Learning Mechanism Research Articles

An enhanced salp swarm algorithm with chaotic mapping and dynamic learning for optimizing purge process of proton exchange membrane fuel cell systems

Swarm intelligence algorithm is now a research focus for the energy optimization of fuel cell systems to improve hydrogen utilization efficiency. However, conventional algorithms exhibit poor robustness and low execution efficiency since complex constraints are involved in the fuel cell cathode purge process. Given that, this paper innovatively introduces chaotic mapping and dynamic learning mechanisms into the Salp swarm algorithm to enhance the algorithm's global search and local exploitation capabilities under complex constraints. First, we developed a three-stage nonlinear cathode purge model to describe the different water transfer behaviors in the three purging stages. Then, an optimizing function is created under the complex staged constraints to accurately reveal the energy consumption mechanism. Finally, the enhanced Salp swarm algorithm enables the rapid and accurate determination of the three-stage dynamic purge flows based on minimizing energy consumption. Compared with the fixed flow purge scheme, we found that energy consumption is reduced by 3.6 %–8.2 % under various constraints. The enhanced Salp swarm algorithm proposed in this work demonstrates an outstanding performance in the energy consumption optimization of fuel cell systems.

Emergent coordination in temporal partitioning congestion games.

In this article we study the social dynamic of temporal partitioning congestion games (TPGs), in which participants must coordinate an optimal time-partitioning for using a limited resource. The challenge in TPGs lies in determining whether users can optimally self-organize their usage patterns. Reaching an optimal solution may be undermined, however, by a collectively destructive meta-reasoning pattern, trapping users in a socially vicious oscillatory behavior. TPGs constitute a dilemma for both human and animal communities. We developed a model capturing the dynamics of these games and ran simulations to assess its behavior, based on a 2×2 framework that distinguishes between the players' knowledge of other players' choices and whether they use a learning mechanism. We found that the only way in which an oscillatory dynamic can be thwarted is by adding learning, which leads to weak convergence in the no-information condition and to strong convergence in the with-information condition. We corroborated the validity of our model using real data from a study of bats' behaviour in an environment of water scarcity. We conclude by examining the merits of a complexity-based, agent-based modelling approach over a game-theoretic one, contending that it offers superior insights into the temporal dynamics of TPGs. We also briefly discuss the policy implications of our findings.

Enhanced Indoor Positioning System Using Ultra-Wideband Technology and Machine Learning Algorithms for Energy-Efficient Warehouse Management

The following article presents a proprietary real-time localization system using temporal analysis techniques and detection and localization algorithms supported by machine learning mechanisms. It covers both the technological aspects, such as proprietary electronics, and the overall architecture of the system for managing human and fixed assets. Its origins lie in the ever-increasing degree of automation in the management of company processes and the energy optimization associated with reducing the execution time of tasks in an intelligent building supported by in-building navigation. The positioning and tracking of objects in the presented system was realized using ultra-wideband radio tag technology. An exceptional focus has been placed on reducing the energy requirements of the components in order to maximize battery runtime, generate savings in terms of more efficient management of other energy consumers in the building and increase the equipment’s overall lifespan.

Virtual multiplexed immunofluorescence staining from non-antibody-stained fluorescence imaging for gastric cancer prognosis

Multiplexed immunofluorescence (mIF) staining, such as CODEX and MIBI, holds significant clinical value for various fields, such as disease diagnosis, biological research, and drug development. However, these techniques are often hindered by high time and cost requirements. Here we present a Multimodal-Attention-based virtual mIF Staining (MAS) system that utilises a deep learning model to extract potential antibody-related features from dual-modal non-antibody-stained fluorescence imaging, specifically autofluorescence (AF) and DAPI imaging. The MAS system simultaneously generates predictions of mIF with multiple survival-associated biomarkers in gastric cancer using self- and multi-attention learning mechanisms. Experimental results with 180 pathological slides from 94 patients with gastric cancer demonstrate the efficiency and consistent performance of the MAS system in both cancer and noncancer gastric tissues. Furthermore, we showcase the prognostic accuracy of the virtual mIF images of seven gastric cancer related biomarkers, including CD3, CD20, FOXP3, PD1, CD8, CD163, and PD-L1, which is comparable to those obtained from the standard mIF staining. The MAS system rapidly generates reliable multiplexed staining, greatly reducing the cost of mIF and improving clinical workflow. Stanford 2022 HAI Seed Grant; National Institutes of Health 1R01CA256890.

