semantic information and adjust the attention weights between semantic and visual information automatically, which achieves more accurate captions compared with semantic enhancement methods and outperforms state-of-the-art methods.

Image Restoration via Frequency Selection.

Image restoration aims to reconstruct the latent sharp image from its corrupted counterpart. Besides dealing with this long-standing task in the spatial domain, a few approaches seek solutions in the frequency domain by considering the large discrepancy between spectra of sharp/degraded image pairs. However, these algorithms commonly utilize transformation tools, e.g., wavelet transform, to split features into several frequency parts, which is not flexible enough to select the most informative frequency component to recover. In this paper, we exploit a multi-branch and content-aware module to decompose features into separate frequency subbands dynamically and locally, and then accentuate the useful ones via channel-wise attention weights. In addition, to handle large-scale degradation blurs, we propose an extremely simple decoupling and modulation module to enlarge the receptive field via global and window-based average pooling. Furthermore, we merge the paradigm of multi-stage networks into a single U-shaped network to pursue multi-scale receptive fields and improve efficiency. Finally, integrating the above designs into a convolutional backbone, the proposed Frequency Selection Network (FSNet) performs favorably against state-of-the-art algorithms on 20 different benchmark datasets for 6 representative image restoration tasks, including single-image defocus deblurring, image dehazing, image motion deblurring, image desnowing, image deraining, and image denoising.