Earticle

This paper proposes a binarized neural network (BNN) processor supporting residual networks. The processor was fabricated in UMC 40-nm CMOS technology. Test results show that performance and energy efficiency are 1036.8 GOPS, and 66.5 GOPS/mW at 200MHz, respectively.

In this paper, we developed a cycle-accurate system simulator using Gem5, integrating an AMD GPU simulator and the ROCm toolchain to emulate NPU environments effectively. This integration is crucial for understanding the interaction between CPU and NPU components in advanced computing systems. The simulator's reliability and accuracy were validated by developing and testing a Matrix Transpose kernel compiled with the ROCm toolchain, demonstrating its potential as a valuable tool for research and development in integrated CPU/NPU systems.

As research continues for the commercialization of 5G and 6G communications, the study of MU-MIMO systems in the mm-Wave band is one of the most key research areas in the field of communication algorithm research. MU-MIMO systems utilize a very wide bandwidth in the tens to hundreds of GHz frequency band, which has a significant disadvantage that the power consumption of the RF at the Rx end is very high. In this paper, a hybrid precoding algorithm in a hybrid beamforming and combining system is studied to solve these drawbacks. The proposed hybrid precoding algorithm defines a problem formulation that can eliminate interference by setting constraint conditions and designing an unconstrained precoder, and then derives a hybrid precoder by projecting a square circle into the null space. The proposed algorithm reduces the computational complexity by about 40% or more by reducing the size and number of SVDs compared to existing works.

A hybrid memory integrates various types of memory to enhance performance, control costs, and improve energy efficiency. In this study, we proposed a reconfigurable crossbar for hybrid memory design. To implement the hybrid memory, high bandwidth memory (HBM), dynamic random access memory (DRAM), and block ram (BRAM) were utilized as volatile memories, while magnetoresistive random access memory (MRAM) was used as non-volatile memory. Each memory controller was individually designed and connected to the crossbar. Subsequently, we conducted an analysis of the hybrid memory performance.

This paper analyzed the required specification for implementing a real-time object detection system using the YOLO network. The YOLO network is one of the object detection schemes that strike a suitable balance between performance and required resources, making it widely applicable across various domains. In this paper, we measured the performance based on different activations for an efficient implementation of the YOLO network. Also, we conducted a comparative analysis of the YOLO network's performance concerning input characteristics, aiming to analyze the appropriate application method for real-time object detection.

As many LLMs have been released, modified network layers based on transformer have been researched to improve performance. However, it is essential to design LLMs in a large size for performance, and as a result, current LLMs can only be executed on large servers, and various attempts have been made to reduce the amount of computation. In this paper, we present a method to reduce the amount of computation by using the data attribute of the SwiGLU layer used by meta and google. Since SwiGLU contains an activation function, it generates a large number of near-zero values, and we try to reduce the amount of computation by skipping unnecessary operations. Our experiments show that our algorithm can reduce the computation by 13.3% when there are 20% zeros from activation function.

Stereo vision is actively researched as a solution for distance measurement in autonomous driving. This technique involves triangulation-based distance calculation using left and right images acquired from two image sensors. A kernel window is employed to determine the disparity between these left and right images. To achieve optimal disparity in various environments, it is essential to support different kernel sizes. Moreover, there is a need for research to integrate high-throughput memory, such as high bandwidth memory (HBM), for processing real-time highresolution stereo vision images from sensors. In this paper, we propose a stereo vision accelerator structure utilizing HBM, which supports various kernel sizes.

In this paper, we propose a PTQ (Post Training Quantization) static with QO (Quantization Only) technique to efficiently deploy and execute deep learning models. A comparative performance evaluation of PTQ static with QDQ (Quantize and DeQuantize) and the proposed quantization method was conducted using the MNIST (Modified National Institute of Standards and Technology database) dataset and 8- bit quantization. Experimental results indicate that the PTQ static with QO method reduces the size of the model by approximately 33%, increases the inference speed by 1.5 times, and minimizes the accuracy loss, similar to the PTQ static with QDQ method. The proposed PTQ static with QO method offers a significant technical enhancement to facilitate the efficient deployment and execution of AI (Artificial Intelligence) models through the quantization of deep learning models. We have shown that the PTQ static with QO method is a beneficial and efficient approach to decrease the size and computation of deep learning models. This study makes novel contributions to the quantization of deep learning models. The practical potential of the PTQ static with QO method lies in its ability to be more suitably deployed for the purposes of AI hardware.

