Language：English   Publishing type：Research paper (scientific journal)  

This paper discusses the visual mechanism of global orientation detection and the realization of a mechanism-based artificial visual system for two-dimensional orientation detection tasks. For interpretation and practicability, we introduce the visual mechanism into the design of a detection system. We first propose an orientation detection mechanism according to the color-orientation jointly selectivity cortical neuron character. We assume that part of the orientation detection tasks is completed by the color-orientation jointly selective cells that are only responsible for orientation detection locally. Each cell can only be activated by stimuli with a specific orientation angle and the preferred color. We realize these cells by the McCulloch–Pitts neuron model and extend them to a two-dimensional version. In each local receptive field, there are four separate color-orientation jointly selective cells responsible for orientation detection, and their optimal responsive color corresponds to the central location's color. Every local region connects such a set of cells. Subsequently, by these sets of these cells, we can collect all local information and obtain the global orientation according to the local activations. The type of local orientation angle recognized the most corresponds to the global orientation. Finally, a mechanism-based artificial visual system (AVS) is implemented. Several simulations and comparative experiments are provided to verify the effectiveness and generalization of the proposed orientation detection scheme and the superiority of the AVS to popular classification networks in orientation detection tasks. In addition, the feature extraction ability of AVS is shown to accelerate the learning and noise immunity of neural networks.

DOI： 10.1016/j.knosys.2022.109715

Web of Science

Scopus

researchmap

Language：English   Publishing type：Research paper (scientific journal)  

Research on dendrites has been conducted for decades, providing valuable information for the development of dendritic computation. Creating an ideal neuron model is crucial for computer science and may also provide robust guidance for understanding our brain's underlying mechanisms and principles. This paper aims to review the related studies regarding a newly emerging, non-spiking and biologically inspired model, the dendritic neuron model (DNM). By mimicking the biological phenomena of neurons in vivo, the DNM incorporates a neural pruning scheme to eliminate superfluous synapses and dendrites, simplifying its architecture and forming unique neuron morphology for a specific task. Furthermore, the simplified structure can be transformed into logic circuits consisting of the comparators and logic AND, OR and NOT gates, without sacrificing model accuracy. The rapidity of binary operations in hardware implementation gives the DNM a distinct advantage to handle high-speed data streams. The advent of the big data era has led to an exponential explosion in the amount and variety of available information. The appealing properties of the DNM lead us to believe that it is worthy of more attention and that it might be a promising data mining technique. This article presents an in-depth analysis of the pruning and transformation mechanisms and a comprehensive review of the learning algorithms and real-world applications of the DNM. It also presents an empirical comparison of the optimization performance of different algorithms. Finally, we outline some critical issues and future works of the DNM. All the source code of DNM is available at http://www.dnm.net.cn/.

DOI： 10.1016/j.neucom.2021.08.153

Web of Science

Scopus

researchmap

Language：English   Publishing type：Research paper (scientific journal)  

The visual system plays a vital role when the brain receives and processes information. Approximately ninety percent of the information received by the brain comes from the visual system, and motion detection is a crucial part of processing visual information. To further understand the generation of direction selectivity, we propose a novel apparent motion detection mechanism using direction-selective ganglion cells (DSGCs). Considering the simplicity of neural computation, each neuron is responsible for detection in a specific direction. For example, eight neurons are employed to detect movements in eight directions, and local information is collected by scanning. The global motion direction is obtained according to the degree of activation of the neurons. We report that this method not only has striking biological similarities with hypercomplex retinal cells, but can also make accurate discriminations. The pioneering mechanism may lead to a new technique for understanding more complex principles of the visual nervous system.

