Reduce Training Time Research Articles

Granularity Optimization of Travel Trajectory Based on Node2vec: A Case Study on Urban Travel Time Prediction

Intersections are known to cause significant changes in traffic states. However, existing link-level trajectory optimization methods often overlook intersection information, making it challenging to preserve key traffic state features during the optimization process. To address this limitation, a novel approach is proposed that integrates node2vec and K-means algorithms. First, the role of intersections in linking road segments is considered. The node2vec algorithm is employed to capture the deep spatial similarity between links while weakening the adjacency relationship between links before and after intersections. This process generates feature representations for each link. Subsequently, clustering centers are initialized at the intersections, and K-means clustering is applied based on these link feature representations. Through this method, consecutive links within a trajectory that belong to the same cluster are merged, thus optimizing the granularity of the trajectory. Finally, experimental analysis and validation are conducted using link-level travel trajectory data from Shenzhen. The results demonstrate that, under optimal conditions, the mean absolute error (MAE), the mean absolute percentage error (MAPE), and the root mean square error (RMSE) values are reduced by 8.91%, 9.44%, and 8.96%, respectively, while computational efficiency is increased by 30.08%. The proposed trajectory granularity optimization method, which accounts for the existence of intersections, not only effectively retains the key traffic state features from the original trajectory but also significantly reduces training time while improving the model’s prediction accuracy.

Transfer Learning for Wildlife Classification: Evaluating YOLOv8 against DenseNet, ResNet, and VGGNet on a Custom Dataset

This study evaluates the performance of various deep learning models, specifically DenseNet, ResNet, VGGNet, and YOLOv8, for wildlife species classification on a custom dataset. The dataset comprises 575 images of 23 endangered species sourced from reputable online repositories. The study utilizes transfer learning to fine-tune pre-trained models on the dataset, focusing on reducing training time and enhancing classification accuracy. The results demonstrate that YOLOv8 outperforms other models, achieving a training accuracy of 97.39% and a validation F1-score of 96.50‘%. These findings suggest that YOLOv8, with its advanced architecture and efficient feature extraction capabilities, holds great promise for automating wildlife monitoring and conservation efforts.

Separable physics-informed DeepONet: Breaking the curse of dimensionality in physics-informed machine learning

The deep operator network (DeepONet) has shown remarkable potential in solving partial differential equations (PDEs) by mapping between infinite-dimensional function spaces using labeled datasets. However, in scenarios lacking labeled data, the physics-informed DeepONet (PI-DeepONet) approach, which utilizes the residual loss of the governing PDE to optimize the network parameters, faces significant computational challenges, particularly due to the curse of dimensionality. This limitation has hindered its application to high-dimensional problems, making even standard 3D spatial with 1D temporal problems computationally prohibitive. Additionally, the computational requirement increases exponentially with the discretization density of the domain. To address these challenges and enhance scalability for high-dimensional PDEs, we introduce the Separable physics-informed DeepONet (Sep-PI-DeepONet). This framework employs a factorization technique, utilizing sub-networks for individual one-dimensional coordinates, thereby reducing the number of forward passes and the size of the Jacobian matrix required for gradient computations. By incorporating forward-mode automatic differentiation (AD), we further optimize computational efficiency, achieving linear scaling of computational cost with discretization density and dimensionality, making our approach highly suitable for high-dimensional PDEs. We demonstrate the effectiveness of Sep-PI-DeepONet through three benchmark PDE models: the viscous Burgers’ equation, Biot’s consolidation theory, and a parametrized heat equation. Our framework maintains accuracy comparable to the conventional PI-DeepONet while reducing training time by two orders of magnitude. Notably, for the heat equation solved as a 4D problem, the conventional PI-DeepONet was computationally infeasible (estimated 289.35 h), while the Sep-PI-DeepONet completed training in just 2.5 h. These results underscore the potential of Sep-PI-DeepONet in efficiently solving complex, high-dimensional PDEs, marking a significant advancement in physics-informed machine learning.