Impaired Face Feature-to-Location Statistical Learning and Single-Feature Discrimination in Developmental Prosopagnosia.

Individuals with developmental prosopagnosia (DP) experience severe face memory deficits that are often accompanied by impairments in face perception. Images of human facial features are better discriminated between when they are presented in the locations on the visual field that they typically appear in while viewing human faces in daily life, than in locations which they do not typically appear (i.e., better performance for eyes in the upper visual field, and better performance for mouths in the lower visual field). These feature-to-location tuning effects (FLEs) can be explained by a retinotopically organised visual statistical learning mechanism. We had a large group of DP participants (N = 64), a control group (N = 74) and a group of individuals with a mild form of DP (N = 58) complete a single-feature discrimination task to determine whether face perception deficits in DP can be accounted for by an impairment in face feature-to-location tuning. The results showed that individuals with DP did not have significant FLEs, suggesting a marked impairment in the underlying visual statistical learning mechanism. In contrast, the mild DP group showed normal FLE effects which did not differ from the control group. Both DP groups had impaired single-feature processing (SFP) as compared to the control group. We also examined the effects of age on FLEs and SFP.

Associations of sugar-related food parenting practices and parental feeding styles with prospective dietary behavior of children and adolescents: a systematic review of the literature from 2017 to 2023.

High consumption of sugar-rich foods and beverages has been associated with increased overweight and obesity in children and adolescents. Dietary behavior is influenced by learned mechanisms that originate in childhood and is associated with food parenting practices (FPP) and parental feeding styles (PFS). This systematic review aimed to narratively synthesize FPP and PFS concerning sugar-rich foods and beverages and their associations with the prospective dietary behavior of children and adolescents to derive evidence-based recommendations for health professionals and parents to promote healthy behaviors. A systematic literature search was conducted using PubMed and Web of Science databases covering the publication years 2017-2023. The results were narratively synthesized, and exposure-outcome matrices were used for visual representation. The review included 15 peer-reviewed studies from different geographical regions that investigated FPP or PFS regarding the handling of sugar-rich foods and beverages in children's diets and reported the associations with the prospective dietary behavior of children and adolescents. The findings indicate that highly controlling parental practices were associated with the development of unhealthy eating behaviors and preferences for sugar-rich foods and beverages over time. Conversely, parental practices that emphasized structure and balance in dietary choices yielded more positive long-term outcomes, associated with reduced preferences for sugar-rich foods and drinks. The results underscore the significance of fostering a healthy home environment and active parental role modeling in promoting healthier dietary behaviors among children and adolescents.

Parameter Calibration and Verification of Elastoplastic Wet Sand Based on Attention-Retention Fusion Deep Learning Mechanism

The discrete element method (DEM) is a vital numerical approach for analyzing the mechanical behavior of elastoplastic wet sand. However, parameter uncertainty persists within the mapping between constitutive relationships and inherent model parameters. We propose a Parameter calibration neural network based on Attention, Retention, and improved Transformer for Sequential data (PartsNet), which effectively captures the nonlinear mechanical behavior of wet sand and obtains the optimal parameter combination for the Edinburgh elasto-plastic adhesion constitutive model. Variational autoencoder-based principal component ordering is employed by PartsNet to reduce the high-dimensional dynamic response and extract critical parameters along with their weights. Gated recurrent units are combined with a novel sparse multi-head attention mechanism to process sequential data. The fusion information is delivered by residual multilayer perceptron, achieving the association between sequential response and model parameters. The errors in response data generated by calibrated parameters are quantified by PartsNet based on adaptive differentiation and Taylor expansion. Remarkable calibration capabilities are exhibited by PartsNet across six evaluation indicators, surpassing seven other deep learning approaches in the ablation test. The calibration accuracy of PartsNet reaches 91.29%, and MSE loss converges to 0.000934. The validation experiments and regression analysis confirmed the generalization capability of PartsNet in the calibration of wet sand. The improved sparse attention mechanism optimizes multi-head attention, resulting in a convergence speed of 21.25%. PartsNet contributes to modeling and simulating the precise mechanical properties of complex elastoplastic systems and offers valuable insights for diverse engineering applications.

