Quasi-recurrent Neural Network Research Articles

Automated EEG based Emotion Detection using Bonobo Optimizer with Deep Learning on Human Computer Interaction

Recently, Emotion detection utilizing EEG signals develops popularity in domain of Human-Computer Interaction (HCI). EEG (electroencephalography) is a non-invasive approach, which processes electrical action from the brain through electrodes located in the scalp. An emotion recognition approach could not only be significant for healthy people among them disabled persons for detecting emotional changes and is utilized for different applications. It is significant to realize that emotion recognition in EEG indications is a difficult task owing to difficult and subjective nature of emotions. In recent times, Machine learning (ML) algorithms like Random Forests or Support Vector Machines (SVM) and Deep Learning (DL) systems namely Recurrent Neural Network (RNN) or Convolutional Neural Network (CNN) are trained on EEG feature extracted and connected emotional labels for classifying the user emotional state. This study presents an Automated EEG-based Emotion Detection using Bonobo Optimizer with Deep Learning (AEEGED-BODL) technique on HCI applications. The goal of the study is to analyze the EEG signals for the classification of several kinds of emotions in HCI applications. To achieve this, the AEEGED-BODL technique uses Higuchi fractal dimension (HFD) approach for extracting features in the EEG signals. Besides, the AEEGED-BODL technique makes use of the quasi-recurrent neural network (QRNN) approach for the detection and classification of distinct kinds of emotions. Furthermore, the BO system was demoralized for optimum hyperparameter selection of QRNN model, which helps in attaining an improved detection rate. The simulation validation of AEEGED-BODL algorithm was simulated on EEG signal database. The comprehensive result stated best outcome of the AEEGED-BODL algorithm over other recent approaches

Modelling of Whale Optimization Algorithm with Deep Learning for Skin Cancer Diagnosis on Dermoscopic Images

Skin cancer diagnoses through the analysis of dermoscopy images have undergone a revolutionary change with the incorporation of advanced Machine Learning (ML) approaches. By using contemporary models such as Convolutional Neural Network (CNN) and Deep Learning (DL) structures, this technique allows automated and accurate detection of malignant and benign skin tumours. This model learns intricate patterns and features directly from the image, which allows them to distinguish between different types of skin tumours based on colour, texture, and structural characteristics. This manuscript presents the Whale Optimization Algorithm with Deep Learning for Skin Cancer Diagnoses on Dermoscopy Images (WOADL-SCDDI) technique. The objective of the WOADL-SCDDI method is to analyze the Dermoscopic images for classifying and recognizing skin tumours. In the proposed WOADL-SCDDI technique, median filtering-based preprocessing and SegNet-based segmentation are involved. In addition, the WOADL-SCDDI technique offers a ResNet50 model for feature extraction with WOA-based hyperparameter tuning. Finally, Quasi Recurrent Neural Network (QRNN) model can be applied for skin tumour recognition and classification process. A comprehensive set of simulations have been conducted to highlight the better efficiency of the WOADL-SCDDI technique. The accomplished outcomes portrayed the improvements of the WOADL-SCDDI method compared to existing DL models.

Automated Lung Disease Detection and Classification Using Quantum Glowworm Swarm Optimizer with Quasi Recurrent Neural Network on Chest X-Ray Images

Lung diseases or otherwise called respiratory diseases are airborne diseases that affect the lungs and the other tissues of the lungs. Tuberculosis, Coronavirus Disease 2019 (COVID-19), and Pneumonia are a few instances of lung diseases. If the lung disease is diagnosed and treated in the initial stage, the chances of recovery rate and long-term survival rates can be increased. Usually, lung disease is identified by Chest X-Ray (CXR) image examination, skin test, sputum sample test, Computed Tomography (CT) scan examination, and blood test. Because of its non-invasive and convenient evaluation for overall outcomes of the chest situation, Lung disease can be detected by specialized radiologists on CXR images. In recent times, Deep Learning (DL) applies to medical images for disease detection and has proved an effective technique for detecting disease. The recent advancement of DL supports the detection and classification of lung diseases in medicinal imaging. This article presents an Automated Lung Disease Detection Using Quantum Glowworm Swarm Optimization with Quasi Recurrent Neural Network (QGSO-QRNN) model on CXR imaging. The presented QGSO-QRNN technique focuses on the identification of lung diseases using DL concepts. To accomplish this, the presented QGSO-QRNN technique initially performs image pre-processing by the use of the Gaussian Filtering (GF) technique. Besides, the Faster SqueezeNet approach is exploited for feature vector generation. Finally, the QRNN model is applied for precise classification of lung diseases with the QGSO technique as a hyperparameter optimizer. The investigational assessment of the QGSO-QRNN technique is examined by employing standard medical datasets and the outputs display the promising performance of the QGSO-QRNN technique over other existing techniques by means of diverse measures.