DOI： 10.1016/j.knosys.2022.108205

Web of Science

Scopus

researchmap

Language：English   Publishing type：Research paper (scientific journal)  

The scale-free network is an important type of complex network. The node degrees in a scale-free network follow the power-law distribution. In the skeleton of a scale-free network, there exists a few nodes which own huge neighborhood size and play an important role in information transmission of the entire network, while most of the network nodes have few connections whose influences of information exchange are limited to a relatively low level. In this paper, we introduce a scale-free population topology into the cuckoo search (CS) algorithm to propose a novel variant, which is termed the scale-free cuckoo search (SFCS) algorithm. Unlike other CS algorithms where the individuals exchange information randomly, two properties of the scale-free network can improve the SFCS algorithm in two aspects: the possibility that the information of competent individuals quickly floods the whole population is reduced significantly, which guarantees population diversity; and the corrupt individuals can learn from competent individuals with greater probability, which is beneficial for convergence. Thus, SFCS can obtain a better trade-off between exploitation and exploration during the search process. To evaluate the effectiveness of the proposed SFCS, 58 benchmark functions with different dimensions (10-D, 30-D, and 50-D), and 21 real-world optimization problems are employed in our experiment. We compare SFCS with the basic CS algorithm, two CS variants, and five state-of-the-art optimization algorithms, and the experimental results and statistical analysis verify the superiority of SFCS in terms of solution quality and convergence speed. Furthermore, we compare SFCS with a scale-free fully informed particle swarm optimization algorithm (SFIPSO) and the results prove our scale-free idea is effective despite its simplicity. We also introduce the scale-free population topology into the differential evolution (DE) and the firefly algorithm (FA) and the experimental results show that the scale-free population topology enhance the search ability of the DE and FA. These lead us to believe that the scale-free population topology may be a new technique for improving the performance of the population-based algorithms.

DOI： 10.1016/j.eswa.2021.116049

Web of Science

Scopus

researchmap

Authorship：Lead author   Publishing type：Research paper (international conference proceedings)  

As a recurrent spiking neural network, liquid state machine (LSM) has attracted more and more attention in neuromorphic computing due to its biological plausibility, computation power, and hardware implementation. However, the neural architecture of LSM, such as hidden neuron number, synaptic density, percentage connectivity, and connection state, has significant impact on its model performance. Manually defining a neural architecture will be ineffective and laborious in most cases. Therefore, based on a state-of-the-art differential evolution algorithm, an evolutionary neural architecture design methodology is proposed to automatically build suitable model topologies for LSM in this study, without any prior knowledge. The effectiveness of the proposed method has been validated on commonly-used image classification tasks.

DOI： 10.1109/ICASSP43922.2022.9747040

Web of Science

Scopus

researchmap

Language：English   Publishing type：Research paper (scientific journal)  

DOI： 10.1016/j.knosys.2021.107509

Language：English   Publishing type：Research paper (scientific journal)  

DOI： 10.1016/j.asoc.2021.107683

Language：English   Publishing type：Research paper (scientific journal)  

DOI： 10.1016/j.engappai.2020.103627

Language：English   Publishing type：Research paper (scientific journal)  

DOI： 10.1016/j.ins.2018.09.034

Publisher：Electronics (Switzerland)  

Pneumonia has long been a significant concern in global public health. With the advancement of convolutional neural networks (CNNs), new technological methods have emerged to address this challenge. However, the application of CNNs to pneumonia diagnosis still faces several critical issues. First, the datasets used for training models often suffer from insufficient sample sizes and imbalanced class distributions, leading to reduced classification performance. Second, although CNNs can automatically extract features and make decisions from complex image data, their interpretability is relatively poor, limiting their widespread use in clinical diagnosis to some extent. To address these issues, a novel weighted cross-entropy loss function is proposed, which calculates weights via an inverse proportion exponential function to handle data imbalance more efficiently. Additionally, we employ a transfer learning approach that combines pretrained CNN model parameter fine-tuning to improve classification performance. Finally, we introduce the gradient-weighted class activation mapping method to enhance the interpretability of the model’s decisions by visualizing the image regions of focus. The experimental results indicate that our proposed approach significantly enhances CNN performance in pneumonia diagnosis tasks. Among the four selected models, the accuracy rates improved to over 90%, and visualized results were provided.