A swarm anomaly detection model for IoT UAVs based on a multi-modal denoising autoencoder and federated learning

The widespread application of unmanned aerial vehicle (UAV) swarms has posed unique challenges for anomaly detection. Multi-modal noise from multi-source heterogeneous sensors during UAV swarm communication affects data quality, and limited data sharing between different UAV organisations restricts training a unified anomaly detection model. To address these problems, this study proposes a UAV swarm anomaly detection model based on a multi-modal denoising autoencoder and federated learning (L-MDAE). First, L-MDAE simulates noise by adding perturbations to the original data during UAV swarm communication. Second, according to the characteristics of UAV data noise, this study designs a new MSE loss function (normalised mean square error, NMSE) based on the normalised correlation coefficient. Furthermore, heterogeneous neural networks with NMSE are constructed to enhance the multi-modal noise–removal capability of the model. Finally, this study considers the UAV control node as the client and the ground control station as the server. Using a federated learning mechanism, L-MDAE is trained on a client dataset, and its parameters are integrated and distributed on the server. In this way, each UAV can effectively detect abnormal data using L-MDAE. Experimental results on five datasets, including ALFA, TLM and ITS, demonstrate that L-MDAE outperforms baseline and related models. When using ALFA, L-MDAE achieved an accuracy of 0.9919 and a swarm anomaly detection accuracy of 0.9901, approximately 2% higher than that of the baseline model.

Cognitive mechanisms of learning in sequential decision-making under uncertainty: an experimental and theoretical approach.

Learning to make adaptive decisions involves making choices, assessing their consequence, and leveraging this assessment to attain higher rewarding states. Despite vast literature on value-based decision-making, relatively little is known about the cognitive processes underlying decisions in highly uncertain contexts. Real world decisions are rarely accompanied by immediate feedback, explicit rewards, or complete knowledge of the environment. Being able to make informed decisions in such contexts requires significant knowledge about the environment, which can only be gained via exploration. Here we aim at understanding and formalizing the brain mechanisms underlying these processes. To this end, we first designed and performed an experimental task. Human participants had to learn to maximize reward while making sequences of decisions with only basic knowledge of the environment, and in the absence of explicit performance cues. Participants had to rely on their own internal assessment of performance to reveal a covert relationship between their choices and their subsequent consequences to find a strategy leading to the highest cumulative reward. Our results show that the participants' reaction times were longer whenever the decision involved a future consequence, suggesting greater introspection whenever a delayed value had to be considered. The learning time varied significantly across participants. Second, we formalized the neurocognitive processes underlying decision-making within this task, combining mean-field representations of competing neural populations with a reinforcement learning mechanism. This model provided a plausible characterization of the brain dynamics underlying these processes, and reproduced each aspect of the participants' behavior, from their reaction times and choices to their learning rates. In summary, both the experimental results and the model provide a principled explanation to how delayed value may be computed and incorporated into the neural dynamics of decision-making, and to how learning occurs in these uncertain scenarios.

The Impact of Educational Informatics on School Management Decision-Making in the Context of Big Data

In the domain of private education, particularly vocational education, school management encounters a complex and variable environment with diverse demands. As the advancement of educational informatics progresses, schools have accumulated a substantial amount of data, including student learning behavior, teaching data, and classroom utilization statistics. However, the efficient utilization of these data for scientific management decisions to enhance the quality of education and resource efficiency presents an urgent challenge. The study of the impacts of educational informatics on school management decision-making is of significant importance. It aids administrators in comprehensively understanding and analyzing the allocation and usage of educational resources, thereby enhancing management efficiency and teaching quality and further promoting the development of educational informatics. Despite some progress in data analysis and decision-making methods within educational informatics, deficiencies remain. Traditional methods often overlook comprehensive considerations of spatiotemporal data, failing to accurately reflect the dynamic changes in educational resources. Existing decision-making methods, predominantly based on single-agent models, lack studies on multi-agent collaborative decision-making, resulting in suboptimal decision outcomes. This study comprises three main components: first, the data modeling of educational informatics based on a spatiotemporal data model; second, an intelligent decision-making framework for school management under a reinforcement learning mechanism; and third, the implementation of an intelligent school management decision-making method based on multi-agent reinforcement learning. Through these investigations, the scientific nature and efficiency of school management are enhanced, providing new ideas and methods for management decision-making in other educational domains.

