Applied and Computational Engineering

- The Open Access Proceedings Series for Conferences

Volume Info.

  • Title

    Proceedings of the 4th International Conference on Signal Processing and Machine Learning

    Conference Date






    978-1-83558-333-3 (Print)

    978-1-83558-334-0 (Online)

    Published Date



    Marwan Omar, Illinois Institute of Technology


  • Open Access | Article 2024-03-15 Doi: 10.54254/2755-2721/46/20241043

    Latency reduction with compression-aware training for efficient distributed computing of Convolution Neural Networks

    To decrease workload on lightweight devices, this project accelerates the computation of Convolution Neural Networks (CNNs) and preserves accuracy through modifying the CNNs’ training process. First, this research implements distributed computing to optimally divide the network workload onto both devices and the cloud. To reduce communication latency between devices and the cloud, this research introduces feature pruning by setting elements in the communicated feature to 0. However, naively pruning the feature causes a significant accuracy drop. To compensate for this limitation, this research applies pruning-aware training to preserve the CNNs’ task performance. This research evaluates the proposed methods on multiple datasets and CNN models, like VGG-11 and ResNet-18 with PyTorch. Empirical results demonstrate that the methods can reduce the computational latency by 50-75% with a negligible 1% accuracy loss. Specifically, this research first identifies the system bottleneck by comparing on-device, on-cloud, and communication latencies (on-device: 14.8%, on-cloud: 1.7%, communication: 83.5%). Then, this research compares multiple pruning strategies and observe the superiority of magnitude-based pruning. At 0.992 sparsity, magnitude-based pruning outperforms other strategies by 45% in accuracy. Finally, this research verifies the effectiveness of the proposed pruning-aware training method by comparing it with the baseline at various splitting points and networks. Pruning-aware training decreases the accuracy loss by up to 26% at 0.998 sparsity. In conclusion, even though distributed computing accelerates applications on lightweight devices, compressing the communication cost is crucial and challenging. This research proposed methods effectively reduce communication latency without sacrificing accuracy, conserving the effectiveness of CNN.

  • Open Access | Article 2024-03-15 Doi: 10.54254/2755-2721/46/20241044

    The impact of big data on the sports industry: Enhancing athlete training, evaluation, and minority empowerment

    The sports industry has witnessed a transformative shift with the advent of big data technologies. This paper explores the profound influence of big data on athlete training and evaluation, examining how it has emerged as a catalyst for empowering athletes from minority backgrounds. By leveraging vast amounts of data, sports organizations and athletes gain valuable insights, make informed decisions, and optimize performance. Traditional, subjective methods of athlete evaluation are being replaced by objective, data-driven approaches that provide a more accurate assessment of performance. Big data also promotes inclusivity within the sports industry by identifying talented individuals from minority backgrounds who may have been overlooked in the past. Several case studies highlight the role of big data in revolutionizing sports, such as its use in basketball for improved decision-making and soccer for optimized training regimens. In conclusion, big data transforms athlete training, evaluation, and inclusivity in sports while shaping the future of the industry.

  • Open Access | Article 2024-03-15 Doi: 10.54254/2755-2721/46/20241053

    An investigation on strategies for optimizing consumer trust in chatbots

    The advancement of artificial intelligence (AI) gave rise to chatbots, which is a type of AI-powered software that communicates via natural language. Chatbots have been used in diverse contexts, delivering significant convenience to the consumers. Nonetheless, this technology encounters ambivalent attitudes from consumers. Some aspects of the chatbot technology are evoking distrustful attitudes among consumers, while the others are cultivating a sense of trust. Thus, the objective of the current paper is to outline and analyze key factors that affect consumer trust and elucidate strategies that firms can adopt to optimize trust. According to recent studies, consumer distrust primarily stems from algorithmic bias, privacy and security concerns, and the lack of algorithmic transparency; on the other hand, consumer trust is formed due to anthropomorphic attributes of chatbots, particularly warmth and competence. To reduce consumer distrust, companies are advised to first identify and minimize existing real risks in their products, then deliver transparency to the public to establish a trustworthy image. To increase trust, companies are suggested to improve upon the anthropomorphic attributes of chatbots. Contributions and limitations of the paper are also discussed to highlight areas that require further investigation in the field of chatbots as well as AI in general.