DOI： 10.3390/electronics13152929

Web of Science

Scopus

Publishing type：Research paper (scientific journal)  

The protein–ligand docking problem plays an essential role in structure-based drug design. The challenge for a protein–ligand docking method is how to execute an efficient conformational search to explore a well-designed scoring function. In this study, we improved the artificial bee colony (ABC) algorithm and proposed an approach called ABC-EDM to solve the protein–ligand docking problem. ABC-EDM employs the scoring function of the classical AutoDock Vina to evaluate a solution during docking simulation. ABC-EDM adopts the search framework of the canonical ABC algorithm to execute conformational search. By further investigating the characteristics of the protein–ligand docking problem, a proprietary search mechanism inspired by estimation of distribution algorithm, i.e., estimation of distribution mechanism (EDM), is designed to enhance the performance of ABC-EDM. To verify the effectiveness of the proposed ABC-EDM, we compare it with three variants of the ABC algorithm, three evolutionary computation algorithms, and AutoDock Vina. The experimental results show that ABC-EDM can effectively solve the protein–ligand docking problem, and it can achieve a success rate 5% higher than AutoDock Vina on the GOLD dataset. This study reveals that taking advantage of problem-specific information about the protein–ligand docking problem to enhance a docking method contributes to solving this problem.

DOI： 10.1016/j.asoc.2024.111732

Web of Science

Scopus

researchmap

Publishing type：Research paper (scientific journal)  

In recent years, Siamese network-based trackers have achieved significant improvements in real-time tracking. Despite their success, performance bottlenecks caused by unavoidably complex scenarios in target-tracking tasks are becoming increasingly non-negligible. For example, occlusion and fast motion are factors that can easily cause tracking failures and are labeled in many high-quality tracking databases as challenging attributes. In addition, Siamese trackers tend to suffer from high memory costs, which restricts their applicability to mobile devices with tight memory budgets. To address these issues, we propose a Specialized teachers Distilled Siamese Tracker (SDST) framework to learn a student tracker, which is small, fast, and has enhanced performance in challenging attributes. SDST introduces two types of teachers for multi-teacher distillation: general teacher and specialized teachers. The former imparts basic knowledge to the students. The latter is used to transfer specialized knowledge to students, which helps improve their performance in challenging attributes. For students to efficiently capture critical knowledge from the two types of teachers, SDST is equipped with a carefully designed multi-teacher knowledge distillation model. Our model contains two processes: general teacher-student knowledge transfer and specialized teachers-student knowledge transfer. Extensive empirical evaluations of several popular Siamese trackers demonstrated the generality and effectiveness of our framework. Moreover, the results on Large-scale Single Object Tracking (LaSOT) show that the proposed method achieves a significant improvement of more than 2–4% in most challenging attributes. SDST also maintained high overall performance while achieving compression rates of up to 8x and framerates of 252 FPS and obtaining outstanding accuracy on all challenging attributes.

DOI： 10.1007/s00138-024-01578-4

Web of Science

Scopus

researchmap

Authorship：Corresponding author   Publishing type：Research paper (scientific journal)  

Assessing building energy consumption is of paramount significance in sustainability and energy efficiency (EE) studies. The development of an accurate EE prediction model is pivotal for optimizing energy resources and facilitating effective building planning. Traditional physical modeling approaches are encumbered by high complexity and protracted modeling cycles. In this paper, we introduce a novel evolutionary dendritic neural regression (EDNR) model tailored to forecasting residential building EE. Acknowledging the vast landscape and complexity of the EDNR weight space, coupled with the inherent susceptibility of traditional optimization algorithms to local optima, we propose a complex network-guided strategy-based differential evolution algorithm for training the EDNR model. This strategy adeptly strikes a balance between exploration and exploitation during the search process, significantly enhancing the predictive and generalization capacities of EDNR. To our knowledge, this study represents the inaugural application of dendritic neural regression in real-world prediction scenarios. Extensive experimental findings demonstrate the efficacy of EDNR in accurately predicting building EE with commendable performance. Furthermore, the results of two nonparametric statistical tests affirm the validity and stability of EDNR. Consequently, our proposed methodology exhibits high potential and competitiveness in machine learning applications within the energy domain.

