Current AI system development processes often follow a fragmented approach, where non-ML and ML components are typically developed in isolation. This separation poses significant challenges, particularly in managing changes effectively. Our study delves into these challenges, highlighting how the lack of co-development can obscure and exacerbate issues within the change management process. Through an in-depth exploratory case study, we meticulously model, execute, and monitor the current (As-Is) development process. Our findings lay the groundwork for an integrated future (To-Be) process, aimed at enhancing change management practices
Today, many software applications have been developed based on a microservices architecture composed of multiple components that can be updated individually without affecting the entire application. Such applications can be deployed on Kubernetes clusters which provide flexibility, scalability, and ease of management. State-of-the-art strategies for deploying microservices in Kubernetes clusters, such as ramped, bluegreen, canary, A/B, or shadow strategy are designed with high availability, stability, and easy rollback. However, these strategies are not energy-efficient because they require keeping both the original and the update version of the application running simultaneously. In this paper, we address this issue by proposing a solution to reduce the update time of microservices applications, thereby decreasing the interval of time in which two versions of the application on the infrastructure. We design an optimized scheduling mechanism for updating microservices components as a Mixed Integer Linear Programming (MILP) model. We mathematically formulate an energy consumption model for the update process and then propose a Greedy algorithm to find the solution in real-time. Simulation results show that our Greedy algorithm approximates the optimal solution provided by a mathematical solver in only a fraction of time.
In recent years, Legal Judgment Prediction (LJP) has attracted a lot of interest from both academic and legal practitioners. Generally, LJP research has three subtasks, i.e., applicable law article prediction, charge/judgment prediction, and term of penalty prediction. Among those three subtasks, the majority of LJP research focuses on charge prediction as one of three tasks of LJP. Charge prediction itself is a single label prediction problem. Unlike the majority, this research proposes a multi-label classification on LJP that predicts violations of relevant law articles in a legal case with the help of Natural Language Processing (NLP). Data used in this research is collected from the publicly accessible legal document in the European Court of Human Rights (ECHR). In detail, the data includes “facts” of each legal case that describes how the events have occurred. Legal BERT (LBERT) embedding is utilized to generate word embedding and Label Powerset is used for problem transformation. Lastly, we apply conventional Machine Learning (ML) classifiers such as Logistic Regression (LR), Decision Tree (DT), and Random Forest (RF) to predict which relevant law articles have been violated. In total, there are 13 law articles in the ECHR dataset. The proposed model achieves over 86.51% precision, 81.77% recall, and 82.89% F1score on ECHR dataset.
The parameterized knowledge within large language models (LLMs), like ChatGPT, offers a significant opportunity for modelling domain knowledge base from text. However, LLMs' context sensitivity can hinder obtaining precise and taskaligned outcomes, thus requiring a suitable design for leveraging prompt engineering. This study explores the efficacy of different prompting methods for RDF knowledge graph construction from medical documents as our preliminary investigation, aiming to develop an efficient pipeline for a large-scale automatic knowledge graph construction according to semantic web standards and technologies. The results show that leveraging in-context learning within LLMs is capable of extracting an array of precise RDF triples from text. We perform a qualitative analysis of the extracted triples with different prompt templates, giving insights that could guide potential development in the research field.
Happiness relates to life satisfaction and can impact an individual's activities in daily life. For students, it impacted their learning activities in and outside the school environment. The report by OECD in 2019 has indicated some East Asian countries and economies did not achieve the average life-satisfaction. Despite not achieving the average lifesatisfaction score, East Asian students has been top performers in Programme for International Student Assessment (PISA). This paper focuses on studying aspects related to happiness with respect to the academic achievement of East Asian students in PISA. The data from PISA 2018 is extracted, integrated and analyzed to identify factors that may have impact towards happiness among East Asian students. Secondly is to ascertain the relationship between level of happiness and academic achievements across Reading, Math and Science domains in PISA. The main findings reveal that the top three correlated factors are frequency of being happy, subjective well-being and frequency of being joyful with each of the three factors having a correlation value of above 0.4 which indicate positive correlation with happiness. From this study it is suggested that activities in and outside the school environment that led to the top 3 correlated factors can be planned to inculcate a sense of happiness among students.