Enhanced Sea Horse Optimization with Deep Learning-based Multimodal Fusion Technique for Rice Plant Disease Segmentation and Classification

The detection of diseases in rice plants is an essential step in ensuring healthy crop growth and maximizing yields. A real-time and accurate plant disease detection technique can assist in the development of mitigation strategies to ensure food security on a large scale and economical rice crop protection. An accurate classification of rice plant diseases using DL and computer vision could create a foundation to achieve a site-specific application of agrochemicals. Image investigation tools are efficient for the early diagnosis of plant diseases and the continuous monitoring of plant health status. This article presents an Enhanced Sea Horse Optimization with Deep Learning-based Multimodal Fusion for Rice Plant Disease Detection and Classification (ESHODL-MFRPDC) technique. The proposed technique employed a DL-based fusion process with a hyperparameter tuning strategy to achieve an improved rice plant disease detection performance. The ESHODL-MFRPDC approach used Bilateral Filtering (BF)-based noise removal and contrast enhancement as a preprocessing step. Furthermore, Mayfly Optimization (MFO) with a Multi-Level Thresholding (MLT) based segmentation process was used to recognize the diseased portions in the leaf image. A fusion of three DL models was used for feature extraction, namely Residual Network (ResNet50), Xception, and NASNet. The Quasi-Recurrent Neural Network (QRNN) was used for the recognition of rice plant diseases, and its hyperparameters were set using the ESHO method. The performance of the ESHODL-MFRPDC method was validated using the rice leaf disease dataset from the UCI database. An extensive comparison study demonstrated the promising performance of the proposed method over others.

An efficient approach for data-imbalanced hate speech detection in Arabic social media

The use of hateful language in public debates and forums is becoming more common. However, this might result in antagonism and conflicts among individuals, which is undesirable in an online environment. Countries, businesses, and educational institutions are exerting their greatest efforts to develop effective solutions to manage this issue. In addition, recognizing such content is difficult, particularly in Arabic, due to a variety of challenges and constraints. Long-tailed data distribution is often one of the most significant issues in actual Arabic hate speech datasets. Pre-trained models, such as bidirectional encoder representations from transformers (BERT) and generative pre-trained transformers (GPT), have become more popular in numerous natural language processing (NLP) applications in recent years. We conduct extensive experiments to address data imbalance issues by utilizing oversampling methods and a focal loss function in addition to traditional loss functions. Quasi-recurrent neural networks (QRNN) are employed to fine-tune the cutting-edge transformer-based models, MARBERTv2, MARBERTv1, and ARBERT. In this context, we suggest a new approach using ensemble learning that incorporates best-performing models for both original and oversampled datasets. Experiments proved that our proposed approach achieves superior performance compared to the most advanced methods described in the literature.

Blockchain-Assisted Machine Learning with Hybrid Metaheuristics-Empowered Cyber Attack Detection and Classification Model

Cyber attack detection is the process of detecting and responding to malicious or unauthorized activities in networks, computer systems, and digital environments. The objective is to identify these attacks early, safeguard sensitive data, and minimize the potential damage. An intrusion detection system (IDS) is a cybersecurity tool mainly designed to monitor system activities or network traffic to detect and respond to malicious or suspicious behaviors that may indicate a cyber attack. IDSs that use machine learning (ML) and deep learning (DL) have played a pivotal role in helping organizations identify and respond to security risks in a prompt manner. ML and DL techniques can analyze large amounts of information and detect patterns that may indicate the presence of malicious or cyber attack activities. Therefore, this study focuses on the design of blockchain-assisted hybrid metaheuristics with a machine learning-based cyber attack detection and classification (BHMML-CADC) algorithm. The BHMML-CADC method focuses on the accurate recognition and classification of cyber attacks. Moreover, the BHMML-CADC technique applies Ethereum BC for attack detection. In addition, a hybrid enhanced glowworm swarm optimization (HEGSO) system is utilized for feature selection (FS). Moreover, cyber attacks can be identified with the design of a quasi-recurrent neural network (QRNN) model. Finally, hunter–prey optimization (HPO) algorithm is used for the optimal selection of the QRNN parameters. The experimental outcomes of the BHMML-CADC system were validated on the benchmark BoT-IoT dataset. The wide-ranging simulation analysis illustrates the superior performance of the BHMML-CADC method over other algorithms, with a maximum accuracy of 99.74%.

