Computational Resource Requirements Research Articles

A study of human pose estimation in low-light environments using YOLOv8 model

Human pose estimation is a formidable task in the field of computer vision., often constrained by limited training samples and various complexities encountered during target detection, including complex backgrounds, object occlusion, crowded scenes, and varying perspectives. The primary objective of this research paper is to explore the performance disparities of the recently introduced YOLOv8 model in the context of human pose estimation. We conduct a comprehensive evaluation of six different models with varying complexities on the same low-light photograph to assess their precision and speed. The objective is to determine the suitability of each model for specific environmental contexts. The experimental results reveal that our findings demonstrate a partial regression in accuracy for the yolov8s-pose and yolov8m-pose models when tested on our sampled images. The increase in model layers indicates enhanced complexity and expressive power, while additional parameters signify improved learning capabilities at the expense of increased computational resource requirements.

Application And Challenges of Alphazero Algorithm in Chinese Chess

In the field of artificial intelligence, AlphaZero has achieved significant success in games like Go and Shogi. However, its application in Chinese Chess remains relatively unexplored. This study aims to unveil the potential and limitations of AlphaZero in Chinese Chess and explore strategies for optimizing its performance. In this paper, we begin by reviewing the fundamental principles and algorithmic structure of AlphaZero, emphasizing its deep reinforcement learning and self-play characteristics. Subsequently, we delve into the specific application of AlphaZero in Chinese Chess, encompassing aspects such as chess rule representation, training procedures, position evaluation, and strategy prediction. By analyzing successful instances of AlphaZero in Chinese Chess, we showcase its potential in the domain of board games. Furthermore, this paper elucidates the challenges that AlphaZero encounters in the context of Chinese Chess, including issues related to computational complexity and position representation. We particularly underscore the potential cost escalation due to high computational resource requirements and the risk of data loss or excessive time consumption associated with extended computation periods. IN conclusion, this research provides valuable insights into the application of AlphaZero in the realm of Chinese Chess and offers recommendations for future research directions to further expand its utility in this domain.

Enhancing Radar Echo Extrapolation by ConvLSTM2D for Precipitation Nowcasting.

Precipitation nowcasting in real-time is a challenging task that demands accurate and current data from multiple sources. Despite various approaches proposed by researchers to address this challenge, models such as the interaction-based dual attention LSTM (IDA-LSTM) face limitations, particularly in radar echo extrapolation. These limitations include higher computational costs and resource requirements. Moreover, the fixed kernel size across layers in these models restricts their ability to extract global features, focusing more on local representations. To address these issues, this study introduces an enhanced convolutional long short-term 2D (ConvLSTM2D) based architecture for precipitation nowcasting. The proposed approach includes time-distributed layers that enable parallel Conv2D operations on each image input, enabling effective analysis of spatial patterns. Following this, ConvLSTM2D is applied to capture spatiotemporal features, which improves the model's forecasting skills and computational efficacy. The performance evaluation employs a real-world weather dataset benchmarked against established techniques, with metrics including the Heidke skill score (HSS), critical success index (CSI), mean absolute error (MAE), and structural similarity index (SSIM). ConvLSTM2D demonstrates superior performance, achieving an HSS of 0.5493, a CSI of 0.5035, and an SSIM of 0.3847. Notably, a lower MAE of 11.16 further indicates the model's precision in predicting precipitation.

DiffMat: Latent diffusion models for image-guided material generation

Creating realistic materials is essential in the construction of immersive virtual environments. While existing techniques for material capture and conditional generation rely on flash-lit photos, they often produce artifacts when the illumination mismatches the training data. In this study, we introduce DiffMat, a novel diffusion model that integrates the CLIP image encoder and a multi-layer, cross-attention denoising backbone to generate latent materials from images under various illuminations. Using a pre-trained StyleGAN-based material generator, our method converts these latent materials into high-resolution SVBRDF textures, a process that enables a seamless fit into the standard physically based rendering pipeline, reducing the requirements for vast computational resources and expansive datasets. DiffMat surpasses existing generative methods in terms of material quality and variety, and shows adaptability to a broader spectrum of lighting conditions in reference images.