  • Open Access | Article 2024-03-15 Doi: 10.54254/2755-2721/46/20241064

    The investigation of application related to deep learning on brain tumor diagnosis

    Brain tumor has been a serious disease to human beings for a long time. Brain tumors have posed a significant health threat to humanity for many years. If left untreated in its early stages, a brain tumor can become malignant, drastically reducing survival chances. Throughout the decades, numerous individuals have endured the hardships of brain tumors, and tragically, some have succumbed to this condition. However, deep learning techniques offer a promising avenue for precise and efficient brain tumor diagnosis. Utilizing this technology enables the early detection and treatment of benign tumors, potentially saving lives and preventing unnecessary loss. In this review paper, two previous research on how different deep learning models perform on the brain tumor diagnosis would be illustrated. In the first research, the performance of five models would be compared with each other. In the second research, Convolutional Neural Network (CNN) and Artificial Neural Network (ANN) would be compared with each other. Furthermore, the examination of two research methods will delve into how various techniques can enhance model performance. Deep learning techniques also find numerous real-life applications. The two important applications are Home Diagnosis and In-Hospital Assistance, and the benefits of applying deep learning techniques in these two areas would also be illustrated. In addition, several suggestions would be proposed based on the applications of deep learning technique.

  • Open Access | Article 2024-03-15 Doi: 10.54254/2755-2721/46/20241068

    Prediction of cardiovascular and cerebrovascular diseases based on machine learning models

    Recently we had the fact that cardiovascular disease has become one of the major threats to human life, which leads to the significance of the research around the prevention and cure of such disease. Recently, machine learning algorithms are utilized for the prediction of a certain person who has an illness or not. To verify the effectiveness of predicting cardiovascular disease using machine learning methods, we predict cardiovascular disease given features of a person’s life habits and illness history from the Behavioral Risk Factor Surveillance System. Therefore, 5 models are selected, including SVM, logistic regression, decision tree, fully connected network, and XGBoost to evaluate the performance via confusion matrix and ROG curves. Plus, the dataset is highly unbalanced, so we also implemented SMOTETomek re-sampling algorithms to evaluate the models’ performance on such kinds of datasets. Results exhibited that XGBoost performs the best on the given dataset, hence deep research on improving the performance using XGBoost is highly recommended.

  • Open Access | Article 2024-03-15 Doi: 10.54254/2755-2721/46/20241071

    Machine learning-based readmission risk prediction for diabetic patients

    Among the numerous hospitalized patients with chronic diseases, diabetes patients are under a higher readmission rate, which poses challenges and pressures to both patients and the healthcare system. To predict the likelihood of diabetic patients being readmitted within a short amount of time, this paper utilizes various machine learning-based models for performance analysis and comparison. By selecting appropriate datasets, cleaning and preprocessing data, the models were trained to forecast the probability of patient readmission. The paper compares the performance metrics of six classifiers: XGBoost, logistic regression, GBDT, decision tree, random forest, and deep neural network. The metrics include accuracy, f1 score, precision, recall, and ROC curve. Experimental results demonstrate that XGBoost exhibits better adaptability to complex data and achieves higher mean values of Accuracy (64.43%), f1 score (59.16%), recall (55.9%), and ROC (70.14%) in readmission prediction.

  • Open Access | Article 2024-03-15 Doi: 10.54254/2755-2721/46/20241090

    Prediction of cardiovascular disease based on machine learning

    Cardiovascular disease is one of the deadliest diseases worldwide, causing millions of deaths every year. Major risk factors include hypertension, hyperlipidemia, smoking, unhealthy diet, and lack of physical activity. To achieve a simple and effective prediction of cardiovascular disease, a study comparing the performance of common machine learning algorithms was conducted. The dataset used in this research consists of a population of 70,000 individuals from Kaggle. During the data processing phase, abnormal values within the feature variables were removed, and a BMI feature variable was added to the dataset to visualize the relationships between the data more intuitively. Deep neural networks were used to predict cardiovascular disease and were compared with eight traditional machine learning algorithms with respect to accuracy, F1 score, PR and ROC. The results indicated that the deep neural network (DNN) is the optimal model for predicting cardiovascular disease.