On the Performance of a Data-Driven Backward Compatible Physics-Informed Neural Network for Prediction of Flow Past a Cylinder

Abstract This paper discusses a physics-informed surrogate model aimed at reconstructing the flow field from sparse datasets under a limited computational budget. A benchmark problem of 2D unsteady laminar flow past a cylinder is chosen to evaluate the performance of the surrogate model. Earlier studies were focused on forward problems with well-defined data. The present study attempts to develop models capable of reconstructing the flow-field data from sparse datasets mirroring real-world scenarios. We demonstrated the performance of data-driven models in reconstructing the flow field and compared the effectiveness of various training methodologies. The proposed surrogate model successfully reconstructed the flow field while also extracting pressure as a latent variable. The proposed surrogate model significantly outperformed data-driven models in accuracy, even under a limited computational budget. Furthermore, transfer learning of parameters of a pretrained model for different Reynolds numbers has reduced training time.

Improving Model Robustness through Hybrid Adversarial Training: Integrating FGSM and PGD Methods

Abstract. With the widespread use of deep learning models in various applications. People are gradually realizing the vulnerability of these models to adversarial attacks. Adversarial training is an effective strategy to defend against adversarial attacks. Based on the advantages and disadvantages of the current mainstream Fast Gradient Sign Method (FGSM) adversarial training and Projected Gradient Descent (PGD) adversarial training, this paper proposes a hybrid adversarial training that integrates FGSM and PGD methods and uses the ResNet-18 model and SVHN dataset for testing. Experimental results show that hybrid adversarial training can effectively reduce training time. Its accuracy on the original data set is higher than that of PGD adversarial training, which is improved by about 2%. The performance when facing FGSM attacks is almost the same as that of single FGSM adversarial training. The performance when facing PGD attacks decreases more significantly, which is about 2% to 3% lower than that of PGD adversarial training. This study not only helps to understand the robustness of hybrid adversarial training to models facing adversarial attacks but also helps in studying new adversarial training strategies.

Real-time Error Compensation Transfer Learning with Echo State Networks for Enhanced Wind Power Prediction

Accurate wind power forecasting is essential for efficient energy management and grid stability, enabling energy providers to balance supply and demand, optimize renewable energy integration, reduce operational costs, and enhance power grid reliability. The Echo State Network (ESN) is widely used for modeling nonlinear dynamic systems due to its simple and rapid training process. However, ESNs can be prone to system errors, leading to inaccurate models when handling high-order nonlinear complexities. To overcome this, we developed the Error Compensation Transfer Learning Echo State Network (ETL-ESN), which combines a computing layer based on ESN and a compensation layer using transfer learning. Our model identifies error auto-correlation as a key factor that increases variance in ESN predictions, and addresses this with an error compensation layer to reduce system errors. We further leverage transfer learning to prevent overfitting within the error domain. Extensive experiments using real-world wind power data demonstrate that the ETL-ESN model reduces training time from 65 s to 2 s compared to LSTM, while lowering MAE by up to 95%. The ETL-ESN achieves a 95% to 98% improvement in prediction accuracy across different turbines compared to traditional models. The code and datasets used in this study are available at GitHub repository https://github.com/zhuyingqin/Error-Transfer-ESN for further research and replication.