Spiking neural networks (SNN), employing eventbased spike computation, can be implemented in hardware where on-chip learning and inference are supported in a powerand area-efficient manner. Although many SNN hardware have been proposed for energy-efficient designs using relatively shallow networks, SNN algorithms that support multi-layer learning need to be implemented in hardware to handle more complex datasets. However, multi-layer learning requires more complicated functions like softmax activation, which makes energy-efficient hardware design difficult. In this paper, we present a zero-spike prediction method to skip the complicated function in the convolution layer. Decomposing the original algorithm, the proposed method skips at least 76.90% of softmax activation operations without classification accuracy degradation.

Global-Local Path Network is a monocular depth estimation network. It presents a new method for integrating global features from an encoder and local features from a decoder through a Selective Feature Fusion module. In this paper, we propose that replacing the SegFormer encoder with the Swin Transformer leads to an improved GLPN, called Swin Transformer-Global-Local-Path-Network. We train the network with modified NYU Depth V2 datasets. Therefore, with the 0.034 RMSE, 0.075 AbsRel, 0.033 log10, 0.951 Delta 1, 0.994 Delta 2, 0.999 Delta 3, our network using a tiny version of Swin Transformer outperforms the previous GLPN model.

Integrating unmanned aerial vehicles (UAVs) and computer vision techniques has contributed to enhancing the accuracy and speed of monitoring for surveillance and warning systems. This paper presents an application of human detection in a beach-warning-system using drone-captured images and YOLO. Our research focuses on detecting critical objects and anomalies on the beach, such as people or buoys. By leveraging the real-time capabilities of YOLO, our system processes highresolution drone images to swiftly identify and classify objects, enabling rapid responses in emergencies. We conducted evaluations using several methods to validate the model's effectiveness. The results showcase its potential to enhance beach warning systems and quick warning of dangerous situations.

3D object detection is widely applied in robotics and autonomous driving; since 3D scenes in autonomous driving are typically outdoor environments, current methods exhibit substantial computational wastage and significant time delays when using convolution directly in the backbone network. This paper proposes a backbone network based on sparse convolutional spatial-semantic fusion modules to solve this problem. High-level semantic features and low-level spatial features extracted through sub-manifold sparse convolution and sparse convolution are fused to enhance feature representation capabilities. Our proposed backbone network achieves excellent performance on the KITTI dataset.

Road torments are ample cause of deterioration of pavements structure. As they are not assessed and recovered on time. Progress can be easily assessed by the strong infrastructure of roads. But this infrastructure requires a lot of effort not only in construction but also in maintenance. In this paper our cause to detect main cause of road destruction because road cracks are the main cause of roads destruction. These cracks are assessed both manually and automatically. On a wide level, manual method of crack detection is used. These cracks are differentiated according to their shapes and severity. This is the crucial part where it is necessary to have a system for the assessment of the crack type. So, the processing of that image is done according to the nature of the crack. As different techniques are applicable to different cracks to assess the severity and nature. On the basis of which precautionary measures can be taken. In this paper I used fuzzy inference system here to process the image through segmentation. So that the most accurate and quick observation can be done on the basis of appropriate segmentation technique applied on the certain image.

The challenge of defect detection in Liquid Crystal Display (LCD) manufacturing is significant. This study proposes a data augmentation technique utilizing Generative Adversarial Networks (GAN) to improve defect identification accuracy. By generating synthetic image data with GAN, the original dataset is expanded, making it more diverse. This augmentation approach aims to improve the model's generalization capability and robustness with real-world data. Unlike traditional data augmentation, GAN-synthesized data provides more realistic and varied data. Experiments show that merging GAN-generated data with the original dataset improves the detection accuracy of critical defects in LCD manufacturing, compared to using the original dataset alone. This method suggests a viable data augmentation strategy for better quality control in LCD production.