  • Open Access | Article 2024-03-15 Doi: 10.54254/2755-2721/46/20241096

    Federated learning-based machine learning for predicting brain tumor

    The swift advancements in artificial intelligence (AI) and machine learning have profoundly impacted the realm of medical research, particularly in the realm of diagnosing and treating intricate conditions such as brain tumors. These tumors, characterized by unregulated cell proliferation, pose significant challenges. The complexities inherent in brain tumor diagnosis stem from the intricate nature of these tumors, symptom overlap with other ailments, and the inherent complexity of the brain itself. Nevertheless, the application of an advanced machine learning algorithm known as Federated Learning (FL) has demonstrated its potential to address data privacy concerns and enhance diagnostic accuracy in this context. This essay discusses the application of FL which is a decentralized training strategy in brain tumor research. FL allows multiple institutions to train the model collaboratively without data sharing. The key advancement includes the improved U-Net model implementation and the utilization of Convolutional Neural Network (CNN) Ensemble Architectures for brain tumor identification. This paper also discusses the potential of FL in optimizing weight sharing for model aggregation in heterogeneous data. Furthermore, it underscores the important role of FL in modern healthcare since FL also solves the privacy concern in smart healthcare. However, challenges such as communication lag, data heterogeneity, and computational cost still exist.

  • Open Access | Article 2024-03-15 Doi: 10.54254/2755-2721/46/20241098

    Federated learning in autonomous driving: Progress, challenges, and outlook in perception, prediction, and communication

    In the ever-evolving field of autonomous driving, vehicles have evolved into mobile computing centres, accumulating and processing vast amounts of data, including environmental variables, driver behaviours, and preferences. Conventional centralized data processing methods face privacy and security vulnerabilities. To address these challenges, federated learning technology has emerged as a promising alternative, with its decentralized, privacy-preserving architecture. This review explores the application of federated learning in autonomous driving, focusing on perception, prediction, and communication scenarios, including research such as using federated learning to enhance the vehicle’s ability to predict steering angle, object detection, and multimodal sensor data fusion. In addition, this review investigates the improvement of communication efficiency through techniques such as Distributed Federated Learning (DFL), Selective Federated Reinforcement Learning (SFRL), and Vehicle-to-everything (V2X) communication. The analysis indicated that federated learning holds great promise in autonomous driving, significantly enhancing vehicle performance in perception, prediction, and communication. However, challenges like data heterogeneity and communication costs persist. Future research should prioritize refining aggregation algorithms, minimizing communication overhead, and adapting federated learning to evolving autonomous driving technologies.

  • Open Access | Article 2024-03-15 Doi: 10.54254/2755-2721/46/20241106

    Deep learning methods and corresponding applications in medical imaging

    Deep learning is a subfield of artificial intelligence and machine learning, and it is becoming a popular topic in recent years. It is a powerful tool in solving complex tasks and achieve state-of-the-art results in many areas, including language processing, computer vision, and more. This paper briefly introduced two main deep learning models, Convolutional Neural Networks (CNNs) and Generative Adversary Networks (GANs) and their applications in medical imaging. CNNs are often used in image recognition tasks, like separating different organs in one medical image. While GANs are better doing medical image generation tasks, like creating an X-ray image of chest. This study introduced some deep learning methods for image segmentation, image classification and image generation. Not only are the basic CNNs and GANs architectures used, but also some improvements and modification involved. These methods greatly expand the existing medical image datasets. They also save lots of time for doctors and radiologists from labeling and recognizing those medical images. Deep learning methods are super strong in processing complex and numerous medical images. However, there are still some limitations caused by the lack of training datasets and learning models.