A Hyper-Parameter Optimizer Algorithm Based on Conditional Opposition Local-Based Learning Forbidden Redundant Indexes Adaptive Artificial Bee Colony Applied to Regularized Extreme Learning Machine

Finding the best configuration of a neural network’s hyper-parameters may take too long to be feasible using an exhaustive search, especially when the cardinality of the search space has a big combinatorial number of possible solutions with various hyper-parameters. This problem is aggravated when we also need to optimize the parameters of the neural network, such as the weight of the hidden neurons and biases. Extreme learning machines (ELMs) are part of the random weights neural network family, in which parameters are randomly initialized, and the solution, unlike gradient-descent-based algorithms, can be found analytically. This ability is especially useful for metaheuristic analysis due to its reduced training times allowing a faster optimization process, but the problem of finding the best hyper-parameter configuration is still remaining. In this paper, we propose a modification of the artificial bee colony (ABC) metaheuristic to act as parameterizers for a regularized ELM, incorporating three methods: an adaptive mechanism for ABC to balance exploration (global search) and exploitation (local search), an adaptation of the opposition-based learning technique called opposition local-based learning (OLBL) to strengthen exploitation, and a record of access to the search space called forbidden redundant indexes (FRI) that allow us to avoid redundant calculations and track the explored percentage of the search space. We set ten parameterizations applying different combinations of the proposed methods, limiting them to explore up to approximately 10% of the search space, with results over 98% compared to the maximum performance obtained in the exhaustive search in binary and multiclass datasets. The results demonstrate a promising use of these parameterizations to optimize the hyper-parameters of the R-ELM in datasets with different characteristics in cases where computational efficiency is required, with the possibility of extending its use to other problems with similar characteristics with minor modifications, such as the parameterization of support vector machines, digital image filters, and other neural networks, among others.

Artificial intelligence-aided electrochemical sensors for capturing and analyzing fingerprint profiles of medicinal materials

This study explores the application of artificial intelligence-aided electrochemical sensors for authenticating medicinal materials, focusing on sika deer antler cap powder. Utilizing differential pulse voltammetry and graphene-modified screen-printed electrodes, we developed a novel method to capture unique electrochemical fingerprints of authentic, counterfeit, and adulterated samples. Three machine learning models—Support Vector Machine (SVM), Random Forest (RF), and Extreme Learning Machine (ELM)—were evaluated using both full voltammogram and principal component analysis (PCA) reduced features. The SVM model with PCA-reduced features emerged as the optimal approach, achieving a classification accuracy of 97.9 % while reducing training time by 65.6 % (from 3.2 s to 1.1 s) and prediction time by 71.4 % (from 0.07 s to 0.02 s per sample) compared to using full voltammogram features. This reduction in computational complexity was achieved by decreasing the input dimensionality from 601 to 5 features through PCA, while maintaining high classification performance across all sample categories. This model demonstrated high sensitivity (>97 %) and specificity (>98 %) across all sample categories, with a notably low limit of detection for adulteration at 2.8 %. Characteristic peaks, such as the pantocrin peak at 0.25 V for authentic samples, provided a robust basis for differentiation. The method's effectiveness in detecting subtle adulterations was evidenced by its ability to identify samples with as low as 5 % adulteration. Furthermore, the approach showed excellent generalization, maintaining 97.0 % accuracy on an independent validation set. These findings highlight the potential of this technique for rapid, accurate, and cost-effective authentication of medicinal materials, addressing the growing challenge of counterfeit products in the pharmaceutical industry.

Unsupervised Learning-Assisted Acoustic-Driven Nano-Lens Holography for the Ultrasensitive and Amplification-Free Detection of Viable Bacteria.

Bacterial infection is a crucial factor resulting in public health issues worldwide, often triggering epidemics and even fatalities. The accurate, rapid, and convenient detection of viable bacteria is an effective method for reducing infections and illness outbreaks. Here, an unsupervised learning-assisted and surface acoustic wave-interdigital transducer-driven nano-lens holography biosensing platform is developed for the ultrasensitive and amplification-free detection of viable bacteria. The monitoring device integrated with the nano-lens effect can achieve the holographic imaging of polystyrene microsphere probes in an ultra-wide field of view (∽28.28mm2), with a sensitivity limit of as low as 99nm. A lightweight unsupervised learning hologram processing algorithm considerably reduces training time and computing hardware requirements, without requiring datasets with manual labels. By combining phage-mediated viable bacterial DNA extraction and enhanced CRISPR-Cas12a systems, this strategy successfully achieves the ultrasensitive detection of viable Salmonella in various real samples, demonstrating enhanced accuracy validated with the qPCR benchmark method. This approach has a low cost (∽$500) and is rapid (∽1h) and highly sensitive (∽38CFUmL-1), allowing for the amplification-free detection of viable bacteria and emerging as a powerful tool for food safety inspection and clinical diagnosis.