The rising prevalence of breast cancer across the globe requires the application of advanced diagnostic techniques for early detection and treatment. This research uses the Wisconsin Breast Cancer Dataset to explore the efficacy of various machine-learning algorithms and ensemble techniques in predicting breast cancer. The study encompasses three significant steps: data retrieval from Kaggle, data preparation through exploratory data analysis, and predictive model formulation and evaluation. Various machine learning algorithms, including Support Vector Machine (SVM) and Random Forest (RF) were employed alongside ensemble techniques like Bagging, Boosting, and a Voting Classifier that integrates multiple models. Feature selection emerged as a pivotal task, enhancing model performance by focusing on significant attributes, thus addressing challenges like high dimensionality and overfitting while promoting model interpretability. The Voting Classifier exhibited the highest accuracy of 98.25%, with varying performance across different feature sets. The insights garnered from feature selection and machine learning models demonstrate promising capabilities for early breast cancer diagnosis, emphasizing the critical role of machine learning in advancing medical data analytics for better healthcare outcomes. This research not only underscores the potential of machine learning in medical diagnostics but also provides a comprehensive exploration of feature selection and ensemble learning in achieving superior predictive accuracy in breast cancer detection.

Automatic skin lesion segmentation in terms of skin lesion analysis is very important. However, it is still a challenging task due to the irregular shapes of the skin lesion. Traditional CNN-based methods usually cannot achieve a satisfactory segmentation performance. We present a novel network with a feature enhanced Transformer for skin lesion segmentation. Unlike earlier CNN-based U-net models, our model utilizes Transformer blocks to capture global and local features, improving the performance of medical image segmentation. By incorporating feature enhancement module at every skip connection layer, we substantially enhance feature fusion capabilities and improve the efficiency of the encoderdecoder structure. In the FEM, a squeeze and excitation attention module is introduced to enhance important feature and suppress unnecessary information. The experimental results show that our proposed model demonstrated the effectiveness on PH2 dataset.

Globally, chronic diseases have a significant impact on health. The diagnosis of chronic diseases has seen extensive usage of machine learning techniques. Early disease detection and treatment lower the risk of increasing disease severity and, consequently, related mortality. The major goal of this research is to provide a technique that increases classification accuracy while also shortening computing time. This comparative research shows the impact of distinct model architectures and features on disease prediction accuracy in addition to assessing the advantages and disadvantages of each technique. These discoveries have implications for personalized healthcare, allowing medical professionals to select the best models for various chronic conditions. Additionally, this research can direct the creation of better forecasting technologies, as well as influence healthcare legislation and budget allocation. In our study comparative analysis of the state-of-the-art approaches has been presented. Using a hybrid model combination of CNN and RNN could be more beneficial. In conclusion, our comparison research improves our comprehension of the potential of deep machine learning for chronic disease prediction, highlighting the significance of adjusting model selection to certain disease types. To progress the field of chronic disease prediction, future research should concentrate on improving these models, and further explore their applicability across various and larger datasets.

Since the onset of the Coronavirus outbreak in December 2019, the virus has infected over six hundred million individuals, resulting in more than six million confirmed deaths, as reported by the World Health Organization (WHO). COVID- 19 is attributed to Severe Acute Respiratory Syndrome Coronavirus 2 (SARS-CoV-2) and is recognized as a respiratory ailment, characterized by symptoms including fever, myalgia, dry cough, headache, sore throat and chest pain. As of October 2022, substantial efforts have been directed toward understanding and combatting the disease, particularly in the domains of vaccination and diagnosis. This paper focuses on the diagnosis of COVID-19 using X-ray images and leverages deep learning technologies. Specifically, we concentrate on employing three convolutional neural network models: ResNet50, InceptionV3 and MobileNetV2. The primary objective is to evaluate their performance in diagnosing COVID-19 from Xray images. During our research, we subjected these models to testing with unseen data. The results revealed that ResNet50 achieved an accuracy of 82.5%, outperforming InceptionV3 with 62.5% and MobileNetV2 with 65% accuracy. The adoption of these models not only alleviates the decision-making burden on medical experts but also enhances the precision of disease classification. The significance of this study lies in its contribution to fine-tuning diagnostic algorithms, paving the way for further research and advancements in the field.