Federated and Online Dynamic Spectrum Access for Mobile Secondary Users

Users in dynamic spectrum access (DSA) with federated reinforcement learning (FRL) autonomously access channels, avoiding centralized coordination and protecting users’ privacy. However, existing FRL-based DSA mechanisms are limited to ideal network states, i.e., assuming that channel states and users’ interference relationships are unchanged. Besides, users should upload intermediate results simultaneously for federated aggregation. The above conditions are impractical for mobile users since their network states and locations are unstable. Meanwhile, newly connected users have to train their models through local data with numerous computing resources since global models are unsuitable for them. We propose FRDSA, an FRL-based secure and lightweight channel selection mechanism in DSA for mobile users under dynamic network states. An independent channel selection environment with a virtual group strategy is presented to avoid interference between users under unstable channel states. Furthermore, an asynchronous parameter aggregation method in FRDSA dynamically adjusts the aggregation factors without users simultaneously uploading intermediate results. Simulations based on real trajectory data show that FRDSA significantly reduces approximately 60% interference between mobile users under unstable network states. Newly connected users can directly apply the well-trained global model to access channels autonomously instead of retraining a model, effectively reducing mobile users’ computing resource requirements.

Skin cancer detection using lightweight model souping and ensembling knowledge distillation for memory-constrained devices

In contemporary times, the escalating prevalence of skin cancer is a significant concern, impacting numerous individuals. This work comprehensively explores advanced artificial intelligence-based deep learning techniques for skin cancer detection, utilizing the HAM10000 dataset. The experimental study fine-tunes two knowledge distillation teacher models, ResNet50 (25.6M) and DenseNet161 (28.7M), achieving remarkable accuracies of 98.32% and 98.80%, respectively. Despite their notable accuracy, the training and deployment of these large models pose significant challenges for implementation on memory-constrained medical devices. To address this issue, we introduce TinyStudent (0.35M), employing knowledge distillation from ResNet50 and DenseNet161, yielding accuracies of 85.45% and 85.00%, respectively. While TinyStudent may not achieve accuracies comparable to the teacher models, it is 82 and 73 times smaller than DenseNet161 and ResNet50, respectively, implying reduced training time and computational resource requirements. This significant reduction in the number of parameters makes it feasible to deploy the model on memory-constrained edge devices. Multi-teacher distillation, incorporating knowledge from both models, results in a competitive student accuracy of 84.10%. Ensembling methods, such as average ensembling and concatenation, further enhance predictive performances, achieving accuracies of 87.74% and 88.00%, respectively, each with approximately 1.05M parameters. Compared to DenseNet161 and ResNet50, these lightweight ensemble models offer shorter inference times, suitable for medical devices. Additionally, our implementation of the Greedy method in Model Soup establishes an accuracy of 85.70%.

An Epileptic EEG Detection Method Based on Data Augmentation and Lightweight Neural Network.

Epilepsy, an enduring neurological disorder, afflicts approximately 65 million individuals globally, significantly impacting their physical and mental wellbeing. Traditional epilepsy detection methods are labor-intensive, leading to inefficiencies. Although deep learning techniques for brain signal detection have gained traction in recent years, their clinical application advancement is hindered by the significant requirement for high-quality data and computational resources during training. The neural network training initially involved merging two datasets of different data quality, namely Bonn University datasets and CHB-MIT datasets, to bolster its generalization capabilities. To tackle the issues of dataset size and class imbalance, we employed small window segmentation and Synthetic Minority Over-sampling Technique (SMOTE). algorithms to augment and equalize the data. A streamlined neural network architecture was then proposed, drastically reducing the model's training parameters. Notably, a model trained with a mere 9,371 parameters yielded impressive results. The three-classification task on the combined dataset delivered an accuracy of 98.52%, sensitivity of 97.99%, specificity of 99.35%, and precision of 98.44%. The experimental findings of this study underscore the superiority of the proposed method over existing approaches in both model size reduction and accuracy enhancement. As a result, it is more apt for deployment in low-cost, low computational hardware devices, including wearable technology, and various clinical applications. Clinical and Translational Impact Statement- This study is a Pre-Clinical Research. The lightweight neural network is easily deployed on hardware device for real-time epileptic EEG detection.