Decoding Continuous Tracking Eye Movements from Cortical Spiking Activity.

Eye movements are the primary way primates interact with the world. Understanding how the brain controls the eyes is therefore crucial for improving human health and designing visual rehabilitation devices. However, brain activity is challenging to decipher. Here, we leveraged machine learning algorithms to reconstruct tracking eye movements from high-resolution neuronal recordings. We found that continuous eye position could be decoded with high accuracy using spiking data from only a few dozen cortical neurons. We tested eight decoders and found that neural network models yielded the highest decoding accuracy. Simpler models performed well above chance with a substantial reduction in training time. We measured the impact of data quantity (e.g. number of neurons) and data format (e.g. bin width) on training time, inference time, and generalizability. Training models with more input data improved performance, as expected, but the format of the behavioral output was critical for emphasizing or omitting specific oculomotor events. Our results provide the first demonstration, to our knowledge, of continuously decoded eye movements across a large field of view. Our comprehensive investigation of predictive power and computational efficiency for common decoder architectures provides a much-needed foundation for future work on real-time gaze-tracking devices.

Resource Allocation in UAV-D2D Networks: A Scalable Heterogeneous Multi-Agent Deep Reinforcement Learning Approach

In unmanned aerial vehicle (UAV)-assisted device-to-device (D2D) caching networks, the uncertainty from unpredictable content demands and variable user positions poses a significant challenge for traditional optimization methods, often making them impractical. Multi-agent deep reinforcement learning (MADRL) offers significant advantages in optimizing multi-agent system decisions and serves as an effective and practical alternative. However, its application in large-scale dynamic environments is severely limited by the curse of dimensionality and communication overhead. To resolve this problem, we develop a scalable heterogeneous multi-agent mean-field actor-critic (SH-MAMFAC) framework. The framework treats ground users (GUs) and UAVs as distinct agents and designs cooperative rewards to convert the resource allocation problem into a fully cooperative game, enhancing global network performance. We also implement a mixed-action mapping strategy to handle discrete and continuous action spaces. A mean-field MADRL framework is introduced to minimize individual agent training loads while enhancing total cache hit probability (CHP). The simulation results show that our algorithm improves CHP and reduces transmission delay. A comparative analysis with existing mainstream deep reinforcement learning (DRL) algorithms shows that SH-MAMFAC significantly reduces training time and maintains high CHP as GU count grows. Additionally, by comparing with SH-MAMFAC variants that do not include trajectory optimization or power control, the proposed joint design scheme significantly reduces transmission delay.

LLM GPT-3.5 study for sentiment analysis across Utkarsh server, Ohio supercomputer, Google Colab and PC

The major objective of the present research is to inspect sentiment analysis models that have been trained on Twitter corpus by utilising the Large Language Model (LLM) gpt-3–5-turbo-16k version of the Generative Pretrained Transformer (GPT 3.5) model. Such trained models include the Bidirectional long short-term memory neural network (BiLSTM), Convolutional Neural Networks, Gated Recurrent Unit and Recurrent Neural Network which were used to perform computational tasks on the Ohio Supercomputer and Utkarsh Server, in comparison to work conducted on Google Colab and a Personal Computer (PC).This research also looks at the performance as well as the computational aspects of these models in terms of accuracy, recall, F1-score, time/memory complexity and resource requirements (CPU/GPU) throughout the training and testing phases of each model. The training accuracies in this case were concentrated between 49.91 % and 99.98 % while those of the testing accuracies of the models accounts for about 50.00–75.00 %. For instance, models including BiLSTM and RNN usually exhibit more time complexity because of the nature of the models (sequential computation), on the contrary, CNNs are less time-consuming and are more effective in terms of storage modifying layered architecture. The use of supercomputers and specialized servers reduces training time, but resource constraints on platforms such as personal computers or Colab cause considerable divergence.

