Beyond Patches: Mining Interpretable Part-Prototypes for Explainable AI

Humans easily apply learned skills to different situations, a flexibility that AI systems still struggle to achieve. Current AI models are often confined to their training setup, leading to isolated developments and a narrow scope of application. This largely restricts the creation of flexible and general-purpose AI systems. Deep Model Reuse presents a novel solution. Imagine tapping into a vast library of pre-trained models, each a master in its specialized domain. Our approach re-purposes these existing models, extracting and transforming their knowledge for the development of novel AI systems. In this talk, we explore the essential techniques of this transformative process, highlighting the shift towards versatile and efficient AI that mirrors human cognition's adaptability. We introduce three foundational pillars of deep model reuse: understanding, composing, and refining. First, we investigate the internal behavior of neural networks—using language models as explainers and analyzing the representation space of diffusion models—to uncover how and what models have learned. Second, we develop methods to transform and compose models through weight mapping, knowledge distillation, and model dissection, enabling the creation of new capabilities by reassembling existing expertise. Third, we enhance reliability by editing model behaviors and mitigating biases, ensuring robustness in complex and dynamic environments. We demonstrate the power of this paradigm in generative AI, where model reuse leads to efficient diffusion models free from spectral bias, improved compositional understanding in video generation, and the repurposing of 2D/3D models for 3D/4D content creation. By shifting from training from scratch to intelligently reusing and recombining models, we move closer to adaptive, scalable, and human-like AI systems—ushering in a new era of sustainable and general intelligence.

Modern AI applications have caused broad societal implications across key public domains. While previous research primarily focuses on individual user perspectives regarding AI systems, this study expands our understanding to encompass general public perceptions. Through a survey (N = 1506), we examined public trust across various tasks within education, healthcare, and creative arts domains. The results show that participants vary in their trust across domains. Notably, AI systems’ abilities were evaluated higher than their benevolence across all domains. Demographic traits had less influence on trust in AI abilities and benevolence compared to technology-related factors. Specifically, participants with greater technological competence, AI familiarity, and knowledge viewed AI as more capable in all domains. These participants also perceived greater systems’ benevolence in healthcare and creative arts but not in education. We discuss the importance of considering public trust and its determinants in AI adoption.

Named entity recognition is an important basic task in the field of natural language processing. The current mainstream named entity recognition methods are mainly based on the deep neural network model. The vulnerability of the deep neural network itself leads to a significant decline in the accuracy of named entity recognition when there is adversarial text in the text. In order to improve the robustness of named entity recognition under adversarial conditions, this paper proposes a Chinese named entity recognition model based on fusion graph embedding. Firstly, the model encodes and represents the phonetic and glyph information of the input text through graph learning and integrates above-multimodal knowledge into the model, thus enhancing the robustness of the model. Secondly, we use the Bi-LSTM to further obtain the context information of the text. Finally, conditional random field is used to decode and label entities. The experimental results on OntoNotes4.0, MSRA, Weibo, and Resume datasets show that the F1 values of this model increased by 3.76%, 3.93%, 4.16%, and 6.49%, respectively, in the presence of adversarial text, which verifies the effectiveness of this model.

Python Powered Intelligence And ML is designed to be your essential companion in your journey through the world of Artificial Intelligence and Python programming. We understand the importance of building a solid foundation in AI concepts, as well as mastering the tools and techniques needed to implement AI solutions effectively. What You’ll Find Inside: Foundation of Artificial Intelligence: In Chapter 1, we lay the groundwork for your AI education, providing a strong understanding of the fundamentals. Knowledge Presentation: Chapter 2 delves into how knowledge is represented in AI systems, a crucial element for creating intelligent machines. Informed / Heuristic Search Strategies: Chapter 3 explores strategies for problem-solving and decision-making, crucial in the AI domain. Natural Language Processing: In Chapter 4, we dive into the world of language understanding and processing, a key area of AI. Soft Computing: Chapter 5 introduces the concept of soft computing, which enables AI systems to work with uncertainty and imprecision. Neural Networks: Chapter 6 covers neural networks, a fundamental technology in modern AI, inspired by the human brain. Fuzzy Systems: Chapter 7 is all about fuzzy logic, which allows AI systems to deal with vagueness and uncertainty. History of Genetic Algorithms: Chapter 8 takes you through the fascinating history and principles of genetic algorithms. Regression: In Chapter 9, we explain regression techniques for predictive modeling, a crucial tool in AI. Python Programming: The second part of Python Powered Intelligence And ML focuses on Python, one of the most versatile and popular programming languages. Numpy: Chapter 11 introduces the powerful library for numerical computing in Python. Pandas: In Chapter 12, we explore Pandas, a tool for data manipulation and analysis. Matplotlib: Chapter 13 introduces you to data visualization using Matplotlib. Regression: We revisit regression in Chapter 14, providing more insights and applications. Our aim is to empower you with the knowledge and skills to excel in Artificial Intelligence and Python programming. Whether you are a beginner or an experienced programmer, this book is your go-to resource for AI and Python. We hope you enjoy this journey with us, and may it inspire you to explore the exciting world of Artificial Intelligence and Python programming. Happy learning!