МЕТАОБУЧЕНИЕ ДЛЯ АВТОМАТИЧЕСКОГО ПРОЕКТИРОВАНИЯ ОПТИМАЛЬНЫХ АРХИТЕКТУР НЕЙРОСЕТЕЙ

The article discusses the use of meta-learning for automatic design of optimal neural network architectures. An approach based on two-level optimization is proposed, where meta-learning is used to search for universal parameters and strategies that accelerate the process of generating effective architectures. It is shown that meta-learning reduces computational costs and provides high adaptability to new problems, which makes it preferable to traditional methods such as manual design and network architecture search (NAS). Experimental results confirming the advantage of the approach are presented, and the main limitations are identified, including the requirements for computational resources and challenges in generalizing to new problems. Promising directions of development are considered, including integration with transformers, hybrid methods, and adaptation for specialized areas such as medicine and finance. The proposed approach contributes to the acceleration of the development of neural networks, improving their performance and availability for application in a wide range of problems.

Interpretable dimensionality reduction and classification of mass spectrometry imaging data in a visceral pain model via non-negative matrix factorization.

Mass spectrometry imaging (MSI) is a powerful scientific tool for understanding the spatial distribution of biochemical compounds in tissue structures. In this paper, we introduce three novel approaches in MSI data processing to perform the tasks of data augmentation, feature ranking, and image registration. We use these approaches in conjunction with non-negative matrix factorization (NMF) to resolve two of the biggest challenges in MSI data analysis, namely: 1) the large file sizes and associated computational resource requirements and 2) the complexity of interpreting the very high dimensional raw spectral data. There are many dimensionality reduction techniques that address the first challenge but do not necessarily result in readily interpretable features, leaving the second challenge unaddressed. We demonstrate that NMF is an effective dimensionality reduction algorithm that reduces the size of MSI datasets by three orders of magnitude with limited loss of information, yielding spatial and spectral components with meaningful correlation to tissue structure that may be used directly for subsequent data analysis without the need for additional clustering steps. This analysis is demonstrated on an MSI dataset from female Sprague-Dawley rats for an animal model of comorbid visceral pain hypersensitivity (CPH). We find that high-dimensional MSI data (∼ 100,000 ions per pixel) can be reduced to 20 spectral NMF components with < 20% loss in reconstruction accuracy. The resulting spatial NMF components are reproducible and correlate well with H&E-stained tissue images. These components may also be used to generate images with enhanced specificity for different tissue types. Small patches of NMF data (i.e., 20 spatial NMF components over 20 × 20 pixels) provide an accuracy of ∼ 87% in classifying CPH vs naïve control subjects. This paper presents the novel data processing methodologies that were used to produce these results, encompassing novel data processing pipelines for data augmentation to support training for classification, ranking of features according to their contribution to classification, and image registration to enhance tissue-specific imaging.

Sports-ACtrans Net: research on multimodal robotic sports action recognition driven via ST-GCN.

Accurately recognizing and understanding human motion actions presents a key challenge in the development of intelligent sports robots. Traditional methods often encounter significant drawbacks, such as high computational resource requirements and suboptimal real-time performance. To address these limitations, this study proposes a novel approach called Sports-ACtrans Net. In this approach, the Swin Transformer processes visual data to extract spatial features, while the Spatio-Temporal Graph Convolutional Network (ST-GCN) models human motion as graphs to handle skeleton data. By combining these outputs, a comprehensive representation of motion actions is created. Reinforcement learning is employed to optimize the action recognition process, framing it as a sequential decision-making problem. Deep Q-learning is utilized to learn the optimal policy, thereby enhancing the robot's ability to accurately recognize and engage in motion. Experiments demonstrate significant improvements over state-of-the-art methods. This research advances the fields of neural computation, computer vision, and neuroscience, aiding in the development of intelligent robotic systems capable of understanding and participating in sports activities.

AAUConvNeXt: Enhancing Crop Lodging Segmentation with Optimized Deep Learning Architectures.