An attention-based CNN model integrating observational and simulation data for high-resolution spatial estimation of urban air quality

Machine learning, especially deep learning, can outperform traditional atmospheric models in air quality assessment, offering enhanced efficiency and accuracy without relying on detailed emission inventories and atmospheric chemical mechanisms. Despite their predictive power, deep learning models often grapple with the perception of being “black boxes” due to their intricate architectures. Here, we develop an attention-based convolutional neural network (CNN-attention) model that incorporates observational data, the parallelized large-eddy-simulation model (PALM), and urban morphology data for high-resolution spatial estimation of urban air quality. Our findings indicate that the CNN-attention model outperforms traditional CNN with higher accuracy and efficiency, achieving R2 = 0.987 and root mean square error (RMSE) = 0.15 mg/m3, while significantly reducing training time and memory usage. Compared to traditional machine learning models, the CNN exhibits higher R2 values and lower RMSE, showcasing its adeptness at capturing complex nonlinear patterns. The inclusion of attention layer further improves the model's performance by dynamically assigning attention scores to key features, enabling the model to focus on areas of critical emissions and distinctive urban features such as highways, arterial roads, intersections, and dense building clusters. This approach also reveals fluid dynamical principles, highlighting the significant disparities in pollutant concentration across roadways caused by atmospheric turbulence, and the distinct plume formations influenced by land use and topography. When applied to various urban settings, the CNN-attention model exhibits superior generalizability and transferability. This study provides valuable scientific insights and technical support for urban planning, air quality management, and exposure risk evaluation.

CGLight: An effective indoor illumination estimation method based on improved convmixer and GauGAN

Illumination consistency is a key factor for seamlessly integrating virtual objects with real scenes in augmented reality (AR) systems. High dynamic range (HDR) panoramic images are widely used to estimate scene lighting accurately. However, generating environment maps requires complex deep network architectures, which cannot operate on devices with limited memory space. To address this issue, this paper proposes CGLight, an effective illumination estimation method that predicts HDR panoramic environment maps from a single limited field-of-view (LFOV) image. We first design a CMAtten encoder to extract features from input images, which learns the spherical harmonic (SH) lighting representation with fewer model parameters. Guided by the lighting parameters, we train a generative adversarial network (GAN) to generate HDR environment maps. In addition, to enrich lighting details and reduce training time, we specifically introduce the color consistency loss and independent discriminator, considering the impact of color properties on the lighting estimation task while improving computational efficiency. Furthermore, the effectiveness of CGLight is verified by relighting virtual objects using the predicted environment maps, and the root mean square error and angular error are 0.0494 and 4.0607 in the gray diffuse sphere, respectively. Extensive experiments and analyses demonstrate that CGLight achieves a balance between indoor illumination estimation accuracy and resource efficiency, attaining higher accuracy with nearly 4 times fewer model parameters than the ViT-B16 model.

Preventing Allogeneic Stem Cell Transplant-Related Cardiovascular Dysfunction: ALLO-Active Trial.