According to new WHO statistics, heart disease is the top reason of death worldwide, killing 17.9 million people each year. This is a growing number. One of the most wellknown issues in clinical offices is that no two professionals have the same knowledge and talent when serving their patients. Researchers are utilizing data mining and machine learning techniques to overcome these difficulties by using predictive analytics to anticipate the risk of heart problems. This study examines the accuracy of various machine learning methods, including Logistic Regression, Naive Bayes, Decision Trees, Support Vector Machines, Neural Networks, and Stochastic Gradient Descent in the prediction of heart disease based on various factors and symptoms such as gender, age, chest pain, and blood sugar using appropriate data. The research entails applying a typical data mining approach to accurately uncover relationships between numerous data sources to predict heart disease. These machine learning algorithms take less time and are more accurate at predicting heart illness, which will lower the global convergence of essential life.

Crop diseases and pests are one of the agricultural disasters that adversely affect the yield and quality of crops. To prevent and control them, many researchers have been working on deep learning-based disease and pest recognition. Most of these studies use classification techniques to output one class with the highest probability from a predefined list of pests. However, the accuracy of classification models is not perfect, and they can produce enough incorrect results to require additional aids. In this study, we proposed a novel disease and pest diagnosis pipeline that combines object detection with a similarity-based retrieval model. In our proposed pipeline, we first detect the damaged region in the image and then classify the class to which it belongs. In the similarity-based retrieval model, the detected region image can be used to further show the user the most similar damage symptom images to help them make a final decision. The pipeline proposed in this study was first applied to three diseases: fire blight, scab, and black necrotic leaf spot.

The rapid of electric vehicles (EVs) is challenges and opportunities for energy grid management and infrastructure planning. This research is aims to fill the knowledge gap by employing advanced analytical methods on a 503-day time series dataset from Thailand's EV charging stations. The dataset includes information on date, station name, connector type, energy consumed in kWh, payment in Baht, vehicle brand and model, as well as customer ID. This study focuses on three main objectives: (1) Forecasting daily energy demand with a focus on the top 5 stations in terms of kWh consumption to identify seasonality and trends, (2) Predicting daily revenue based on energy consumption, and (3) Conducting a Geo-Spatial Analysis to recommend optimal locations for installing new EV charging stations. The insights derived are expected to assist in efficient grid management, revenue planning, and strategic infrastructure deployment.

Typical Text-to-SQL research copes with the problem of parsing natural language questions into executable SQL queries. While new datasets emerge progressively, most are committed to exploring the robustness and generalizability of models, leaving the research of model's capability in solving geography-specific questions blank. To this end, in this work, we approach this problem by proposing an improved abstract syntax tree to represent spatial SQL queries. Following this proposition, we build SpatialSpider, a new Text-to-SQL dataset in the geography domain. Our dataset introduces the following challenges: 1) Understanding of geographical semantics. 2) Prediction of spatial functions. In summary, we propose an improved abstract syntax tree and a dataset to tackle the spatial Text-to-SQL problem and obtain promising experimental results.

Automatic identification of residential areas from remote sensing images is beneficial to tasks such as urban planning and disaster assessment. Currently, residential area extraction tasks are primarily based on deep learning methods using single-modal data, and the information that a single modality can express is limited. Therefore, this paper proposes an end-to-end multi-modal semantic segmentation model that extracts features of remote sensing images and mobile phone signaling data through a dual-branch encoder, fuses the two features, and concatenates them with the feature map of the decoder stage. Experimental results show that our proposed method outperforms other models and can effectively identify residential areas.

This paper introduces a comprehensive approach to dataset standardization aimed at enhancing the effectiveness and reliability of solar power forecasting models. Leveraging multiple datasets, this study incorporates additional attributes such as atmospheric pressure and sunshine duration. These enrichments bridge critical gaps in meteorological and environmental data, facilitating more robust and precise solar power forecasting. The paper underscores the significance of these attributes, furnishes detailed equations for their computation, and presents the outcomes of their integration. It underscores their pivotal role in enabling solar energy stakeholders to make informed decisions and optimize energy production effectively.

In this paper, we introduce a novel solar tracking algorithm designed to optimize power generation by dynamically adjusting the angle of foldable solar panels. This method ensures the optimal capture of sunlight throughout the day, thereby generating the maximum possible solar power by maintaining the incident angle of the solar rays on the panels. The algorithm enables the foldable solar panels to periodically change their orientation based on the time of day, thus maintaining the angle of incident sunlight in the most suitable position for power generation. Comprehensive comparisons between foldable and fixed configurations indicate that foldable bifacial panels demonstrate superior efficiency compared to their fixed counterparts. Additionally, results from the Mann- Whitney U test and correlation coefficients provide critical insights of optimizing solar panel configurations for maximum power output.