  • Open Access | Article 2024-03-15 Doi: 10.54254/2755-2721/46/20241109

    A comparative analysis and investigation of Attn-GAN and SSA-GAN for text-to-image generation

    Text-to-image generation has emerged as a captivating and intricate challenge within the field of artificial intelligence. This paper provided an extensive comparative evaluation of two cutting-edge Generative Adversarial Network (GAN) models, namely Attn-GAN and SSA-GAN, in the context of text-to-image generation. The significance of this problem extends to a multitude of applications, encompassing content generation, advertising, and virtual reality. Attn-GAN, an acronym for Attentional Generative Adversarial Network, leverages attention mechanisms to align textual descriptions with their corresponding image regions.This approach aims to detect feature details and ensure semantic consistency in the generated images. In contrast, SSA-GAN, or Semantic-Spatial Aware GAN, integrates spatial and semantic information into the image generation process to produce visually plausible and semantically meaningful images.This paper provides a detailed examination of the architectures and working principles of both Attn-GAN and SSA-GAN, followed by a comparative evaluation covering image quality, semantic fidelity, and computational efficiency. The results reveal that SSA-GAN excels in generating images with superior semantic consistency and fine-grained details, while Attn-GAN produces diverse and visually appealing images. Furthermore, the study includes practical recommendations for selecting the appropriate model based on specific project requirements. This study conducted experiments on the COCO and CUB datasets, showcasing the strengths and weaknesses of each model in different scenarios. The findings emphasize the importance of understanding these trade-offs when choosing between Attn-GAN and SSA-GAN for text-to-image generation tasks.

  • Open Access | Article 2024-03-15 Doi: 10.54254/2755-2721/46/20241124

    Risk prediction of acute myocardial infarction based on deep neural network

    The cardiovascular disease situation in China is becoming more and more serious, among which the prevalence and mortality rate of acute myocardial infarction (AMI) have been increasing in recent years, so it is urgent to prevent cardiovascular disease. Based on the personal behavior data and basic physical data covering 30,000 respondents within 30,000 days, a deep learning algorithm was used to predict whether the respondents would have acute myocardial infarction in the near future. The prediction accuracy and generalization ability of different algorithms, including logistic regression, decision tree, XGBoost, and GNB algorithms, are compared. The results show that it is feasible to predict the short-term AMI disease split line based on the patient's personal basic information, and the DNN algorithm can achieve the highest prediction accuracy. The research results can assist clinical medical staff in the prevention and treatment of acute myocardial infarction to a certain extent, and also have good civilian characteristics, which can serve medical insurance and other industries.

  • Open Access | Article 2024-03-15 Doi: 10.54254/2755-2721/46/20241164

    Enhancing medical image classification based on super-resolution techniques: A comparative study

    This paper explores the application of super-resolution techniques in the context of medical image classification, focusing on the impact of various super-resolution models on classification model performance. With the burgeoning influence of deep learning in the field of medical image analysis. This research discusses the potential of image super-resolution to enhance the precision of medical image categorization. This paper compares three super-resolution methods: bicubic interpolation, Super Resolution Residual network (SRResnet), and Super Resolution Generative Adversarial Networks (SRGAN), using the OCT2017 dataset for optical coherence tomography (OCT) images. The experiments reveal that deep learning-based super-resolution methods surpass bicubic interpolation in boosting classification model accuracy. Notably, SRResnet stands out as the preferred choice for enhancing classification accuracy, despite yielding less visually appealing results compared to SRGAN. This suggests that while Generative Adversarial Networks (GANs) and perceptual loss functions hold promise in super-resolution, their translation to improved classification model performance may vary. This paper’s findings provide guidance for practitioners and researchers in the field, emphasizing that selecting an appropriate super-resolution method could optimize classification models for medical image analysis.