Allogeneic stem cell transplantation (allo-SCT) is an efficacious treatment for hematologic malignancies but can be complicated by cardiac dysfunction and exercise intolerance impacting quality of life and longevity. We conducted a randomized controlled trial testing whether a multicomponent activity intervention could attenuate reductions in cardiorespiratory fitness and exercise cardiac function (co-primary end points) in adults undergoing allo-SCT. Sixty-two adults scheduled for allo-SCT were randomized to a 4-month activity program (n=30) or usual care (UC; n=32). Activity comprised multicomponent exercise training (3 days/week) and sedentary time reduction (≥30 min/day) program and was delivered throughout hospitalization (≈4 weeks) and for 12 weeks after discharge. Physiological assessments conducted before admission and at 12 weeks after discharge included cardiopulmonary exercise testing to quantify peak oxygen uptake ([Formula: see text]), exercise cardiac magnetic resonance imaging for peak cardiac volume (CIpeak) and stroke volume (SVIpeak) index, echocardiography-derived left ventricular ejection fraction and global longitudinal strain, and cardiac biomarkers (cTn-I [troponin-I] and BNP [B-type natriuretic peptide]). Fifty-two participants (84%) completed follow-up (25 activity and 27 UC); median (interquartile range [IQR]) adherence to the activity program was 74% (41-96%). There was a marked decline in [Formula: see text] in the UC program (-3.4 mL‧kg-1‧min-1 [95% CI, -4.9 to -1.8]) that was attenuated with activity (-0.9 mL‧kg-1‧min-1 [95% CI, -2.5 to 0.8]; interaction P=0.029). Activity preserved exercise cardiac function, with preservation of CIpeak (0.30 L‧min-1‧m-2 [95% CI, -0.34 to 0.41]) and SVIpeak (0.6 mL/m2 [95% CI, -1.3 to 2.5]), both of which declined with UC (CIpeak, -0.68 L‧min-1‧m-2 [95% CI, -1.3 to -0.32]; interaction P=0.008; SVIpeak, -2.7 mL/m2 [95% CI, -4.6 to -0.9]; interaction P=0.014). There were no treatment effects of activity on cardiac biomarkers or echocardiographic indices. Multicomponent activity intervention during and after allo-SCT is beneficial for preserving patient cardiorespiratory fitness and exercise cardiac function. These results may have important implications for cardiovascular morbidity and mortality after allo-SCT. URL: https://anzctr.org.au/; Unique identifier: ACTRN12619000741189.

Optimal feature selection for heart disease prediction using modified Artificial Bee colony (M-ABC) and K-nearest neighbors (KNN)

Heart disease is a complex and widespread illness that affects a significant number of people worldwide. Machine learning provides a way forward for early heart disease diagnosis. A classification model has been developed for the present study to predict heart disease. The attribute selection was done using a modified bee algorithm. Using the proposed model, practitioners can accurately predict heart disease and make informed decisions about patient health. In our study, we have proposed a framework based on Modified Artificial Bee Colony (M-ABC) and k-Nearest Neighbors (KNN) for predicting the optimal feature selection to obtain better accuracy. Using a modified bee algorithm, this paper focuses on identifying the optimal subset of attributes from the dataset. Specifically, during the classification-training phase, only the features that provide significant information are retained. The proposed study not only improves classification accuracy but also reduces training time for classifiers.

Cross-coupled prompt learning for few-shot image recognition

Prompt learning based on large models shows great potential to reduce training time and resource costs, which has been progressively applied to visual tasks such as image recognition. Nevertheless, the existing prompt learning schemes suffer from either inadequate prompt information from a single modality or insufficient prompt interaction between multiple modalities, resulting in low efficiency and performance. To address these limitations, we propose a Cross-Coupled Prompt Learning (CCPL) architecture, which is designed with two novel components (i.e., Cross-Coupled Prompt Generator (CCPG) module and Cross-Modal Fusion (CMF) module) to achieve efficient interaction between visual and textual prompts. Specifically, the CCPG module incorporates a cross-attention mechanism to automatically generate visual and textual prompts, each of which will be adaptively updated using the self-attention mechanism in their respective image and text encoders. Furthermore, the CMF module implements a deep fusion to reinforce the cross-modal feature interaction from the output layer with the Image–Text Matching (ITM) loss function. We conduct extensive experiments on 8 image datasets. The experimental results verify that our proposed CCPL outperforms the SOTA methods on few-shot image recognition tasks. The source code of this project is released at: https://github.com/elegantTechie/CCPL.

