Search

Keypoint-based deep learning detectors have proven highly effective in object detection tasks by predicting specific keypoints to determine object classification and location. Examples include CornerNet, CenterNet, ExtremeNet, RetinaNet, FCOS, and ObjectBox. Despite their strengths, these methods are particularly susceptible to class imbalance, which can result in poor detection performance for less frequent classes. Popular solutions such as hard sampling, soft sampling, and sampling-free methods have been proposed to tackle this issue. However, these approaches often have inherent limitations, including sensitivity to hyperparameters and neglecting the gradient dynamics of the loss function. To address these challenges, this paper proposes Adaptive Focal Loss (AFL), which combines the strengths of Focal Loss (FL) and Class-Balanced Loss (CBL) while introducing an Adaptive Gradient Function. This function is specifically designed to mitigate the impact of large gradients during the early stages of training. Extensive experiments demonstrate that AFL significantly outperforms classical sampling methods across key metrics on the MS COCO, LVIS, and PASCAL VOC datasets. Notably, on the LVIS dataset, AFL achieves an average improvement of over 3% $AP_{r}$ compared to the baseline Focal Loss across multiple keypoint-based detectors, showcasing its effectiveness in enhancing rare class accuracy. Furthermore, AFL delivers substantial improvements in the $AP_{0.5\sim 0.95}$ metric, surpassing the baseline Focal Loss by an average of over 2%, 1%, and 0.2% on the LVIS, MS COCO, and PASCAL VOC datasets, respectively. This highlights AFL’s capability to enhance the overall accuracy of foreground objects. By providing a more balanced and robust solution to class imbalance, AFL demonstrates superior performance in challenging scenarios, making it a valuable advancement in keypoint-based detection.

Financial investment data decision-making is part of corporate financial management. Deep Learning Algorithms (DLAs) have represented significant and meaningful promise in many financial investment applications, leveraging their capability and verity to automatically learn hierarchical interpretations from large and complicated datasets. The major influence of this application is to derive the most preferable and the most dominant kind of deep learning architecture, used in financial investment data analysis. In this study, we concentrate on evaluating some new conditional aggregation operators based on Aczel-Alsina and power operators under the presence of Type-2 Fuzzy (T2F) DLAs. For this, we derive the Aczel-Alsina operational laws for T2F values and also simplify their related results with the help of some suitable examples. Thus, we evaluate the novel theory of conditional aggregation operators, such as the T2F Aczel-Alsina Weighted Power Averaging (T2FAAWPA) operator and the T2F Aczel-Alsina Weighted Power Geometric (T2FAAWPG) operator. Some properties are the above two proposed operators are also initiated. Additionally, we analyze the problems of DLAs for financial investment data analysis under the presence of the proposed operators. For this, we discuss the Multi-Attribute Decision-Making (MADM) procedure based on initiated operators for T2F uncertainty. Finally, we illustrate or demonstrate some numerical examples for evaluating the comparison between the proposed theory with some existing techniques to show the supremacy and validity of the initiated techniques for financial investment data decision-making.

In this article, a maiden attempt have been taken for the online detection of faults, classification of faults, and identification of the fault locations of a grid-connected Micro-grid (MG) system. A deep learning algorithm-based Long Short Term Memory (LSTM) network is proposed, for the first time, for the online detection of faults and their classifications of the considered MG system to overcome the issues that persist in the existing algorithms. Also, a combination of an LSTM network and feed-forward neural network (FFNN) with a back-propagation algorithm (BPA) is proposed to identify the exact locations of the faults since the identification of fault locations is more challenging than fault categorizations. To select a suitable deep learning network with multiple hidden layers for achieving the aforesaid objectives, a rigorous analysis has been done. To study the accuracy of the proposed techniques, different types of faults with different parameters are considered in this paper. An extensive simulation has been done in MATLAB/Simulink platform to study the performance of the system with the proposed techniques. To validate the effectiveness of the proposed techniques, the entire system is implemented in the real-time platform using the OPAL-RT digital simulator. Comparison has also been done for the results obtained using ANN and proposed techniques. The results show that the proposed techniques based on the deep learning network effectively detect, classify, and identify the location of different faults of an MG system with acceptable performances.

Some teenagers actively participate in cyberbullying, which is a pattern of online harassment of others. Many teenagers are unaware of the risks posed by cyberbullying, which can include depression, self-harm, and suicide. Because of the serious harm it can cause to a person’s mental health, cyberbullying is an important problem that needs to be addressed. This research aimed to develop a technique to identify the severity of bullying using a deep learning algorithm and fuzzy logic. In this task, Twitter data (47,733 comments) from Kaggle were processed and analyzed to flag cyberbullying comments. The comments embedded by Keras were fed into a long short-term memory network, composed of four layers, for classification. After that, fuzzy logic was applied to determine the severity of the comments. Experimental results suggest that the proposed framework provides a suitable solution to detect bulling with values of 93.67%, 93.64%, 93.62% achieved for the accuracy, F1-score, and recall, respectively.

The accessibility of online teaching makes it popular in various teaching scenarios. Most of these researches about online interaction mainly focuse on the communication network stability, facial expression/gesture recognition algorithm and statistical description analysis, and its expertise mainly comes from the fields of computer technology and algorithm engineering. The evaluation and analysis of the effect of online teaching is an important test for the adaptability of such tools in the field of education. However, from the perspective of teachers, there is still a lack of literature on data interpretation after the application of this technology. An experiment based on real online teaching was carried out in this paper. This study uses image recognition technology to process video and extract five kinds of head movement data from dozens of student samples, and then develop statistical description interpretation. Some novel and interesting conclusions indicate that diversified behaviors occurred in real-time online learning. This study obtained data of five high-frequency online learning behaviors, including blinking, yawning, nodding, shaking head and leaving. These behaviors are related to learning state and time. Teaching features, students’ personal characteristics and learning environment have a comprehensive impact on online learning behaviors. The result provides a basis for personalized learning and teaching scheme design in the future. It also helps to enrich online teaching evaluation methods and accelerate the construction of online education framework and rules.