Rice lodging, a phenomenon precipitated by environmental factors or crop characteristics, presents a substantial challenge in agricultural production, notably impacting yield prediction and disaster assessment. Despite that the application of conventional methodologies like visual assessment, mathematical models, and satellite remote sensing technologies has been employed in the segmentation of crop lodging, these approaches are still constrained in precision, immediacy, and capacity for large-scale evaluation. This study introduces an innovative convolutional neural network architecture, AFOA + APOM + UConvNeXt, that integrates intelligent optimization algorithms for automatic selection of optimal network parameters, thereby enhancing the accuracy and efficiency of crop lodging segmentation. The proposed model, empirically validated, outperforms recent state-of-the-art models in crop lodging segmentation, demonstrating higher accuracy, lower computational resource requirements, and greater efficiency, thereby markedly reducing the cost of segmentation. In addition, we investigated the segmentation on half lodging rice, and the results indicate that the model exhibits commendable performance on the half lodging dataset. This outcome holds significant implications for the prediction of rice lodging trends. The fusion of deep learning with intelligent optimization algorithms in this study offers a new effective tool for crop lodging monitoring in agricultural production, providing strong technical support for accurate crop phenotypic information extraction, and is expected to play a significant role in agricultural production practices.

Needle tracking and segmentation in breast ultrasound imaging based on spatio-temporal memory network.

Ultrasound-guided needle biopsy is a commonly employed technique in modern medicine for obtaining tissue samples, such as those from breast tumors, for pathological analysis. However, it is limited by the low signal-to-noise ratio and the complex background of breast ultrasound imaging. In order to assist physicians in accurately performing needle biopsies on pathological tissues, minimize complications, and avoid damage to surrounding tissues, computer-aided needle segmentation and tracking has garnered increasing attention, with notable progress made in recent years. Nevertheless, challenges remain, including poor ultrasound image quality, high computational resource requirements, and various needle shape. This study introduces a novel Spatio-Temporal Memory Network designed for ultrasound-guided breast tumor biopsy. The proposed network integrates a hybrid encoder that employs CNN-Transformer architectures, along with an optical flow estimation method. From the Ultrasound Imaging Department at the First Affiliated Hospital of Shantou University, we developed a real-time segmentation dataset specifically designed for ultrasound-guided needle puncture procedures in breast tumors, which includes ultrasound biopsy video data collected from 11 patients. Experimental results demonstrate that this model significantly outperforms existing methods in improving the positioning accuracy of needle and enhancing the tracking stability. Specifically, the performance metrics of the proposed model is as follows: IoU is 0.731, Dice is 0.817, Precision is 0.863, Recall is 0.803, and F1 score is 0.832. By advancing the precision of needle localization, this model contributes to enhanced reliability in ultrasound-guided breast tumor biopsy, ultimately supporting safer and more effective clinical outcomes. The model proposed in this paper demonstrates robust performance in the computer-aided tracking and segmentation of biopsy needles in ultrasound imaging, specifically for ultrasound-guided breast tumor biopsy, offering dependable technical support for clinical procedures.

A Comparative Analysis of Modern Object Detection Algorithms: YOLO vs. SSD vs. Faster R-CNN

In recent years, object detection has become a crucial component in various computer vision applications, including autonomous driving, surveillance, and image recognition. This study provides a comprehensive comparative analysis of three prominent object detection algorithms: You Only Look Once (YOLO), Single Shot MultiBox Detector (SSD), and Faster Region-Based Convolutional Neural Networks (Faster R-CNN). The background of this research lies in the growing need for efficient and accurate object detection methods that can operate in real-time. YOLO is known for its speed, SSD for its balance between speed and accuracy, and Faster R-CNN for its high detection accuracy, albeit at a slower pace. The methodology involves implementing these algorithms on a standardized dataset and evaluating their performance based on various metrics, including detection accuracy, processing speed, and computational resource requirements. Each algorithm is tested under similar conditions to ensure a fair comparison. The results indicate that while YOLO excels in real-time applications due to its high speed, SSD offers a middle ground with respectable accuracy and speed, making it suitable for applications requiring a balance of both. Faster R-CNN demonstrates superior accuracy, making it ideal for scenarios where detection precision is paramount, despite its slower performance. This comparative analysis highlights the strengths and weaknesses of each algorithm, providing valuable insights for researchers and practitioners in selecting the appropriate object detection method for their specific needs.