Expansive learning in the learning assistant model: how instructors’ goals lead to differences in implementation and development of LAs’ practices

BackgroundThe learning assistant (LA) model supports student success in undergraduate science courses; however, variation in outcomes has led to a call for more work investigating how the LA model is implemented. In this research, we used cultural historical activity theory (CHAT) to characterize how three different instructors set up LA-facilitated classrooms and how LAs’ understanding and development of their practices was shaped by the classroom activity. CHAT is a sociocultural framework that provides a structured approach to studying complex activity systems directed toward specific objects. It conceptualizes change within these systems as expansive learning, in which experiencing a contradiction leads to internalization and critical self-reflection, and then externalization and a search for solutions and change.ResultsThrough analyzing two semi-structured retrospective interviews from three professors and eleven LAs, we found that how the LA model was implemented differed based on STEM instructors’ pedagogical practices and goals. Each instructor leveraged LA-facilitated interactions to further learning and tasked LAs with emotionally supporting students to grapple with content and confusions in a safe environment; however, all three had different rules and divisions of labor that were influenced by their perspectives on learning and their objects for the class. For LAs, we found that they had multiple, sometimes conflicting, motives that can be described as either practical, what they described as their day-to-day job, or sense-making, how they made sense of the reason for their work. How these motives were integrated/separated or aligned/misaligned with the collective course object influenced LAs’ learning in practice through either a mechanism of consonance or contradiction. We found that each LA developed unique practices that reciprocally shaped and were shaped by the activity system in which they worked.ConclusionsThis study helps bridge the bodies of research that focus on outcomes from the LA model and LA learning and development by describing how LA learning mechanisms are shaped by their context. We also show that variation in the LA model can be described both by classroom objects and by LAs’ development in dialogue with those objects. This work can be used to start to develop a deeper understanding of how students, instructors, and LAs experience the LA model.

Pre-training enhanced spatio-temporal graph neural network for predicting influent water quality and flow rate of wastewater treatment plant: Improvement of forecast accuracy and analysis of related factors

Efficient management of wastewater treatment plants (WWTPs) necessitates accurate forecasting of influent water quality parameters (WQPs) and flow rate (Q) to reduce energy consumption and mitigate carbon emissions. The time series of WQPs and Q are highly non-linear and influenced by various factors such as temperature (T) and precipitation (Precip). Conventional models often struggle to account for long-term temporal patterns and overlook the complex interactions of parameters within the data, leading to inaccuracies in detecting WQPs and Q. This work introduced the Pre-training enhanced Spatio-Temporal Graph Neural Network (PT-STGNN), a novel methodology for accurately forecasting of influent COD, ammonia nitrogen (NH3-N), total phosphorus (TP), total nitrogen (TN), pH and Q in WWTPs. PT-STGNN utilizes influent data of the WWTP, air quality data and meteorological data from the service area as inputs to enhance prediction accuracy. The model employs unsupervised Transformer blocks for pre-training, with efficient masking strategies to effectively capture long-term historical patterns and contextual information, thereby significantly boosting forecasting accuracy. Furthermore, PT-STGNN integrates a unique graph structure learning mechanism to identify dependencies between parameters, further improving the model's forecasting accuracy and interpretability. Compared with the state-of-the-art models, PT-STGNN demonstrated superior predictive performance, particularly for a longer-term prediction (i.e., 12 h), with MAE, RMSE and MAPE at 12-h prediction horizon of 2.737 ± 0.040, 4.209 ± 0.060 and 13.648 ± 0.151 %, respectively, for the algebraic mean of each parameter. From the results of graph structure learning, it is observed that there are strong dependencies between NH3-N and TN, TP and Q, as well as Precip, etc. This study innovatively applies STGNN, not only offering a novel approach for predicting influent WQPs and Q in WWTPs, but also advances our understanding of the interrelationships among various parameters, significantly enhancing the model's interpretability.