  • Open Access | Article 2024-03-15 Doi: 10.54254/2755-2721/46/20241177

    Hyperparameter search and performance assessment of convolutional neural networks for image classification

    As the fields of deep learning, computer vision gain increasing popularity. Image classification is a fundamental task in computer vision that aims to understand images holistically and classify them into specific categories. Various types of image classification algorithms based on convolutional neural networks (CNNs) have been developed and evaluated on multiple image datasets. During the training of many image classification algorithms, several factors can significantly impact the results and even the learning and training process. These factors include the complexity of the image inputs, the optimizer method used, and the parameter tuning techniques applied. In this article, the author conducted experiments using ResNet-18, a residual network learns residual functions to avoid overfitting, with different combinations of learning rates and optimizers. After analyzing the experimental results, it could be observed that the extent of input complexity can indeed affect the accuracy of the model’s results, as well as its convergence behavior during the training process.

  • Open Access | Article 2024-03-15 Doi: 10.54254/2755-2721/46/20241180

    CNN advancements and its applications in image recognition: A comprehensive analysis and future prospects

    With the rapid advancement of industrialization, it has become increasingly evident that conventional image recognition methods are inadequate to meet the burgeoning demands for processing vast quantities of engineering image data. In response to this challenge, convolutional neural networks (CNNs), a cornerstone of deep learning algorithms, have emerged as powerful tools for feature extraction within the realm of image recognition. This paper endeavors to elucidate the underpinnings of image recognition theory and the intricacies of convolutional neural networks. It not only provides an in-depth exploration of the structural characteristics of three iconic CNN models, along with illustrative real-world applications but also delves into a comparative analysis of the practical efficacy and constraints associated with various advanced optimization techniques applied to CNN models. The paper's central focus lies in assessing the real-world applications of CNNs, particularly in the domains of medical diagnosis, agricultural disease identification, and transportation systems. It scrutinizes the impact and limitations of these advanced optimization techniques in each context. Subsequently, it culminates in a comprehensive discussion of the challenges that lie at the nexus of CNN technology and image recognition, providing insights into the prospects of this dynamic intersection.

  • Open Access | Article 2024-03-15 Doi: 10.54254/2755-2721/46/20241190

    Comparison of deep learning-based models for retinal vessels segmentation

    With the trend of social diversification and the rapid popularization of information, more and more people use mobile phones to deal with matters in life, study and work, as well as a place of entertainment in their leisure time. People use their eyes at high intensity for a long time. In addition, Low eye care attention greatly increases the chance of developing eye diseases. Such lesions are closely linked to eye diseases such as glaucoma and diabetes, and these lesions are the main cause of blindness. The retina has a complex structure, diverse shapes, and high curvature. It is time-consuming and difficult to diagnose diseases from retinal images. Doctors’ subjective judgments on images account for a large proportion. In order to make segmentation performance more efficient and reduce manual segmentation errors, an intelligent system that can automatically extract retinal blood vessels from fundus images has become an urgent problem. Medical picture segmentation and the merging of machine learning have gained popularity as a study area with the growth and attention of deep learning. Various models are applied to accurate segmentation of the retina. Through computer feature extraction and sample training of images, it can be directly learned from many data samples. The development of this technology saves time and cost, and at the same time meets the medical needs in remote areas. Doctors can efficiently screen and diagnose through the vascular status of the fundus to determine diabetic retinopathy, glaucoma, microaneurysms and other eye disease fundus manifestations.

  • Open Access | Article 2024-03-15 Doi: 10.54254/2755-2721/46/20241222

    Comparative analysis of typical UCB algorithm performance

    Amidst the burgeoning Internet service industry, there's an escalating demand for robust recommendation systems. To cater to this need, this study meticulously examines the UCB algorithm, renowned within the Multi-Armed Bandit (MAB) paradigm. Through a meticulous comparative analysis, distinctions between the classic UCB approach and its modern counterpart, the randomized UCB, are drawn, with an emphasis on their performance on real-world datasets. The empirical findings accentuate the proficiency of the randomized UCB. It showcases a measured growth rate and a notably reduced overall regret. These results are more than mere statistical data; they attest to the randomized UCB's unparalleled efficiency in practical environments. Furthermore, the insights gleaned can potentially spur cutting-edge developments in recommendation system algorithms, setting a new benchmark in the domain. Conclusively, as the digital realm remains ever-evolving, this research vehemently advocates for the relentless refinement of algorithms, ensuring they remain adept at navigating the intricacies of the modern digital landscape.