In outdoor optical wireless communication systems, weather-induced turbulence affects optical signals, resulting in distortion, thereby, degradation in communication performance. The channel model including turbulence is used for estimating the performance of optical wireless communication system under turbulence. A deep learning algorithm is developed to classify degree of turbulence. This study is based on channel classification using a convolutional neural network for a 4-PSK optical wireless communication system. The channel characteristics are generated following the gamma-gamma distribution. By labeling each data point and distorted constellation for different degrees of turbulence, the deep learning model is trained, and its classification performance is evaluated.

In response to the demand for impartial and precise personality testing, this study presents a unique multi-modal method for predicting personality traits in collaborative settings. Conventional approaches that depend on surveys frequently create biases, which has led to the investigation of raw, subconscious open writing as a rich source of personality information. This study uses deep learning algorithms in conjunction with the stream of consciousness storytelling approach to uncover personality traits by utilizing both textual and gestural data. We use BERT word embedding to improve contextual understanding and convolutional networks for the textual component. Compared to earlier methods, this methodology offers a more dependable way for text-based personality evaluation. Moreover, we present facial recognition as an extra factor for personality evaluation, providing a whole framework with a wide range of uses. We conducted studies in a collaborative setting to assess the effectiveness of our strategy, and we obtained encouraging findings. This multimodal method changes the way people collaborate and opens doors to a wide range of applications, such as mental health diagnosis, job interviews, and forensic investigations. A thorough grasp of personality features is expected to improve personalization and cooperation, leading to more efficient collaboration and improved decision-making.

This paper investigates the performance of ARIMAX-based stock price forecasting models on highmarketcapitalization technology stocks: Apple Inc. (AAPL), Microsoft Corp. (MSFT), and Alphabet Inc. (GOOGL), listed on NYSE and NASDAQ. These stocks were selected due to their substantial impact on both the technology sector and the broader market. Utilizing historical data, we developed three distinct versions of ARIMAX models. Model performance was assessed using key metrics such as Root Mean Square Error (RMSE) and daily direction forecast accuracy. Our results show that the ARIMAX model with a minimal feature set, referred to as Mode 1, generally produced the lowest RMSE values for these specific stocks, indicating superior predictive accuracy. However, while some versions of the ARIMAX model demonstrated promise in predicting the daily direction of stock prices, their performance varied substantially across the evaluated stocks. This inconsistency suggests that further research is needed to identify a universally optimal ARIMAX model for predicting stock price direction. It should be noted that these models were not validated on stocks from other industries or those with different market capitalizations, limiting their generalizability. Additionally, as the models are based on historical data, caution is advised when applying them to predict future stock movements.

Sentiment Analysis is a crucial area of study within the realm of Computer Science. With the rapid advancement of Information Technology and the prevalence of social media, a substantial volume of textual comments has emerged on web platforms and social networks such as Twitter. Consequently, individuals have become increasingly active in disseminating both general and politically-related information, making it imperative to examine public responses. Many researchers have harnessed the unique features and content of social media to assess and forecast public sentiment regarding political events. This study presents an analytical investigation employing data from general discussions on Twitter to decipher public sentiment regarding the crisis in Pakistan. It involves the analysis of tweets authored by various ethnic groups and influential figures using Machine Learning techniques like the Support Vector Classifier (SVC), Decision Tree (DT), Naïve Bayes (NB) and Logistic Regression. Ultimately, a comparative assessment is conducted based on the outcomes obtained from different models in the experiments.

Hierarchical clustering is a widely-used technique in data analysis. Typically, tools for this method operate on data that is in its original, readable form. This poses privacy concerns when dealing with sensitive data that needs to remain confidential. To tackle this issue, we developed a method that integrates CKKS homomorphic encryption, allowing the clustering process to happen without revealing the raw data. However, a challenge emerges when trying to sort the encrypted distances, a crucial step for single linkage clustering. Given the complexities of sorting encrypted data, we propose a cooperative approach: the data owner aids in the sorting process and shares a list of data positions. Using this list, the server can determine how data points cluster together. Our approach ensures a secure hierarchical single linkage clustering process, grouping data without exposing its original content.