DOI： 10.3390/electronics13101803

Web of Science

Scopus

researchmap

Authorship：Corresponding author   Publishing type：Research paper (scientific journal)  

Studying how objects are positioned is vital for improving technologies like robots, cameras, and virtual reality. In our earlier papers, we introduced a bio-inspired artificial visual system for orientation detection, demonstrating its superiority over traditional systems with higher recognition rates, greater biological resemblance, and increased resistance to noise. In this paper, we propose a hardware-based orientation detection system (ODS). The ODS is implemented by a multiple dendritic neuron model (DNM), and a neuronal pruning scheme for the DNM is proposed. After performing the neuronal pruning, only the synapses in the direct and inverse connections states are retained. The former can be realized by a comparator, and the latter can be replaced by a combination of a comparator and a logic NOT gate. For the dendritic function, the connection of synapses on dendrites can be realized with logic AND gates. Then, the output of the neuron is equivalent to a logic OR gate. Compared with other machine learning methods, this logic circuit circumvents floating-point arithmetic and therefore requires very little computing resources to perform complex classification. Furthermore, the ODS can be designed based on experience, so no learning process is required. The superiority of ODS is verified by experiments on binary, grayscale, and color image datasets. The ability to process data rapidly owing to advantages such as parallel computation and simple hardware implementation allows the ODS to be desirable in the era of big data. It is worth mentioning that the experimental results are corroborated with anatomical, physiological, and neuroscientific studies, which may provide us with a new insight for understanding the complex functions in the human brain.

DOI： 10.3390/electronics13071367

Web of Science

Scopus

researchmap

Language：English   Publishing type：Research paper (scientific journal)  

DOI： https://doi.org/10.3390/bios14030153

Publisher：CEUR Workshop Proceedings  

In recent years, due to the impact of Coronavirus disease (COVID-19), digital platforms have developed rapidly and accumulated a large amount of data. To better utilize the comprehensive and diverse data stored in online platforms for data mining, such as learning behavior analysis or performance prediction, and to provide guidance and valuable feedback for educator became more important. For the current analysis of learning behaviors by time series data with DNN method, the interpretability is not enough. This paper proposes a method based on the simultaneous use of learning behaviors and learning materials to obtain the representation of learned knowledge, and through multiple cross-validations, the effect of this knowledge representation has a certain improvement on the original data, and the interpretability can promote the feedback function.

Scopus

Publisher：Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)  

Learning analytics, blending education theory, psychology, statistics, and computer science, utilizes data about learners and their environments to enhance education. Artificial Intelligence advances this field by personalizing learning and providing predictive insights. However, the opaque ’black box’ nature of AI decision-making poses challenges to trust and understanding within educational settings. This paper presents a novel visual analytics method to predict whether a student is at risk of failing a course. The proposed method is based on a dendritic neuron model (DNM), which not only performs excellently in prediction, but also provides an intuitive visual presentation of the importance of learning behaviors. It is worth emphasizing that the proposed DNM has a better performance than recurrent neural network (RNN), long short term memory network (LSTM), gated recurrent unit (GRU), bidirectional long short term memory network (BiLSTM) and bidirectional gated recurrent unit (BiGRU). The powerful prediction performance can assist instructors in identifying students at risk of failing and performing early interventions. The importance analysis of learning behaviors can guide students in the development of learning plans.

DOI： 10.1007/978-981-97-5495-3_14

Scopus

Publishing type：Research paper (international conference proceedings)  

In recent years, Siamese network-based trackers brought new vitality to the visual object tracking field. However, tracking tasks have always been troubled by complex scenarios. As Siamese trackers become more powerful, the performance bottlenecks caused by complex scenarios become more and more non-negligible. Occlusion is the most common and challenging complex scenario that can easily cause tracking failures. Some high-quality tracking databases provide visible ratio labels to describe occlusion in more detail. In addition, high-performance Siamese trackers can not run efficiently on resource-limited devices due to their high memory cost and complexity. To address these issues, we propose an Adaptive Multi-teacher Knowledge Distillation (AMKD) model to distill lightweight tracker, which is fast and achieves satisfactory performance in low visible ratios scenarios. In AMKD, we adopt the teacher model to transfer adequate knowledge to student. Furthermore, to extract visibility-based knowledge from visible ratios labeled data and transfer it to student efficiently, we introduced assistant teachers which are customed to overcome low visible ratios scenarios. For multiple assistant teachers transfer knowledge to student more efficiently and effectively, the AMKD is equipped with an Adaptive Selection Mechanism (ASM). Experiments of several Siamese trackers on high-quality dataset GOT-10K demonstrated the effectiveness of our method. Moreover, the AMKD distilled student achieve 9 times of compression rates and 6 times of speed up reach 181 FPS while improving accuracy in low visible ratios scenarios and obtaining favorable overall performance.