Multi-Object Detection and Tracking Using Reptile Search Optimization Algorithm with Deep Learning

Multiple-Object Tracking (MOT) has become more popular because of its commercial and academic potential. Though various techniques were devised for managing this issue, it becomes a challenge because of factors such as severe object occlusions and abrupt appearance changes. Tracking presents the optimal outcomes whenever the object moves uniformly without occlusion and in the same direction. However, this is generally not a real scenario, particularly in complicated scenes such as dance events or sporting where a greater number of players are tracked, moving quickly, varying their speed and direction, along with distance and position from the camera and activity they are executing. In dynamic scenes, MOT remains the main difficulty due to the symmetrical shape, structure, and size of the objects. Therefore, this study develops a new reptile search optimization algorithm with deep learning-based multiple object detection and tracking (RSOADL–MODT) techniques. The presented RSOADL–MODT model intends to recognize and track the objects that exist with position estimation, tracking, and action recognition. It follows a series of processes, namely object detection, object classification, and object tracking. At the initial stage, the presented RSOADL–MODT technique applies a path-augmented RetinaNet-based (PA–RetinaNet) object detection module, which improves the feature extraction process. To improvise the network potentiality of the PA–RetinaNet method, the RSOA is utilized as a hyperparameter optimizer. Finally, the quasi-recurrent neural network (QRNN) classifier is exploited for classification procedures. A wide-ranging experimental validation process takes place on DanceTrack and MOT17 datasets for examining the effectual object detection outcomes of the RSOADL–MODT algorithm. The simulation values confirmed the enhancements of the RSOADL–MODT method over other DL approaches.

SATFuzz: A Stateful Network Protocol Fuzzing Framework from a Novel Perspective

Stateful network protocol fuzzing is one of the essential means for ensuring network communication security. However, the existing methods have problems, including frequent auxiliary message interaction, no in-depth state-space exploration, and high shares of invalid interaction time. To this end, we propose SATFuzz, a stateful network protocol fuzzing framework. SATFuzz first prioritizes the states identified by the status codes in response messages, then randomly selects a state to test among the high-priority states, and determines its corresponding optimal test sequence, which is composed of the minimum pre-lead sequence, the test case, and the fittest post-end sequence. Finally, SATFuzz uses a quasi-recurrent neural network (QRNN) to filter the test cases before performing interaction, and only the optimal test sequence, including the valid test case, can be fed to the protocol entity. To verify the proposed framework, we conduct extensive experiments with the state-of-the-art fuzzer on two popular protocols. The results show that the vulnerability discovery efficiency of the proposed approach increases by at least 1.48 times (at most by 3.06 times), making it superior to the rival methods. This not only confirms the effectiveness of SATFuzz in terms of improving the vulnerability discovery efficiency but also shows that SATFuzz has significant advantages.

GS-QRNN: A High-Efficiency Automatic Modulation Classifier for Cognitive Radio IoT

Cognitive radio (CR) has been introduced into the Internet of Things (IoT) domain as a potential solution to the critical spectrum resource shortage caused by a dramatic increase of IoT solutions within a limited operational spectrum. One technique of use within traditional CR processing for sensing communication activity across a spectrum of interest is automatic modulation classification (AMC). However, utilizing AMC within resource-limited cognitive radio-enabled IoT (CR-IoT) devices poses a significant challenge. In this article, we present a high-efficiency automatic modulation classifier architecture whose core is built upon the implementation of stacking quasirecurrent neural network (S-QRNN) layers acting as a feature extraction stage. This architecture utilizes the low-latency feature extraction capability of convolutional layers, and a minimalist recurrent pooling function that mimics recurrent-layer operations to aggregate the extracted features over time steps for higher classification accuracy. Additionally, implementing dense layers between two consecutive QRNN layers helps keep the network growth rate low. Therefore, the proposed S-QRNN classifier overall exhibits higher efficiency, fitting well within the limitations of CR-IoT devices. In order to demonstrate higher performance efficiency of our classifier, we conduct a comprehensive evaluation of our approach against the state-of-the-art AMC classifiers. Our analysis metrics of the classifier performance focuses on trainability, classification accuracy, execution latency, and resulting overall efficiency. Our results demonstrate that our proposed S-QRNN classifier exhibits higher trainability and, on average, a 75.83% higher efficiency, while it has, on average, a 26.54% lower execution latency. We then expand upon S-QRNN’s foundation by introducing gated recurrent units into our classifier to initially extract temporal features of the received constellations. The resulting GS-QRNN classifier demonstrates an efficiency increase by an average of 191% compared to the gated-enhanced recurrent-neural-network-based AMC classifier, with our GS-QRNN classifier average execution latency being 59.31% lower.