Generative artificial intelligence systems remain vulnerable to sophisticated cyber threats and the emerging challenges posed by quantum computing. This study proposes and evaluates a new penetration testing suite to address quantum security concerns. The suite integrates dynamic and static application security testing (DAST and SAST) using OWASP ZAP, Burp Suite, SonarQube, and Fortify to detect and resolve vulnerabilities across application lifecycles. Real-time monitoring through interactive application security testing (IAST) with Contrast Assess near-real-time analysis facilitates pre-emptive remediation and remediation of insecure data handling and encryption flaws. Blockchain-enhanced logging, implemented via Hyperledger Fabric, provides tamper-proof and auditable records of all security activities. Furthermore, quantum-resistant cryptographic protocols, including lattice-based cryptography and RLWE, safeguard against quantum decryption threats, validated through simulated quantum attack scenarios. AI-driven red team simulations emulate adversarial and quantum-assisted attacks, uncovering vulnerabilities overlooked by traditional methods. Key results include the identification and remediation of over 300 vulnerabilities, a 70% reduction in high-severity issues within two weeks of testing, and a 90% resolution efficiency for blockchain-logged vulnerabilities. Quantum-resistant protocols exhibited strong resilience under adversarial conditions against simulated quantum attacks, achieving secure API encryption and data transmission. This research establishes a new protocol for securing generative AI systems, combining advanced tools, methodologies, and industry-tested methods.

This study presents a hybrid federated learning and explainable artificial intelligence framework, termed RobustFL, designed to mitigate adversarial attacks in medical imaging while incorporating privacy-preserving mechanisms. Focusing on chest X-ray analysis, the research systematically examined adversarial vulnerabilities and data privacy challenges in distributed healthcare systems. A simulation-based methodology was implemented using a 500-sample dataset partitioned across multiple clients in a non-independent and identically distributed setting. The framework extends federated averaging by integrating projected gradient descent-based adversarial training, differential privacy stochastic gradient descent, and SHAP-based interpretability. Experimental results demonstrate that RobustFL maintains stable performance under adversarial conditions, achieving 70% accuracy against FGSM and PGD attacks, with an adversarial success rate of approximately 30–31%. A balanced privacy budget of ε = 1.0 produced an F1-score of 0.412, indicating a trade-off between privacy and model utility. Interpretability consistency, measured via SHAP outputs, served as a proxy for trust assessment, while privacy risk indicators were used to infer potential user concern. The framework provides a practical pathway toward secure, transparent, and privacy-aware healthcare AI systems.

ObjectiveWhile the integration of modern AI systems in clinical practice is currently in the process of transforming how medicine is being practiced, the focus of most research activities lies on AI-associated efficacy and safety. However, the interplay between human agents and AI systems will equally shape the actual impact of such systems.MethodsThis study simulated human decision-making using 27 agents characterized by varying levels of competence, certainty, and trust. Agents completed binary and three-option decision tasks, both with and without AI assistance. AI models varied in competence (0.3–0.9) and, in some simulations, included confidence signals to influence human trust dynamically. Each scenario involved 10,000 simulated decisions per agent. In AI-assisted conditions, decisions were modulated by the agent's baseline trust and, in the conditional trust setting, the AI's expressed confidence.ResultsAI support significantly improved decision accuracy for most agents, especially those with high competence but low certainty. In binary tasks, agents showed up to 150% relative improvement in decision accuracy with AI competence ≥0.6. In three-option tasks, even lower-performing AI (e.g., 0.4 competence) enhanced decision results. Conditional trust simulations showed further gains, particularly among agents with moderate baseline trust, as dynamic trust adjustments based on AI confidence reduced over-reliance on poor AI recommendations.DiscussionResults demonstrate that AI assistance, particularly when paired with confidence calibration, enhances human decision-making, especially for uncertain or moderately skilled users. However, over-trusting low-competence AI can impair outcomes for high-performing agents. Tailored AI-human collaboration strategies are essential for optimizing clinical decision support.