DOI： 10.1109/ISAS61044.2024.10552525

Scopus

researchmap

Publisher：Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)  

In educational settings, a challenge is the lack of linked and labeled data, hindering effective analysis. The integration of ontology facilitates the formulation of educational knowledge concepts, student behaviors, and their relations. Traditional ontology creation requires deep domain knowledge and significant manual effort. However, advancements in Large Language Models (LLMs) have offered a novel opportunity to automate and refine this process. In this paper, we propose an LLMs-driven educational ontology learning approach aimed to enhance student performance predictions. We leverage LLMs to process lecture slide texts to identify knowledge concepts and their interrelations, while question texts are used to associate them with the concepts they assess. This process facilitates the generation of the educational ontology that links knowledge concepts and maps to student interactions. Additionally, we deploy a dual-branch Graph Neural Network (GNN) with distance-weighted pooling to analyze both global and local graph information for student performance prediction. Our empirical results demonstrate the effectiveness of using LLMs for ontology-based enhancements in educational settings.

DOI： 10.1007/978-981-97-5498-4_5

Scopus

Publisher：Communications in Computer and Information Science  

This study investigates the effects of generative AI on the knowledge comprehension of university students, focusing on the use of ChatGPT strategies. Data from 81 junior students who used the ChatGPT worksheet were collected and analyzed. Path analysis revealed complex interactions between ChatGPT strategy use, e-book reading behaviors, and students’ prior perceived understanding of concepts. Students’ prior perceived understanding and reading behaviors indirectly affected their final scores, mediated by the ChatGPT strategy use. The mediation effects indicated that reading behaviors significantly influenced final scores through ChatGPT strategies, indicating the importance of the interaction with learning materials. Further regression analysis identified the specific ChatGPT strategy related to verifying and comparing information sources as significantly influenced by reading behaviors and directly affecting students’ final scores. The findings provide implications for practical strategic guidance for integrating ChatGPT in education.

DOI： 10.1007/978-3-031-64315-6_12

Scopus

Publishing type：Research paper (international conference proceedings)  

With the advent of generative artificial intelligence (AI), the scope of data analysis, prediction of performances, real-time feedback, etc. in learning analytics has widened. The purpose of this study is to explore the possibility of using generative AI to analyze educational data. Moreover, the performances of two large language models (LLMs): GPT-4 and text-davinci-003, are compared with respect to different types of analyses. Additionally, a framework, LangChain, is integrated with the LLM in order to achieve deeper insights into the analysis, which can be beneficial for beginner data scientists. LangChain has a component called an agent, which can help study the analysis being performed step-by-step. Furthermore, the impact of the OpenLA library, which pre-processes the data by calculating the number of reading seconds of students, counting the number of operations performed by students, and making page-wise behavior of each student, is also studied. Besides, factors with the most significant impact on students’ performances were also discovered in this analysis. The results show that GPT-4, when using the data pre-processed by OpenLA, provides the best analysis in terms of both, the accuracy of the final answer, and the step-by-step insights provided by LangChain’s agent. Also, we learn the significance of reading time and interactions used (Add marker, bookmark, memo) by students in predicting grades.

Scopus

researchmap

Publishing type：Research paper (international conference proceedings)  

Recently, banks are constantly facing the problem of customers churning. Customer churn not only leads to a decline in bank funds and profits but also reduces its credit capacity and affects the bank’s operational management. As an important component of Customer Relationship Management, predicting customer churn has been increasingly urgent. Inspired by biological neurons, we build up a dendritic neural regression model (DNRM) with four layers, namely the synaptic layer, the dendritic layer, the membrane layer, and the soma layer for bank customer churn prediction. To pursue better prediction performance in this experiment, the Chicken Swarm Optimization (CSO) algorithm is defined as the training algorithm of DNRM. With the ability to balance exploration and exploitation, CSO is implemented to optimize and improve the accuracy of the DNRM. In this paper, we propose a novel dendritic neural regression model called CSO-DNRM for churn prediction, and the experimental results are based on a benchmark dataset from Kaggle. Compared with other algorithms and models, our proposed model obtains the highest accuracy of 92.27% and convergence speed in customer churn prediction. Due to the novel bionic algorithms and the pruning function of the model, it is evident that our proposed model has advantages in accuracy and computational speed in the field of customer churn prediction and can be widely applied in commercial bank customer relationship management.