Today, agreements and contracts play a crucial role in daily life. Early detection of potential risks in contracts can help avoid later disputes and damages for both sides. With the rapid advancements of artificial intelligence in law, a risk detecting system from contracts would reduce the need of legal expertise, save time, and increase the automation. This paper presents our solutions to the problem called “Vietnamese legal warning”. To this end, our work focuses on creating a dataset of Vietnamese contract documents and building an automated system that is capable of identifying and highlighting legal warnings in contracts. Our dataset consists of 1000 annotated contract documents of various types. Our warning detecting system consists of three steps: an optical character recognition (OCR) solution with sophisticated deep learning models to accurately convert scanned legal materials into machinereadable texts; unsupervised techniques to correct spelling errors of the OCR step; and the legal warning detection with different learning methods ranging from rule-based, HMMs, CRFs to recent advanced deep net architectures like LSTMs and BERT. The empirical study showed that our solutions gave competitive results in comparison with the existing methods, therefore demonstrating the effectiveness and potential benefits of automated legal warning detection systems.
This paper investigates the legal question-answering (QA) task in Vietnamese. Different from prior studies that only report results on the task of machine reading comprehension (MRC), we compare the strong QA models in two scenarios: MRC (span extraction) and answer generation (AG) (text generation). To do that, we first created a new dataset, namely ViBidLQA, using the bidding law. The dataset is synthesized by using a large language model (LLM) and corrected by two domain experts. After that, we train a set of robust MRC and AG models on the ViBidLQA dataset and predict on both ALQAC and the test set of ViBidLQA. Experimental results show that for the MRC scenario, vi-mrc-large achieves the best scores while for the AG scenario, ViT5 obtains good performance. The results also indicate that the new ViBidLQA dataset contributes to improving the performance of MRC models for domain adaptation on ALQAC11Code and data: https://github.com/ntphuc149/ViLQA.
Recommender systems (RSs) are crucial in aiding user decisions by providing precise and easily interpretable recommendations. Traditionally, approaches such as contentbased filtering, collaborative filtering, and hybrid methods have been widely adopted in RSs. However, the emergence of Deep Reinforcement Learning (DRL), a branch of machine learning, has demonstrated significant promise in various applications, prompting researchers to explore its potential in RSs. This research applies DRL to improve RS performance by understanding users' preferences and catering to their needs. The study utilizes product information, including features, and historical purchases, to construct a knowledge graph, which forms a basis to use as an environment essential for the Markov Decision Process in the DRL framework. This environment facilitates the generation of a policy that predicts the future products users might be interested in. The findings indicate that DRL-based RSs offer improved performance metrics. By leveraging DRL, RS dynamically adapts to user behavior and preferences, resulting in more personalized recommendations.
C++ is a powerful language that provides numerous flexible and efficient features, however, it also presents significant challenges in detecting and correcting errors, especially in objectoriented projects. While a test case for a procedure program includes a sequence of input values to be provided during execution along with the expected outputs, test cases for classes under test must also consider the state of the object on which the method is invoked. To overcome this difficulty, Genetic Algorithm are utilized to generate test cases for class unit testing automatically. However, due to its reliance on randomness, Genetic Algorithm faces difficulties in solving complex conditions in conditional blocks, resulting in lower code coverage. Meanwhile, solving complex conditions in conditional blocks is precisely the strength of Concolic Execution. This paper proposes a method, named GenColic, that integrates Concolic Execution into Genetic Algorithm to generate automated test data for object-oriented C++ projects. The core focus of this method is to insert a Concolic Execution Phase after the selection process of the Evolutionary loop in Genetic Algorithm. By using Concolic Execution, a new test case that covers an execution path not yet covered by the test suite is added. This results in offspring with better coverage, which continues to undergo Crossover and Mutation. Experiments and comparisons were conducted on Concolic Execution, GenColic, and Genetic Algorithm applied to object-oriented C++ open-source projects. The results show that GenColic can generate test suites with higher branch and statement coverage.
This research focuses on detecting dead laying hens in crowded coops, where manual inspection and monitoring are challenging. To address this issue, the study proposes using thermal cameras to monitor the laying hens' body temperatures. The implemented algorithm consists of two main phases: identifying regions of interest based on supervised learning and laying hens-dead decision-making based on computer vision. Logically, dead laying hens typically have lower body temperatures than living ones, thus a deep learning model based on the YOLOv8 network is used to quickly and accurately identify the location of laying hens at risk of being dead. The results are then fed into the Optical Flow algorithms to detect stationary objects, in order to make a final determination of whether there are any dead laying hens. Experimental research on a dataset consisting of images and videos of laying hens coops achieved an mAP@50 of 0.93. Subsequently, the study discusses the practical deployment of this model in a real farm setting, addressing the challenges and considerations for effective implementation.