Machine learning to probe modal interaction in dynamic atomic force microscopy

Modal interactions are pervasive effects that commonly emerge in nanomechanical systems. The coupling of vibrating modes can be leveraged in many ways, including to enhance sensing or to disclose complex phenomenologies. In this work we show how machine learning and data-driven approaches could be used to capture intermodal coupling. We employ a quasi-recurrent neural network (QRNN) for identifying mode coupling and verify its applicability on experimental data obtained from tapping mode atomic force microscopy (AFM). Hidden units of the QRNN are monitored to trace fingerprints of modes activation and to quantify their contributions over the global distortion of orbits in the phase space. To demonstrate the broad applicability of the method, the trained model is re-applied over different experiments and on diverse materials. Over a range of tip-sample configurations, dynamic AFM possesses features general enough to be seized by the QRNN and it is not required an ad-hoc re-training for the identification of interacting modes. Our study opens up a route for utilizing established machine learning techniques for rapid recognition of nonlinear complex effect such as internal resonances in nanotechnology. The QRNN analysis is meant to assist AFM sensing operations when exploiting modal interaction to enhance the signal-to-noise ratio of higher harmonics and provide high resolution compositional contrast in multi-frequency AFM applications.

Towards a Hybrid Machine Learning Model for Intelligent Cyber Threat Identification in Smart City Environments

The concept of a smart city requires the integration of information and communication technologies and devices over a network for the better provision of services to citizens. As a result, the quality of living is improved by continuous analyses of data to improve service delivery by governments and other organizations. Due to the presence of extensive devices and data flow over networks, the probability of cyber attacks and intrusion detection has increased. The monitoring of this huge amount of data traffic is very difficult, though machine learning algorithms have huge potential to support this task. In this study, we compared different machine learning models used for cyber threat classification. Our comparison was focused on the analyzed cyber threats, algorithms, and performance of these models. We have identified that real-time classification, accuracy, and false-positive rates are still the major issues in the performance of existing models. Accordingly, we have proposed a hybrid deep learning (DL) model for cyber threat intelligence (CTI) to improve threat classification performance. Our model was based on a convolutional neural network (CNN) and quasi-recurrent neural network (QRNN). The use of QRNN not only resulted in improved accuracy but also enabled real-time classification. The model was tested on BoT-IoT and TON_IoT datasets, and the results showed that the proposed model outperformed the other models. Due to this improved performance, we emphasize that the application of this model in the real-time environment of a smart system network will help in reducing threats in a reasonable time.

Short-Term Wind Speed Interval Prediction using LUBE based Quasi-Recurrent Neural Network

Accurate wind speed forecasting is an extremely difficult task and therefore less accurate forecasts may sometimes make the power systems vulnerable. Interval predictions on the other hand may help in maximizing the usage of integrated wind energy, as well as to reduce the adverse effects of the uncertainties introduced by the random fluctuations of wind, to the power systems. In this paper, we propose a novel wind speed interval prediction model by integrating lower upper bound estimation method (LUBE) into a quasi-recurrent neural network (QRNN). LUBE is a non-differentiable loss function requiring usage of meta-heuristic optimization techniques for training the model. By incorporating two objective functions, we proposed a new form of LUBE, using which the network can be trained with conventional stochastic gradient descent. The proposed model is evaluated for eight different datasets distributed among two wind fields. Results obtained from the computational experiments reveals that the proposed method generated narrow intervals with high coverage which helps it in achieving an improvement of 33% in coverage width criterion over the traditional models.