DOI： 10.1007/978-981-99-8184-7_20

Scopus

researchmap

Publisher：2024 IEEE Congress on Evolutionary Computation, CEC 2024 - Proceedings  

In the domain of artificial neural networks, the comprehensive understanding and optimization of neuron dynamics are crucial. The Dendritic Neuron Model (DNM), noted for its distinct architecture and data processing capabilities, exemplifies this. However, the sophistication of the DNM leads to complexities in hyperparameter tuning. Notably, this complexity manifests in the way parameter changes can alter the dimensions of the solution space, a prime example of the Metameric Variable-length Problem. In this study, we have innovatively integrated chaotic maps into the gene expression mechanisms of Evolutionary Algorithms, enabling the incorporation of all DNM hyperparameters into a singular algorithmic framework. This integration allows for iterative adjustments within a variable-dimensional solution space, representing an evolving neuron model that streamlines the tuning process. Our approach, tested against benchmark datasets from the UCI Machine Learning Repository, demonstrates significant improvements in the DNM's performance, highlighting the effectiveness of incorporating biological chaos phenomena into neural network optimization.

DOI： 10.1109/CEC60901.2024.10612087

Scopus

Publishing type：Research paper (scientific journal)  

DOI： 10.1109/ACCESS.2024.3429554

Web of Science

Scopus

researchmap

Publishing type：Research paper (international conference proceedings)  

Handwritten notes by students are a critical data source reflecting their learning status. Traditionally, educators have had to review these notes individually to gauge students' comprehension and mastery of course material. However, this method is time-consuming, inefficient, and often fails to capture and quantify students' learning progress and challenges comprehensively. With advancements in technology, especially in text recognition and machine learning, new avenues have opened up for automating this review process. This allows for a more efficient and systematic analysis of students' learning situations. This study aims to explore how these technologies can be utilized to automatically identify and analyze key information in students' handwritten notes. The goal is to assess students' learning outcomes and understanding more effectively. By automating the recognition and analysis of handwritten notes, the study seeks to provide educators with a powerful tool to monitor students' learning progress more accurately, identify learning obstacles, and offer personalized feedback and support based on individual needs. This paper introduces a technology integrating Attention Multi-task U-Net and GPT-4 for extracting data from handwritten notes. The method facilitates better understanding and analysis of student notes, offering teachers precise learning data and aiding students in receiving personalized learning support. The study underscores its potential in educational technology, particularly in improving teaching quality and student learning outcomes.

DOI： 10.1109/ISAS61044.2024.10552516

Scopus

researchmap

Publishing type：Research paper (international conference proceedings)  

In recent years, with the rapid development of neural networks, visual object tracking is becoming increasingly important in real-world applications such as camera drones and driving assistant systems, even self-driving technology. Neural network-based Siamese trackers achieve satisfactory accuracy in the object tracking field and stand out. However, high-performance Siamese trackers are often designed to be complex and heavy, which hinders their application on resource-limited mobile devices. Thus, compressing the neural network-based tracker to make it lightweight and efficient without obvious performance degradation is of great significance. Inspired by the outstanding work of RepVGG which focuses on compressing multi-branches neural networks without any accuracy cost, we propose the RepSiamses Tracker (RST) which is extremely lightweight and achieves very high tracking speed. In RST, the RepVGG-based backbone depth can be adapted to the different target hardware which benefits from the high flexibility of RepVGG. The experimental results on the high-quality benchmark VOT2018 and LaSOT show that RST achieves satisfactory accuracy and extremely high tracking speed on GPU. More impressively, RST is able to run on the CPU at a hyper-real-time of 69 fps and with very little memory cost of 16.4 MB. Such high tracking speed and low memory cost can bridge the gap between academic algorithms and real-world applications.

DOI： 10.1109/ISAS61044.2024.10552459

Scopus

researchmap

Publishing type：Research paper (international conference proceedings)  

In recent years, due to the impact of Coronavirus disease (COVID-19), digital platforms have developed rapidly and accumulated a large amount of data. To better utilize the comprehensive and diverse data stored in online platforms for data mining, such as learning behavior analysis or performance prediction, and to provide guidance and valuable feedback for educator became more important. For the current analysis of learning behaviors by time series data with DNN method, the interpretability is not enough. This paper proposes a method based on the simultaneous use of learning behaviors and learning materials to obtain the representation of learned knowledge, and through multiple cross-validations, the effect of this knowledge representation has a certain improvement on the original data, and the interpretability can promote the feedback function.