Reinforcement Learning has rapidly evolved from theoretical concepts to practical applications, particularly in high-stakes fields such as autonomous driving, healthcare, and logistics, enhancing both efficiency and decision-making capabilities. However, as these applications engage with more complex environments, they face significant challenges from adversarial inputs. These inputs can jeopardize system reliability and safety by inducing unpredictable and potentially hazardous decisions. In this paper, we examine the vulnerability of reinforcement learning systems to such adversarial threats by presenting our novel mitigation approach. We introduce a concept referred to as an imposter—carefully crafted adversarial noise added to the state space that can, for instance, cause a Lunar Lander to crash or a Bipedal Walker robot to lose balance. Our approach leverages entropy, a measure of information content, to differentiate genuine data from adversarial noise effectively. We assess the efficacy of this method through the evaluation of different entropy measures: single entropy, joint entropy, and Kullback-Leibler divergence, across two practical case studies involving the Lunar Lander and the Bipedal Walker robot. The results demonstrate that the entropy metric is particularly effective in detecting and eliminating imposter features, thereby preserving the integrity and safety of critical missions.
This paper investigates the “Gravity in Mind” theory and its impact on enhancing intelligence through the lens of informational dynamics. Drawing analogies from the physical world where gravity influences the motion of objects, the study explores how informational gravity affects mental processes and decision-making. By integrating the resignation mechanism in AI players, we observed significant improvements in decision-making efficiency and overall game dynamics. The findings suggest that higher informational acceleration (a) enables earlier achievement of minimal objectivity, aligning subjective experiences with objective outcomes. This framework offers new insights into understanding and optimizing decision-making processes, providing a novel approach to enhancing intelligence in both artificial and human contexts.
To ensure the software quality, testing methods aim at both targets of achieving high code coverage and error detection capability. Among various testing methods, the Concolic Testing has proven effective in generating test data that gains high code coverage. However, the error detection capability of the test data is limited, especially in cases such as with boundary values. In this paper, we propose a method named BEC that combines the Concolic Testing method with the Boundary Value Analysis (BVA) and Equivalence Class Partitioning (ECP) to improve the error detection capability of the generated test data by the Concolic testing method. The key idea of the method is to extract the boundary values from simple conditions of the given unit source code. From these values, BEC generates test data using the combination of the BVA and ECP methods. These test data help improve the error detection capability of the test data set generated by the Concolic testing method. Experimental results indicate that BEC can dramatically detect more errors than the Concolic Testing method. We give some discussions about the proposed method in the paper.
The more technology advances, the more scientific articles are published, which places pressure on scientists to absorb information efficiently. Because abstract sections within scientific papers are often influenced by the authors' perspectives, they may not provide a comprehensive view of the document. In this paper, we propose an architecture that generates a comprehensive summary of scientific documents. The main idea is to create summaries that capture the author's perspectives while incorporating insights from the research community. In our architecture, we used both the content of scientific documents and their citation networks as input. Since the citation network holds valuable information curated by humans and mirrors the document's parts of interest to the research community, we used citation sentences to identify significant sentences within the document. Subsequently, we compress the important sentences using a pre-trained summarization model and augment them for the abstract section. Evaluation of our proposed architecture on the CL-SciSumm dataset demonstrates its superior performance. To the best of our knowledge, the proposed architecture outperforms existing models.
This research presents a novel approach to stroke prediction by applying machine learning techniques to electronic health records (EHR). The primary focus is on data preprocessing techniques that enhance the performance of machine learning models. These techniques consist of feature extraction based on stroke domain knowledge, encoding categorical variables, data categorization, and handling missing values. The effectiveness of these preprocessing techniques is evaluated using various machine learning models, including Decision Trees, Random Forests, K-Nearest Neighbors (KNN), and Naive Bayes. Performance metrics, such as accuracy, precision, recall, and F1-Score are used for evaluation. The findings highlight the significant role of data preprocessing in stroke prediction and suggest that this approach could be generalized to other medical conditions. Future work will explore more advanced preprocessing techniques and machine learning models to further enhance stroke prediction.