Our brains sense the future through a new quantum-like implicit learning mechanism

BackgroundImagine if our brains could unconsciously predict future events. This study explores this concept, presenting evidence for an inherent 'foreseeing' ability, termed anomalous cognition (AC). We introduce a new experimentally verifiable approach to explain anomalous information anticipation (AIA), a type of AC, based on an innovative, quantum-like model of implicit learning, grounded in Nonlocal Plasticity Theory (NPT). MethodsOur research involved 203 participants using methods such as continuous flash suppression, random dot motion, and advanced 3D EEG neuroimaging, along with IBM quantum random event generators for precise measurements across 144 trials. These trials tested contingencies between undetectable sensory stimuli and dot movements, focusing on participants' prediction abilities. The design conditions were strictly experimental, violating fundamental classical learning principles, particularly reflex conditioning. If these principles were immutable, their violation would prevent any systematic behavioral changes, resulting in random responses. This violation was implemented through two quantum physics concepts: the mathematical principle of nonlocality and entanglement. ResultsDespite the sensory stimulus being inaccessible, our results showed a significant prediction between the contingencies and an increase in AIA accuracy, with explained variances between 25 % and 48 %. EEG findings supported this, showing a positive link between brain activity in specific regions and AIA success. Electrochemical activations were detected in the posterior occipital cortex, the intraparietal sulcus, and the medial temporal gyri. AIA hits exceeded the threshold value corresponding to one standard deviation above the expected mean, showing moderate effect sizes in the experimental group (Cohen’s d = 0.461). Analyzing the learning curve using the derivation technique, we identified the acceleration point of the wave function, indicating systematic implicit learning. This result showed that from repetition 63 onwards, AIA hits increased significantly. ConclusionsThe results suggest that, despite violating fundamental classical learning principles, cognitive processes produced changes in participants' responses susceptible to neuromodulation, considering quantum physics principles of nonlocality and entanglement (both present in NPT). We discuss (a) why the priming effect does not explain the significant results; (b) the potential discovery of a new form of quantum-like implicit learning, which could scientifically resolve phenomena associated with anomalous cognitions (e.g., AIA); and (c) future research directions, including potential applications and clinical impact.

Application of Optimized Deep Learning Mechanism for Recognition and Categorization of Retinal Diseases

Application of Optimized Deep Learning Mechanism for Recognition and Categorization of Retinal Diseases

Exploring the effectiveness of reward-based learning strategies for second-language speech sounds.

Adults struggle to learn non-native speech categories in many experimental settings (Goto, Neuropsychologia, 9(3), 317-323 1971), but learn efficiently in a video game paradigm where non-native speech sounds have functional significance (Lim & Holt, Cognitive Science, 35(7), 1390-1405 2011). Behavioral and neural evidence from this and other paradigms point toward the involvement of reinforcement learning mechanisms in speech category learning (Harmon, Idemaru, & Kapatsinski, Cognition, 189, 76-88 2019; Lim, Fiez, & Holt, Proceedings of the National Academy of Sciences, 116, 201811992 2019). We formalize this hypothesis computationally and implement a deep reinforcement learning network to map between environmental input and actions. Comparing to a supervised model of learning, we show that the reinforcement network closely matches aspects of human behavior in two experiments - learning of synthesized auditory noise tokens and improvement in speech sound discrimination. Both models perform comparably and the similarity in the output of each model leads us to believe that there is little inherent computational benefit to a reward-based learning mechanism. We suggest that the specific neural circuitry engaged by the paradigm and links between striatum and superior temporal areas play a critical role in effective learning.

Complexity in Systemic Cognition: Theoretical Explorations with Agent-Based Modeling

This paper presents a systemic view of human cognition that suggests complexityis an essential feature of such a system. It draws on the embodied, distributed, and extended cognition paradigms to outline the elements and the mechanisms that define cognition. In doing so, it uses an agent-based computational model (the TS 1.0.5Model) with a focus on learning mechanisms as they reflect on individual competence to gain insights on how cognition works. Results indicate that cognitive dynamics do not depend solely on macro structural elements, nor do they depend uniquely on individual characteristics. Instead, more insights and understanding are available through the consideration of all elements together as they co-evolve and interact over time. This perspective illustrates the essential role of how we define the meso domain and constitutes a clear indication that cognitive systems are indeed complex.