  • Open Access | Article 2024-03-15 Doi: 10.54254/2755-2721/46/20241226

    Comparing the influencing factors of M/G/1 performance indices in queuing theory across different scheduling approaches

    The M/G/1 queue holds significant research value within the realm of queuing theories and systems due to its broad applicability. However, the multifaceted nature of the M/G/1 queue makes its characteristics and scheduling challenges particularly intricate. Key performance metrics include the average response time, average waiting time, system expectancy, and queue expectancy. This study predominantly concentrates on the average response time, average waiting time, and the increasingly emphasized metric, average slowdown. These metrics provide a more holistic view of system performance, unhindered by system size variations. Appropriate scheduling can do more than just decrease the average response and wait times; it holds a specific relevance to achieving equitable scheduling. Among the myriad metrics employed to gauge system efficiency, the fairness index is gaining traction. The crux of this investigation centers on this metric, aiming to delineate the discrepancies in average slowdown between non-preemptive and preemptive scenarios. Furthermore, a delve into the z-transform within the M/G/1 system will unravel the intricacies of the update-reward mechanism, illuminating the performance oscillations of the M/G/1 system across varied timelines.

  • Open Access | Article 2024-03-15 Doi: 10.54254/2755-2721/46/20241243

    Performance comparison of deep learning-based image classification algorithms on ImageNet

    The rapid evolution of Artificial Intelligence (AI) technology has propelled image recognition to the forefront of computational advancements. Since the inception of Convolutional Neural Networks (CNNs), the field has expanded into a multitude of sophisticated models and their derivatives, each tailored to address specific challenges and applications. Image recognition's landscape encompasses foundational tasks such as object and face detection, extending to more specialized applications like emotion analysis, optical character recognition, and complex interpretation of biological imagery. This domain's historical perspectives trace back to models like AlexNet, which set benchmarks with accuracy rates of around 70%. Fast forward to contemporary times, and advanced algorithms consistently achieve accuracy figures beyond the 90% threshold on benchmark datasets like ImageNet. Moreover, the diversification of AI applications has led to the development of models like MobileNet, which are intricately designed for streamlined efficiency on mobile devices, balancing performance with resource constraints. This discourse will navigate the intricate maze of image recognition, primarily leveraging insights from the ImageNet dataset as a canonical reference. By the end of this exploration, this work will discuss several cost-efficient models. Finally, this work will also cover some complex algorithms with high accuracy. All these algorithms use different approaches and obtain good performance in either cost-efficiency or accuracy. This discourse will provide an overview of these algorithms, detailing their novelty, implementation, and experimental results for accuracy and cost-efficiency.

  • Open Access | Article 2024-03-15 Doi: 10.54254/2755-2721/46/20241296

    Machine learning-based hotel occupancy prediction

    The online booking system has tremendously facilitated people's ability to travel and make reservations because of the growth of the Internet. However, the unpredictability of travel led to frequent changes in reservations. Frequent order changes can lead to many problems, such as hotels not being able to get order changes in a timely manner and more in-demand customers not being able to stay, which can lead to lower profits and occupancy rates. In addition to this, there are a number of subjective, such as changes in the trip, reasons for work, and reasons for family, and objective, such as weather changes, natural disasters, and Transportation issues, factors that make it more difficult to predict the occupancy rate. In recent years, machine learning has become an increasingly valuable tool for researchers to analyze data. Based on these, this paper summarizes the machine learning algorithms related to hotel occupancy probability prediction and analyzes and compares them. Finally, it gives an outlook on the research of the hotel occupancy rate prediction.

Copyright © 2023 EWA Publishing. Unless Otherwise Stated