Improved Spike-Based Brain-Machine Interface Using Bayesian Adaptive Kernel Smoother and Deep Learning

Multiunit activity (MUA) has been proposed to mitigate the robustness issue faced by single-unit activity (SUA)-based brain-machine interfaces (BMIs). Most MUA-based BMIs still employ a binning method for estimating firing rates and linear decoder for decoding behavioural parameters. The limitations of binning and linear decoder lead to suboptimal performance of MUA-based BMIs. To address this issue, we propose a method which consists of Bayesian adaptive kernel smoother (BAKS) as the firing rate estimation algorithm and deep learning, particularly quasi-recurrent neural network (QRNN), as the decoding algorithm. We evaluated the proposed method for reconstructing (offline) hand kinematics from intracortical neural data chronically recorded from the primary motor cortex of two non-human primates. Extensive empirical results across recording sessions and subjects showed that the proposed method consistently outperforms other combinations of firing rate estimation algorithm and decoding algorithm. Overall results suggest the effectiveness of the proposed method for improving the decoding performance of MUA-based BMIs.

Autoencoder Quasi-Recurrent Neural Networks for Remaining Useful Life Prediction of Engineering Systems

Remaining useful life (RUL) prediction is a key solution to improve the reliability, availability, and maintainability of engineering systems. Long short-term memory (LSTM) and convolution neural networks (CNN) are the current hotspots in the field of RUL prediction. However, the LSTM-based prognostic approach has a slow loop step to process large-scale time-series data since the dependence of the data processing process at each time on the output of the previous time limits parallelism, and the CNN-based prognostic approach is not fit for time-series data although it can process the data in parallel. In this article, a new autoencoder quasi-recurrent neural networks (AEQRNN) based prognostic approach is proposed for RUL prediction of the engineering systems. The AEQRNN contains convolution components that can process input data in parallel, and pooling components which has two LSTM-like gate structures to process time-series data. In addition, the AEQRNN can automatically extract hidden features from monitoring signals without manual feature design. The effectiveness of the proposed prognostic approach is validated by three prognostic benchmarking datasets, including a turbofan engine dataset, a rolling bearing dataset, and a machining tool dataset. Experimental results demonstrate that this approach has both superior prognostic performance and training speed in comparison with other kinds of recurrent-neural-network-based approaches and various state-of-the-art approaches in the recent literature.

A Parallel Electrical Optimized Load Forecasting Method Based on Quasi-Recurrent Neural Network

Based on massive power big data resources, this paper establishes a new model for short-term load forecasting based on quasi-recurrent neural network (QRNN). QRNN combines the structural advantages of recurrent neural network (RNN) and convolutional neural network (CNN). It takes advantage of RNN’s cyclic connections to deal with the temporal dependencies of the load series, while implementing parallel calculations in both timestep and minibatch dimensions like CNN. The paper detailly describes the design and construction of QRNN, as well as the pre-processing and training steps of the forecasting model. Then, the algorithm is deployed to the big data platform, and an integrated load prediction system integrating data extraction, offline training, online forecasting and data visualization is developed. Finally, the proposed model is compared with some widely used machine learning load forecasting models. The results show that the QRNN based method achieves better prediction accuracy, and greatly improves the computational efficiency of training and testing, which is more practical for real-time and large-scale load forecasting.

Robust and accurate decoding of hand kinematics from entire spiking activity using deep learning

Objective. Brain–machine interfaces (BMIs) seek to restore lost motor functions in individuals with neurological disorders by enabling them to control external devices directly with their thoughts. This work aims to improve robustness and decoding accuracy that currently become major challenges in the clinical translation of intracortical BMIs. Approach. We propose entire spiking activity (ESA)—an envelope of spiking activity that can be extracted by a simple, threshold-less, and automated technique—as the input signal. We couple ESA with deep learning-based decoding algorithm that uses quasi-recurrent neural network (QRNN) architecture. We evaluate comprehensively the performance of ESA-driven QRNN decoder for decoding hand kinematics from neural signals chronically recorded from the primary motor cortex area of three non-human primates performing different tasks. Main results. Our proposed method yields consistently higher decoding performance than any other combinations of the input signal and decoding algorithm previously reported across long-term recording sessions. It can sustain high decoding performance even when removing spikes from the raw signals, when using the different number of channels, and when using a smaller amount of training data. Significance. Overall results demonstrate exceptionally high decoding accuracy and chronic robustness, which is highly desirable given it is an unresolved challenge in BMIs.