Feature enhancement network for cloud removal in optical images by fusing with SAR images

ABSTRACT Presence of cloud-covered pixels is inevitable in optical remote-sensing images. Therefore, the reconstruction of the cloud-covered details is important to improve the usage of these images for subsequent image analysis tasks. Aiming to tackle the issue of high computational resource requirements that hinder the application at scale, this paper proposes a Feature Enhancement Network(FENet) for removing clouds in satellite images by fusing Synthetic Aperture Radar (SAR) and optical images. The proposed network consists of designed Feature Aggregation Residual Block (FAResblock) and Feature Enhancement Block (FEBlock). FENet is evaluated on the publicly available SEN12MS-CR dataset and it achieves promising results compared to the benchmark and the state-of-the-art methods in terms of both visual quality and quantitative evaluation metrics. It proved that the proposed feature enhancement network is an effective solution for satellite image cloud removal using less computational and time consumption. The proposed network has the potential for practical applications in the field of remote sensing due to its effectiveness and efficiency. The developed code and trained model will be available at https://github.com/chenxiduan/FENet.

New approaches to simplification of integrated asset models

The current trend in modeling of the development of oil and gas fields is the transition from models of individual elements of the production system to complex integrated asset models (IAM) of hydrocarbon production fields. The use of such models is especially relevant for the correct forecasting and management of hydrocarbon production in gas, gas condensate and oil and gas condensate fields, where the parameters of facility infrastructure determine the dynamics of production no less than wells and productive reservoirs. The complexity of integrated asset models is associated with the labor-intensive of its creation and the high requirements for computational and time resources required to create and maintain models. This article proposes approaches to increase the efficiency of calculations of integrated asset models while maintaining the quality of forecasting, which helps to increase the value of modeling and the degree of details of development of project solutions. A study of four integrated asset models configurations was carried out. Firstly, the operating features of a detailed integrated asset model are presented, and then methods for simplifying both the reservoir model and the gathering system model are described. For each model, key characteristics are given, as well as calculation algorithms. Through the example of a gas field, a numerical experiment was performed using all the considered configurations; a comparison of the main technological parameters of development was carried out, which showed similar results for all configurations. Based on the study, a conclusion was made about the possibility of using such simplified integrated asset models to perform operational, including multivariate calculations in addition to detailed integrated asset models.

Walking on two legs: Joint service placement and computation configuration for provisioning containerized services at edges

With the development of edge computing and container virtualization, provisioning containerized services at network edges is proposed for high responsiveness and low wide area network (WAN) traffic. However, realizing its full potential faces multiple challenges. First, due to containers’ fine-grained computation resource isolation, finely configuring services’ computation resource requirements is needed, especially for resource-constrained edge nodes. Second, due to edges’ heterogeneity and services’ diversity, system performance highly depends on which edge node to place each service. Third, as the main metric of quality of service, service response time involves the computing-network delay tradeoff and urges to optimize the decisions jointly. Prior works on edge-enabled service placement either ignore computation resource isolation and configuration, or assume computation resource configuration is given manually. To fill this gap, this paper investigates the joint service placement and computation configuration problem for provisioning containerized services at edges. Then based on the convex and submodular optimization techniques, we propose a two-stage greedy and local-search combined algorithm, TeLa for short. Rigorous theoretical analyses demonstrate that TeLa is a polynomial-time algorithm with performance guarantees. Finally, we implement twelve containerized services and an edge computing prototype to realistically evaluate TeLa. The results confirm TeLa’s empirical superiority over state-of-the-art algorithms, in terms of 39% on average reduction on the weighted sum of service response time and WAN traffic.

Improving Scientific Literature Classification: A Parameter-Efficient Transformer-Based Approach

Transformer-based models have been utilized in natural language processing (NLP) for a wide variety of tasks like summarization, translation, and conversational agents. These models can capture long-term dependencies within the input, so they have significantly more representational capabilities than Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN). Nevertheless, these models require significant computational resources in terms of high memory usage, and extensive training time. In this paper, we propose a novel document categorization model, with improved parameter efficiency that encodes text using a single, lightweight, multiheaded attention encoder block. The model also uses a hybrid word and position embedding to represent input tokens. The proposed model is evaluated for the Scientific Literature Classification task (SLC) and is compared with state-of-the-art models that have previously been applied to the task. Ten datasets of varying sizes and class distributions have been employed in the experiments. The proposed model shows significant performance improvements, with a high level of efficiency in terms of parameter and computation resource requirements as compared to other transformer-based models, and outperforms previously used methods.