Environmental data collected from IoT devices can reflect personal information of users in the system. This paper represents a cryptographic lightweight method for data privacypreserving and integrated it to build a privacy-preserving solution for IoT data in a 2-tier IoT network. IoT system performance is guaranteed due to the linearity of the proposed symmetric cryptosystem and cryptosystem security is guaranteed because the encryption key is continuously changed using the proposed fast key generation algorithms. This privacy-preserving approach can be effectively applied in various IoT-based smart home systems, ensuring that sensitive user data, such as location and personal habits, remains secure. Furthermore, it is highly suitable for healthcare IoT networks, where protecting patient data privacy is critical while maintaining real-time system performance.
Templates are a powerful feature in C++ for writing generic programs that can work with different data types. This paper presents a method named ITEC to generate test data automatically for unit testing in C++ projects using Templates, which is an improvement on an existing method to handle Templates. By executing a preprocessing phase before applying the testing method, suitable data types for Template parameters are extracted, which can provide additional information for the testing process. For each set of suitable data types, the Template unit codes become unit codes with different concrete data types, and then ITEC generates test cases automatically for these codes using dynamic symbolic execution method. ITEC has been implemented and tested on several open-source projects. Experimental results show that ITEC achieves a significant increase in the average coverage with an adequate number of test cases for source code using Template compared with the old existing method. In addition, we also give some discussions about ITEC and experimental results in the paper.
This study introduces an enhanced method for Content-based Image Retrieval tailored for fish species identification. The proposed approach integrates Zernike Moments Invariant to capture global shape features and Local Directional Pattern to extract local texture features. Additionally, hue information from the HSV colour space is incorporated to enhance retrieval accuracy. By combining these features, the method achieves invariance to rotation, translation, and scaling, addressing the challenges posed by the varying orientations and sizes of fish images in natural settings. The Fish4Knowledge dataset, comprising 27370 images, served as the test bed for the proposed system. Our method achieved an improvement of Mean Average Precision of 12.69 %, significantly outperforming existing techniques. The use of a comprehensive 168-dimensional feature vector, formed by the fusion of colour, texture, and shape descriptors, contributed to the high retrieval performance. Experimental results demonstrated the robustness of the proposed method in dealing with complex backgrounds and various fish poses. This method offers a promising solution for accurate fish species identification, with potential applications in fisheries management and aquatic research.
3D facial reconstruction from monocular images is a challenging task, with applications ranging from virtual avatars to emotion analysis. While recent methods have shown impressive results in reconstructing facial geometry, accurately capturing and manipulating subtle emotional expressions remains a challenge. This paper presents a novel pipeline for generating 3D faces with new, blended emotional expressions by combining existing expressions. Our pipeline consists of two main phases: (1) Emotion Prediction, using a stacked model to classify emotions from a single image, and (2) Correlation-Driven Expression Blending, a novel method for combining emotional parameters based on their correlations with specific emotions. We demonstrate the effectiveness of our approach in generating realistic and emotionally accurate 3D facial expressions, outperforming traditional linear interpolation techniques. Our method is particularly effective in capturing subtle emotional nuances, especially in the context of Asian facial features, where expressions are often conveyed more subtly through eye movements. However, the accuracy of the emotion prediction phase is crucial for the success of the blending phase, as mislabeling the initial emotion can lead to inaccurate or unexpected results.
Generating high-quality adversarial examples for object detectors is essential for evaluating and enhancing their robustness. This paper presents an efficient method to produce such adversarial examples with minimal computational cost. The key idea is to define an objective function and iteratively adjust the input image to minimize this function. The proposed method then enhances the quality of the adversarial examples in terms of $\left\vert L_{0}\right\vert,\left\vert L_{2}\right\vert$, and SSIM. Experiments conducted on YOLOv5 demonstrate the effectiveness of the proposed method. Results on the COCO dataset reveal substantial improvements. Specifically, $\left\vert L_{0}\right\vert$ is reduced from approximately 400 k to $30 \mathrm{k},\left\vert L_{2}\right\vert$ is reduced from about 4 to 2, and SSIM is increased from around 0.5 to 0.99. These improvements are achieved with significant performance gains. The findings illustrate that the proposed method is both effective and practical for real-world applications.
