Compensating for Scarce Historical Images in Cross-Domain Cultural Heritage Retrieval Using Synthetic Aging
研究通过合成老化图像来补充稀缺的历史图像，以改善跨域文化遗产检索。使用EfficientNetV2-M模型评估了真实与合成图像混合训练的效果。
0 citationsRead paper
Selected work
Representative Papers
Against Explainable Artificial Intelligence In Law: Why Justifiable Ai Matters. A Credit Scoring Example
This study addresses the limitations of existing explainable artificial intelligence (XAI) approaches in legal contexts such as credit scoring, which prioritize technical interpretability while neglecting legal justifiability and thus inadequately safeguard creditors’ rights. To overcome this gap, the paper proposes a novel paradigm—“justifiable AI”—that embeds legitimacy requirements from the European Union’s legal framework directly into the core of AI decision-making. By integrating legal text analysis, model explanation techniques, and compliance assessment, this approach constructs a coherent argumentation system that harmonizes legal and technical reasoning. The proposed method transcends conventional XAI constraints, offering a viable solution for high-risk AI applications that simultaneously ensures legal validity and technical feasibility, thereby substantively strengthening the protection of creditors’ rights.
0 citationsRead paper
These Aren't the Reviews You're Looking For How Humans Review AI-Generated Pull Requests
This study investigates differences in code review practices between AI-generated and human-authored GitHub pull requests (PRs), challenging the assumption that conventional review metrics adequately reflect human oversight in AI-assisted development. Leveraging the AIDev dataset, the authors conduct a large-scale empirical analysis to systematically compare review behaviors for both types of PRs within the same repositories and categorize interaction patterns between human developers and AI agents. The findings reveal that the majority of AI-generated PRs receive no human review; when reviewed, they are predominantly assessed by AI agents. Humans tend to directly evaluate human-authored PRs but engage indirectly with AI-generated code by guiding AI reviewers rather than performing direct assessment themselves. These results indicate a fundamental shift in code review structures within AI-integrated software workflows.
0 citationsRead paper
YOLO Object Detectors for Robotics -- a Comparative Study
This study systematically evaluates the performance and robustness of various YOLO models for object detection in robotic workspaces. By constructing a custom dataset tailored to robotic scenarios, integrating the COCO2017 benchmark, and incorporating image distortions to simulate real-world deployment conditions, the work presents the first comprehensive comparison of different YOLO variants in this specific context. The experimental results reveal significant differences among the models in terms of accuracy, inference speed, and resilience to visual perturbations. These findings provide empirical evidence and practical guidance for selecting appropriate YOLO architectures in robotic vision systems, balancing trade-offs between detection precision, computational efficiency, and robustness under realistic operating conditions.
0 citationsRead paper
Linux Kernel Recency Matters, CVE Severity Doesn't, and History Fades
This study investigates key factors influencing the delay in fixing vulnerabilities in the Linux kernel, with a focus on the roles of Common Vulnerability Scoring System (CVSS) severity ratings and the age of kernel versions. Leveraging survival analysis, CVE metadata mining, Git commit tracing, and patch delay statistics, the research systematically examines the dynamics of vulnerability introduction and remediation. The findings reveal that kernel version recency serves as a strong predictor of patch latency: developers prioritize fixing vulnerabilities in newer kernel versions, while older versions often retain unpatched CVEs for extended periods. In contrast, CVSS severity scores exhibit little to no correlation with repair timelines. These results highlight the distinctive nature of Linux kernel vulnerability management and challenge conventional severity-based prioritization strategies commonly adopted in software maintenance practices.
0 citationsRead paper
Recent publications
Latest Papers
Compensating for Scarce Historical Images in Cross-Domain Cultural Heritage Retrieval Using Synthetic Aging
研究通过合成老化图像来补充稀缺的历史图像，以改善跨域文化遗产检索。使用EfficientNetV2-M模型评估了真实与合成图像混合训练的效果。
0 citationsRead paper
Against Explainable Artificial Intelligence In Law: Why Justifiable Ai Matters. A Credit Scoring Example
This study addresses the limitations of existing explainable artificial intelligence (XAI) approaches in legal contexts such as credit scoring, which prioritize technical interpretability while neglecting legal justifiability and thus inadequately safeguard creditors’ rights. To overcome this gap, the paper proposes a novel paradigm—“justifiable AI”—that embeds legitimacy requirements from the European Union’s legal framework directly into the core of AI decision-making. By integrating legal text analysis, model explanation techniques, and compliance assessment, this approach constructs a coherent argumentation system that harmonizes legal and technical reasoning. The proposed method transcends conventional XAI constraints, offering a viable solution for high-risk AI applications that simultaneously ensures legal validity and technical feasibility, thereby substantively strengthening the protection of creditors’ rights.
0 citationsRead paper
These Aren't the Reviews You're Looking For How Humans Review AI-Generated Pull Requests
This study investigates differences in code review practices between AI-generated and human-authored GitHub pull requests (PRs), challenging the assumption that conventional review metrics adequately reflect human oversight in AI-assisted development. Leveraging the AIDev dataset, the authors conduct a large-scale empirical analysis to systematically compare review behaviors for both types of PRs within the same repositories and categorize interaction patterns between human developers and AI agents. The findings reveal that the majority of AI-generated PRs receive no human review; when reviewed, they are predominantly assessed by AI agents. Humans tend to directly evaluate human-authored PRs but engage indirectly with AI-generated code by guiding AI reviewers rather than performing direct assessment themselves. These results indicate a fundamental shift in code review structures within AI-integrated software workflows.
0 citationsRead paper
YOLO Object Detectors for Robotics -- a Comparative Study
This study systematically evaluates the performance and robustness of various YOLO models for object detection in robotic workspaces. By constructing a custom dataset tailored to robotic scenarios, integrating the COCO2017 benchmark, and incorporating image distortions to simulate real-world deployment conditions, the work presents the first comprehensive comparison of different YOLO variants in this specific context. The experimental results reveal significant differences among the models in terms of accuracy, inference speed, and resilience to visual perturbations. These findings provide empirical evidence and practical guidance for selecting appropriate YOLO architectures in robotic vision systems, balancing trade-offs between detection precision, computational efficiency, and robustness under realistic operating conditions.
0 citationsRead paper
Linux Kernel Recency Matters, CVE Severity Doesn't, and History Fades
This study investigates key factors influencing the delay in fixing vulnerabilities in the Linux kernel, with a focus on the roles of Common Vulnerability Scoring System (CVSS) severity ratings and the age of kernel versions. Leveraging survival analysis, CVE metadata mining, Git commit tracing, and patch delay statistics, the research systematically examines the dynamics of vulnerability introduction and remediation. The findings reveal that kernel version recency serves as a strong predictor of patch latency: developers prioritize fixing vulnerabilities in newer kernel versions, while older versions often retain unpatched CVEs for extended periods. In contrast, CVSS severity scores exhibit little to no correlation with repair timelines. These results highlight the distinctive nature of Linux kernel vulnerability management and challenge conventional severity-based prioritization strategies commonly adopted in software maintenance practices.
0 citationsRead paper