Digital assets have come under various network security threats in the digital age. As a kind of security equipment to protect digital assets, intrusion detection system (IDS) is less efficient if the alert is not timely and IDS is useless if the accuracy cannot meet the requirements. Therefore, an intrusion detection model that combines machine learning with deep learning is proposed in this paper. The model uses the k-means and the random forest (RF) algorithms for the binary classification, and distributed computing of these algorithms is implemented on the Spark platform to quickly classify normal events and attack events. Then, by using the convolutional neural network (CNN), long short-term memory (LSTM), and other deep learning algorithms, the events judged as abnormal are further classified into different attack types finally. At this stage, adaptive synthetic sampling (ADASYN) is adopted to solve the unbalanced dataset. The NSL-KDD and CIS-IDS2017 datasets are used to evaluate the performance of the proposed model. The experimental results show that the proposed model has better TPR for most of attack events, faster data preprocessing speed, and potentially less training time. In particular, the accuracy of multi-target classification can reach as high as 85.24% in the NSL-KDD dataset and 99.91% in the CIC-IDS2017 dataset.

Humans learn incrementally from sequential experiences throughout their lives, which has proven hard to emulate in artificial neural networks. Incrementally learning tasks causes neural networks to overwrite relevant information learned about older tasks, resulting in ‘Catastrophic Forgetting’. Efforts to overcome this phenomenon often utilize resources poorly, for instance, by growing the network architecture or needing to save parametric importance scores, or violate data privacy between tasks. To tackle this, we propose SPACE, an algorithm that enables a network to learn continually and efficiently by partitioning the learnt space into a Core space, that serves as the condensed knowledge base over previously learned tasks, and a Residual space, which is akin to a scratch space for learning the current task. After learning each task, the Residual is analyzed for redundancy, both within itself and with the learnt Core space. A minimal number of extra dimensions required to explain the current task are added to the Core space and the remaining Residual is freed up for learning the next task. We evaluate our algorithm on P-MNIST, CIFAR and a sequence of 8 different datasets, and achieve comparable accuracy to the state-of-the-art methods while overcoming catastrophic forgetting. Additionally, our algorithm is well suited for practical use. The partitioning algorithm analyzes all layers in one shot, ensuring scalability to deeper networks. Moreover, the analysis of dimensions translates to filter-level sparsity, and the structured nature of the resulting architecture gives us up to 5x improvement in energy efficiency during task inference over the current state-of-the-art.

Network traffic analysis and predictions have become vital for monitoring networks. Network prediction is the process of capturing network traffic and examining it deeply to decide what is the occurrence in the network. The accuracy of analysis and estimation of network traffic are increasingly becoming significant in achieving guaranteed Quality of Service (QoS) in the network. The main aim of the presented research is to propose a new methodology to improve network traffic prediction by using sequence mining. The significance of this important topic lies in the urge to contribute to solving the research problem in network traffic prediction intelligently. We propose an integrated model that combines clustering with existing series models to enhance prediction the network traffic. Clustering granules are obtained using fuzzy c-means to analyze the network data for improving the existing time series. The novelty of the proposed research has used the clustering approach to handle the ambiguity from the entire network data for enhancing the existing time series models. Furthermore, we have suggested using the weighted exponential smoothing model as preprocessing stages for increasing the reliability of the proposed model. In this research paper, machine intelligence proposed to predict network traffic. The machine intelligence is working as pre-processing for enhancing the existing time series models. The machine intelligence combines non-crisp Fuzzy-C-Means (FCM) clustering and the weight exponential method for improving deep learning Long Short-Time Memory(LSTM)and Adaptive Neuro-Fuzzy Inference System (ANFIS)time series models. The ANFIS and LSMT time series models are applied to predict network traffic. Two real network traffic traces were conducted to test the proposed time series models. The empirical results of proposed to enhanced LSTM 97.95% and enhanced ANFIS model is R = 96.78% for cellular traffic data, with respect to the correlation indicator. It is observed that the proposed model outperforms alternative time series models. A comparative prediction results between the proposed model and existing time series models are presented. The comparisons indicate that the presented model outperforms the opponent models; the proposed method optimises the deep learning LSTM and ANFIS time series models. The proposed methodology offers more effective approach to the prediction of network traffic.

With a rapid increase in operating frequency and package complexity, conventional analysis methods cannot efficiently cope with current complex electromagnetic interference (EMI) issues. In this paper, a new method built on a deep neural network (DNN) model is proposed to accurately and rapidly predict the maximum radiated electric field at 3-meters of a wire-bond ball grid array (WB-BGA) package. The key hyper-parameters of the DNN model, such as the learning rate and type of optimizers, are discussed in depth so as to attain optimal performance. Predicted radiation results by the DNN model and the results of the full-wave simulation show good agreement. Once DNN is trained, the prediction time is in the order of milliseconds and the model size is in megabytes, which can acquire the predicted radiation quickly and accurately and save storage space. Furthermore, to prevent the radiation from exceeding requirements, package structures are optimized by adjusting those parameters sensitive to radiation and disregarding insensitive parameters. The sensitivity of these WB-BGA package structural parameters to EMI radiation can be analyzed quickly based on data with different deviations generated by the trained DNN model. The sensitive parameters are adjusted according to their correlation with EMI radiation. The effectiveness and feasibility of the optimization method are verified by the WB-BGA package with two sets of different structural parameters.