Rapid and accurate bacteria identification through deep-learning-based two-dimensional Raman spectroscopy

Surface-enhanced Raman spectroscopy (SERS) offers a distinctive vibrational fingerprint of the molecules and has led to widespread applications in medical diagnosis, biochemistry, and virology. With the rapid development of artificial intelligence (AI) technology, AI-enabled Raman spectroscopic techniques, as a promising avenue for biosensing applications, have significantly boosted bacteria identification. By converting spectra into images, the dataset is enriched with more detailed information, allowing AI to identify bacterial isolates with enhanced precision. However, previous studies usually suffer from a trade-off between high-resolution spectrograms for high-accuracy identification and short training time for data processing. Here, we present an efficient bacteria identification strategy that combines deep learning models with a spectrogram encoding algorithm based on wavelet packet transform and Gramian angular field techniques. In contrast to the direct analysis of raw Raman spectra, our approach utilizes wavelet packet transform techniques to compress the spectra by a factor of 1/15, while concurrently maintaining state-of-the-art accuracy by amplifying the subtle differences via Gramian angular field techniques. The results demonstrate that our approach can achieve a 99.64 % and a 90.55 % identification accuracy for two types of bacterial isolates and thirty types of bacterial isolates, respectively, while a 90 % reduction in training time compared to the conventional methods. To verify the model's stability, Gaussian noises were superimposed on the testing dataset, showing a specific generalization ability and superior performance. This algorithm has the potential for integration into on-site testing protocols and is readily updatable with new bacterial isolates. This study provides profound insights and contributes to the current understanding of spectroscopy, paving the way for accurate and rapid bacteria identification in diverse applications of environment monitoring, food safety, microbiology, and public health.

Malware Detection and Classification using Generative Adversarial Network

The Generative Adversarial Networks (GANs) are playing a crucial role in deep-learning-based malware classification to overcome the dataset imbalance and unseen malware. The Generative AI is preferably used in many applications, such as improving image resolution and generating audio, video, and text. The cybercriminals are also using the Generative AI for generating the malware and deepfake videos to harm the targeted person or device. By generating the synthetic data, it makes the deep learning model more robust to detect such types of unseen and adversarial attacks. This work utilizes GANs for generating adversarial malware samples to train a classification and detection model, improving the model’s ability to identify sophisticated malware variants. The performance of the proposed Conditional Generative Adversarial Network (CGAN) model is evaluated on a multiclass malware grayscale image dataset and a binary class malware RGB image dataset. The performance of proposed model is compared with current state-of-the-art. Results indicate a significant improvement in classification accuracy and a reduction in training time and false positives, showcasing GAN’s potential in the dynamic cybersecurity landscape.

Continual learning with high-order experience replay for dynamic network embedding

Dynamic network embedding (DNE) poses a tough challenge in graph representation learning, especially when confronted with the frequent updates of streaming data. Conventional DNEs primarily resort to parameter updating but perform inadequately on historical networks, resulting in the problem of catastrophic forgetting. To tackle such issues, recent advancements in graph neural networks (GNNs) have explored matrix factorization techniques. However, these approaches encounter difficulties in preserving the global patterns of incremental data. In this paper, we propose CLDNE, a Continual Learning framework specifically designed for Dynamic Network Embedding. At the core of CLDNE lies a streaming graph auto-encoder that effectively captures both global and local patterns of the input graph. To further overcome catastrophic forgetting, CLDNE is equipped with an experience replay buffer and a knowledge distillation module, which preserve high-order historical topology and static historical patterns. We conduct experiments on four dynamic networks using link prediction and node classification tasks to evaluate the effectiveness of CLDNE. The outcomes demonstrate that CLDNE successfully mitigates the catastrophic forgetting problem and reduces training time by 80% without a significant loss in learning new patterns.