Seamless Wireless Communication Platform for Internet of Things Applications

The rapid growth of the Internet of Things (IoT) devices resulted in the proliferation of wireless technologies to cater to their increasing data rate requirements and support multiple applications. However, such ever-increasing wireless technologies present numerous challenges such as incompatible wireless standards, increased energy consumption, and insecure communication. The traditional gateways proposed in the literature suffers from limitations such as computational complexity, resource requirements, increased cost, and device size. We vision an era of seamless wireless communication to alleviate the aforementioned challenges in IoT applications. through three inter-dependent functionalities namely detection and identification of wireless technologies, energy-efficient transmit power control, and secure end-to-end communication. To prove the concept, a new gateway is proposed to achieve these three functionalities with only physical layer measurements so that the different communication protocols in the higher layers can be avoided. Novel schemes are conceptualized for resource-limited seamless IoT applications. Moreover, the conceptual seamless IoT platform is validated through software-based computer simulation and software-defined radio-based testbed implementation. The preliminary analysis demonstrates that the proposed platform has great potential in advancing seamless IoT applications.

A High‐Resolution Prediction Network for Predicting Intratumoral Distribution of Nanoprobes by Tumor Vascular and Nuclear Feature

In this study, the critical need for precise and accurate prediction of intra‐tumor heterogeneity related to the enhanced permeability and retention effect and spatial distribution of nanoprobes is addressed for the development of effective nanodrug delivery strategies. Current predictive models are limited in terms of resolution and accuracy, prompting the construction of a high‐resolution prediction network (HRPN) that estimates the microdistribution of quantum dots, factoring in tumor vascular and nuclear features. The HRPN algorithm is trained using 27 780 patches and validated on 4920 patches derived from 4T1 breast cancer whole‐slide images, demonstrating its reliability. The HRPN model exhibits minimal error (mean square error = 1.434, root mean square error = 1.198), satisfactory goodness of fit (R2 = 0.891), and superior image quality (peak signal‐to‐noise ratio = 44.548) when compared to a generative‐adversarial‐network‐structured model. Furthermore, the HRPN model offers improved prediction accuracy, broader prediction intervals, and reduced computational resource requirements. Consequently, the proposed model yields high‐resolution predictions that more closely resemble actual tumor microdistributions, potentially serving as a powerful analytical tool for investigating the spatial relationship between the tumor microenvironment and nanoprobes.

Study of the ensemble of machine learning models in medical diagnostics

The research discusses the use of machine learning models for medical diagnosis, which is of great significance in modern medicine. The focus of the article is on the use of ensemble models in which convolutional neural networks (CNN), models for classification and regression tasks (SVM), and long short-term memory recurrent networks (LSTM) are involved. The effectiveness of these methods is compared both individually and in an ensemble approach in the context of improving the accuracy of medical diagnosis. Despite numerous studies discussed in the article combining multiple models simultaneously to enhance prediction accuracy, the topic of ensemble modeling in machine learning for medical diagnosis remains relevant. Previous research has proposed a wide range of ML models for analyzing electronic medical records, with most studies using k-means, KNN, and SVM. Research results indicate that the use of an ensemble model that combines the mentioned ML methods leads to an increase in diagnostic accuracy. Even though individual ML methods may be effective in specific areas, their combination into an integrated structure allows overcoming limitations that may arise when using only one method. The advantages and disadvantages of the chosen ensemble model are discussed. The main advantage lies in the increased accuracy of medical diagnosis, which can lead to early detection of diseases and, consequently, improve treatment outcomes. On the other hand, drawbacks may include the complexity of developing and tuning the ensemble approach, as well as computational resource requirements. It is important to note that the obtained results suggest that combining different models can significantly improve the quality of disease diagnosis. This is a significant contribution to the development of the medical field, as it can contribute to the early detection and effective treatment of various pathologies.