The Use of Short Stories to Improve Vocabulary: Understanding the Perceptions of EFL Students

The integration of technologies, such as virtual learning environments, web-based collaborative setups, and gamified applications, tend to promote the positive effects of short stories on vocabulary learning in a foreign language. This study investigates the effectiveness of using short stories as a pedagogical tool for enhancing vocabulary acquisition in English as a Foreign Language (EFL) students. Two cohorts of participants took part simultaneously in a course to improve vocabulary through short stories graded according to their level of proficiency. The highly validated tool MSLQ was applied to gather perceptions of motivation and learning strategies throughout the course. Correlational analyses suggest that short stories, combined with appropriate technological and pedagogical strategies, offer a powerful approach to vocabulary learning in diverse educational contexts. The study also emphasizes the importance of self-regulated learning mechanisms, which promote learner autonomy and self-efficacy, leading to better vocabulary outcomes. This research contributes to the understanding of effective vocabulary learning and provides practical insights for educators seeking to enhance EFL students' language mastery.

Neural Mechanisms of Learning and Consolidation of Morphologically Derived Words in a Novel Language: Evidence From Hebrew Speakers

Abstract We examined neural mechanisms associated with the learning of novel morphologically derived words in native Hebrew speakers within the Complementary Learning Systems (CLS) framework. Across four sessions, 28 participants were trained on an artificial language, which included two types of morphologically complex words: linear (root + suffix) with a salient structure, and non-linear (root interleaved with template), with a prominent derivational structure in participants’ first language (L1). A third simple monomorphemic condition, which served as baseline, was also included. On the first and fourth sessions, training was followed by testing in an fMRI scanner. Our behavioural results showed decomposition of both types of complex words, with the linear structure more easily learned than the non-linear structure. Our fMRI results showed involvement of frontal areas, associated with decomposition, only for the non-linear condition, after just the first session. We also observed training-related increases in activation in temporal areas specifically for the non-linear condition, which was correlated with participants’ L1 morphological awareness. These results demonstrate that morphological decomposition of derived words occurs in the very early stages of word learning, is influenced by L1 experience, and can facilitate word learning. However, in contrast to the CLS framework, we found no support for a shift from reliance on hippocampus to reliance on cortical areas in any of our conditions. Instead, our findings align more closely with recent theories showing a positive correlation between changes in hippocampus and cortical areas, suggesting that these representations co-exist and continue to interact with one another beyond initial learning.

Robust online active learning with cluster-based local drift detection for unbalanced imperfect data

With the rapid development of data-driven technologies, a massive amount of actual data emerges from industrial systems, forming data stream. Their data distribution may change over time and outliers may be generated as unbalanced imperfect data due to time-varying working condition, aging equipment, etc. Previous methods struggle with the dual challenges of concept drift and unbalance, however, fail to efficiently distinguishing outliers from a drift under the limited labeling budget, causing the performance degradation. To address the issue, robust online active learning with cluster-based local drift detection is proposed to classify unbalanced imperfect data stream with the above characteristics. The cluster-based local drift detection is first designed to capture a new concept and recognize the corresponding drifted regions. Following that, an improved active learning mechanism is presented to distinguish outliers from a drift, and select most valuable instances for labeling and updating ensemble classifier. Experimental results for eight synthetic and four real-world data streams show that the proposed method outperforms seven comparative methods on classification accuracy and robustness.

DMANet: Dual-modality alignment network for visible–infrared person re-identification

Visible–infrared person re-identification (VI-ReID) is a challenging retrieval task, which aims to match the same pedestrian between visible and infrared modalities. Most existing works achieve performance gains by solving the problem of the inherent cross-modality discrepancies. However, they cannot fully mine the modality information and lead to a poor generalization. In addition, the pedestrian images are unable to align well due to the large inter- and intra- class variations. To tackle the above limitations, we propose a novel dual-modality alignment network (DMANet) for VI-ReID. The core idea of our work is to develop multi-granularity features mutual learning (MGFML) for inadequate perception of modalities information, and to solve modality difference by proposing inter- and intra- modality alignment module (IIMA). Specifically, firstly, an effective multi-granularity features mutual learning module is proposed to mine the multi-granularity features, which combines the domain alignment and self-distillation to relieve modality discrepancy. Further, the maximum mean discrepancy loss and mutual learning loss are presented to enhance the identity-aware ability of the DMANet. Secondly, an effective inter- and intra- modality alignment module is presented to explore the potential alignment relation of inter- and intra- modalities. Finally, joint learning mechanism of multi-granularity features and modality alignment is utilized to improve the VI-ReID accuracy. Extensive experiments on mainstream benchmarks demonstrate that our method is superior to the state-of-the-art methods.