Scopus

researchmap

Publishing type：Research paper (scientific journal)  

As the most studied sensory system, the visual system plays an important role in our understanding of brain functions. Biological researchers have divided the nerve cells in the retina into dozens of visual channels carrying various characteristics based on visual features. Although orientation-selective cells have been identified in the retinas of various animals, the specific neural circuits of such cells have been controversial. In this study, a new simple and efficient orientation detection model based on the perceptron is proposed to restore the neural circuitry of orientationselective cells in the retina. The performance of this model is experimentally compared with that of the convolutional neural network for image orientation recognition, and the results verify that the proposed model offers very good orientation detection. The proposed perceptron-based orientation detection model provides a new perspective to explain the neural circuits of orientation-selective cells.

DOI： 10.3934/era.2023103

Web of Science

Scopus

researchmap

Publishing type：Research paper (scientific journal)  

The mechanism of motion direction detection for direction selective ganglion cells (DSGCs) is still not well-understood and under debate. Recent studies have elaborated the critical experimental evidence that the starburst amacrine cells (SACs) can trigger off the null-direction inhibition to DSGCs. In this study, a simple but effective neural model is introduced for the SACs to solve the motion direction detection problems, based on greyscale images in the visual scene. Virtual simulations demonstrate that the neural model is capable of detecting the motion direction of objects with different shapes, sizes, greyscales, and positions efficiently. To further demonstrate the feasibility and effectiveness of the model, the performance of the proposed model is compared with traditional artificial neural networks (ANNs). Experimental results show it can completely beat ANNs on motion direction detection problems, in terms of recognition accuracy, noise immunity, computational and learning costs, biological soundness, and reasonability.

DOI： 10.1504/IJBIC.2023.130560

Web of Science

Scopus

researchmap

Publishing type：Research paper (scientific journal)  

With the wide application of imaging modalities such as X-ray and Computed Tomography (CT) in clinical practice, Content-based Medical Image Retrieval (CBMIR) has become a current research hotspot. Related studies have shown that hash-based image retrieval algorithms can retrieve relevant images faster and more accurately than traditional image retrieval methods. Therefore, in this paper, we propose a novel deep hashing method for medical image retrieval, called DenseHashNet. Specifically, we first use DenseNet to extract the original image features, and introduce the Spatial Pyramid Pooling (SPP) layer after the last Dense Block so that features at different scales can be extracted and multi-scale features fused with information from multiple regions. Then, the output of the SPP layer is subjected to Power-Mean Transformation (PMT) operation to enhance the nonlinearity of the model and improve the performance of the model. Finally, we map the output of PMT to hash codes through fully connected layers. Experimental results show that our method achieves better performance, compared with some representative methods.

DOI： 10.1109/JRFID.2022.3209986

Web of Science

Scopus

researchmap

Language：English   Publishing type：Research paper (scientific journal)  

DOI： 10.3390/atmos12121647

Language：English   Publishing type：Research paper (scientific journal)  

DOI： 10.3390/a14080239

Language：English   Publishing type：Research paper (scientific journal)  

DOI： 10.3390/electronics10040373

Language：English   Publishing type：Research paper (scientific journal)  

Language：English   Publishing type：Research paper (scientific journal)  

Event date： 2021.8

Language：English   Presentation type：Oral presentation (general)  

Country：China  

Event date： 2021.7

The research on orientation recognition of visual systems is of great significance to improve our understanding of biological vision and the human brain. However, the mechanism of orientation recognition in the visual system is still unknown. In this research, we propose a new mechanism to detect the orientation of the object. The orientation detection ganglion neuron is applied to infer the global detection according to the local information. This mechanism is simulated by a dendritic neuron model and a series of experiments are conducted to verify the effectiveness. The experimental results show the ability of this mechanism to detect the orientation of an object regardless of the size and position, which is consistent with most known physiological experimental results.