Overcoming Interpretability in Deep Learning Cancer Classification.

Since its inception, deep learning has revolutionized the field of machine learning and data-driven science. One such data-driven science to be transformed by deep learning is genomics. In the past decade, numerous genomics studies have adopted deep learning and its applications range from predicting regulatory elements to cancer classification. Despite its dominating efficacy in these applications, deep learning is not without drawbacks. A prominent shortcoming of deep learning is the lack of interpretability. Hence, the main objective of this study is to address this obstacle in the deep learning cancer classification. Here we adopt a feature importance scoring methodology (Gradient-based class activation mapping or Grad-CAM) on a quasi-recurrent neural network model that classify cancer based on FASTA sequencing data. In this study, we managed to formulate a nucleotide-to-genomic-region Grad-CAM scoring methodology, as well as, validate the use this methodology for the chosen model. Consequently, this allows for the utilization of the Grad-CAM scoring methodology for feature importance in deep learning cancer classification. The results from our study identify potential novel candidate genes, genomic elements, and mechanisms for future cancer research.

Bidirectional 3D Quasi-Recurrent Neural Network for Hyperspectral Image Super-Resolution

Hyperspectral imaging is unable to acquire images with high resolution in both spatial and spectral dimensions yet, due to physical hardware limitations. It can only produce low spatial resolution images in most cases and thus hyperspectral image (HSI) spatial super-resolution is important. Recently, deep learning-based methods for HSI spatial super-resolution have been actively exploited. However, existing methods do not focus on structural spatial-spectral correlation and global correlation along spectra, which cannot fully exploit useful information for super-resolution. Also, some of the methods are straightforward extension of RGB super-resolution methods, which have fixed number of spectral channels and cannot be generally applied to hyperspectral images whose number of channels varies. Furthermore, unlike RGB images, existing HSI datasets are small and limit the performance of learning-based methods. In this article, we design a bidirectional 3D quasi-recurrent neural network for HSI super-resolution with arbitrary number of bands. Specifically, we introduce a core unit that contains a 3D convolutional module and a bidirectional quasi-recurrent pooling module to effectively extract structural spatial-spectral correlation and global correlation along spectra, respectively. By combining domain knowledge of HSI with a novel pretraining strategy, our method can be well generalized to remote sensing HSI datasets with limited number of training data. Extensive evaluations and comparisons on HSI super-resolution demonstrate improvements over state-of-the-art methods, in terms of both restoration accuracy and visual quality.

Data-driven prediction and analysis of chaotic origami dynamics

Advances in machine learning have revolutionized capabilities in applications ranging from natural language processing to marketing to health care. Recently, machine learning techniques have also been employed to learn physics, but one of the formidable challenges is to predict complex dynamics, particularly chaos. Here, we demonstrate the efficacy of quasi-recurrent neural networks in predicting extremely chaotic behavior in multistable origami structures. While machine learning is often viewed as a “black box”, we conduct hidden layer analysis to understand how the neural network can process not only periodic, but also chaotic data in an accurate manner. Our approach shows its effectiveness in characterizing and predicting chaotic dynamics in a noisy environment of vibrations without relying on a mathematical model of origami systems. Therefore, our method is fully data-driven and has the potential to be used for complex scenarios, such as the nonlinear dynamics of thin-walled structures and biological membrane systems.

3-D Quasi-Recurrent Neural Network for Hyperspectral Image Denoising.

In this article, we propose an alternating directional 3-D quasi-recurrent neural network for hyperspectral image (HSI) denoising, which can effectively embed the domain knowledge-structural spatiospectral correlation and global correlation along spectrum (GCS). Specifically, 3-D convolution is utilized to extract structural spatiospectral correlation in an HSI, while a quasi-recurrent pooling function is employed to capture the GCS. Moreover, the alternating directional structure is introduced to eliminate the causal dependence with no additional computation cost. The proposed model is capable of modeling spatiospectral dependence while preserving the flexibility toward HSIs with an arbitrary number of bands. Extensive experiments on HSI denoising demonstrate significant improvement over the state-of-the-art under various noise settings, in terms of both restoration accuracy and computation time. Our code is available at https://github.com/Vandermode/QRNN3D.