researchmap

Event date： 2021.3

Language：English  

Country：Other  

Event date： 2021

Advances in the understanding of dendrites promote the development of dendritic computation. For decades, the researchers are committed to proposing an appropriate neural model, which may feedback the research on neurons. This paper aims to employ an effective metaheuristic optimization algorithm as the learning algorithms to train the dendritic neuron model (DNM). The powerful ability of the backpropagation (BP) algorithm to train artificial neural networks led us to employ it as a learning algorithm for a conventional DNM, but this also inevitably causes the DNM to suffer from the drawbacks of the algorithm. Therefore, a metaheuristic optimization algorithm, named the firefly algorithm (FA) is adopted to train the DNM (FADNM). Experiments on twelve datasets involving classification and prediction are performed to evaluate the performance. The experimental results and corresponding statistical analysis show that the learning algorithm plays a decisive role in the performance of the DNM. It is worth emphasizing that the FADNM incorporates an invaluable neural pruning scheme to eliminate superfluous synapses and dendrites, simplifying its structure and forming a unique morphology. This simplified morphology can be implemented in hardware through logic circuits, which approximately has no effect on the accuracy of the original model. The hardwareization enables the FADNM to efficiently process high-speed data streams for large-scale data, which leads us to believe that it might be a promising technology to deal with big data.

researchmap

Event date： 2020.10

Language：English  

Country：Other  

Event date： 2020

This paper proposes a novel plastic neural model (PNM) at the single neuron level and a specified learning algorithm to train it. The dendritic structure of PNM presents its diversity to fulfill each particular task. During the training process, PNM divides the Euclidean space of the training instances into several appropriate hypercubes, which have the same dimensional number. And then, each hypercube is transformed into a corresponding dendritic branch in PNM. A suitable dendritic structure of PNM has been proved to have powerful computational capabilities to solve the classification problems. Both theoretical analysis and empirical study of the proposed model are demonstrated in this paper. Five benchmark problems are employed to verify the effectiveness of PNM in our experiment. The results have verified that PNM can provide competitive classification performance compared with several widely-used classifiers.

researchmap

Event date： 2020

Inspired by the dynamic dendritic architecture of biological neurons, the approximate logic neuron model (ALNM) is a novel model recently proposed by us. ALNM owns four layers, namely, the synaptic layer, the dendritic layer, the membrane layer, and the cell body. Through neural pruning function, the model can omit useless synapses and unnecessary branches of dendrites after the training process. In other words, it will form a unique and simplified dendritic structure for each particular classification task. Further, the simplified dendritic structure can be completely substituted by logic circuits, which makes ALNM be capable of running in hardware. However, although it has satisfactory performances to solve classification problems, it still suffers from some disadvantages owing to its learning algorithm, named batch gradient descent (BGD) algorithm. It is because that, using all the training data for each iteration is time-consuming and it is unsuitable for large scale problems. In addition, BGD cannot adaptively adjust the learning rate during the whole training process, which will converge slowly in the neighborhood of saddle points, and oscillate in the steep region of gradient space. To settle the above issues, we propose a novel stochastic adaptive gradient descent (SAGD) algorithm, which uses stochastic gradient descent information and adaptively adjusts the learning rate, to improve the classification performances of ALNM. In our experiments, ALNM trained by the new algorithm is evaluated on three benchmark classification datasets, and experimental results demonstrate that it performs significantly better than the original model in terms of accuracy and convergence rate.

researchmap

Event date： 2019.12

Gravitational search algorithm (GSA) has attracted more and more attention in dealing with complex optimization problems. However, it still suffers from some major drawbacks, such as poor exploitation. This paper has proposed a novel self-adaptive chaotic search mechanism embedded GSA (SA-GSA), in which different chaotic maps are utilized to guide the local search in a self-adaptive way. Specifically, instead of defining the search range randomly and arbitrarily, the distance between distinct individuals has been utilized in the current population as the search range. Thus the search range will decrease synchronously according to the convergence speed of the population, which is thought to improve the exploitation ability of GSA effectively. To evaluate the performance of SA-GSA, we compare it with classic GSA and the classic chaotic GSA (CGSA) on 23 benchmark optimization problems. The experimental results and statistic analysis verify that SA-GSA is superior to its competitors in terms of convergence speed and solution accuracy rate.

researchmap