Recent advancements in AI have led to the development of specialized systems related to multi-agent systems (MAS). However, the inherently collaborative nature of agents is often overlooked, and many of these specialized systems are used as components by other AI systems. From a software engineering perspective, this context can benefit from aligning the architectural characteristics of distributed systems with the inherently distributed nature of MAS. We propose that introducing a minimal set of agent-related concepts into the Distributed Systems (DS) domain can improve the engineering of modern MAS by leveraging techniques from DS engineering with established agent theory. In this study, we recapitulated the common origins of MAS and DS by drawing architectural parallels to establish a unified engineering approach. We then defined a minimal set of agent concepts to perform two practical studies on leveraging MAS development. First, we incorporated these concepts into a DS architectural pattern to design a distributed MAS. We then used these concepts in a graduate course to teach MAS engineering to students with no prior knowledge of agent theory. The learning outcomes from both courses included successful MAS implementation using DS tools and techniques. Although more than two-thirds of these students had no practical experience in developing distributed systems, the average final grade in both courses was above 80%, thus validating our approach. Finally, we discuss how this study supports the development of advanced systems using modern AI techniques consistently with established agent-related research while leveraging established DS techniques and concepts.

Existing CNN-based methods for semantic segmentation heavily depend on multi-scale features to meet the requirements of both semantic comprehension and detail preservation. State-of-the-art segmentation networks widely exploit conventional scale-transfer operations, i.e., up-sampling and down-sampling to learn multi-scale features. In this work, we find that these operations lead to scale-confused features and suboptimal performance because they are spatial-invariant and directly transit all feature information cross scales without spatial selection. To address this issue, we propose the Gated Scale-Transfer Operation (GSTO) to properly transit spatial-filtered features to another scale. Specifically, GSTO can work either with or without extra supervision. Unsupervised GSTO is learned from the feature itself while the supervised one is guided by the supervised probability matrix. Both forms of GSTO are lightweight and plug-and-play, which can be flexibly integrated into networks or modules for learning better multi-scale features. In particular, by plugging GSTO into HRNet, we get a more powerful backbone (namely GSTO-HRNet) for pixel labeling, and it achieves new state-of-the-art results on multiple benchmarks for semantic segmentation including Cityscapes, LIP and Pascal Context, with negligible extra computational cost. Moreover, experiment results demonstrate that GSTO can also significantly boost the performance of multi-scale feature aggregation modules like PPM and ASPP.

This paper proposes two robust control methods for swing-up and stabilization of a rotary inverted pendulum. A coupled sliding mode controller and an integral backstepping sliding mode controller are proposed here. These controllers provide aggressive swing-up and stabilization without involving a switching scheme between the swing-up mode and the stabilization mode. Robustness of both controllers are analyzed against matched and unmatched uncertainties. The coupled sliding mode control method is robust against matched uncertainties alone. Whereas the integral backstepping sliding mode control method ensures robustness against both matched and unmatched uncertainties with a continuous and chattering free control law. Simulations are carried out to compare the performance of the proposed controllers.

Auto-encoder (AE)-based condition monitoring (CM) methods for fault detection of wind turbines have received considerable attention due to their powerful feature extraction ability. However, traditional AE-based monitoring methods can only learn point-to-point features by minimizing reconstruction errors, which leads to a low sensitivity to anomaly data and weak robustness to noise data. To this end, we introduce a novel deep generative method based on the convolutional neural network (CNN)-conditional variational auto-encoder (CVAE). The key idea of CNN-CVAE is to unify the representation learning capacity of the CVAE and CNN. Specifically, CVAE can learn a probability distribution model by being trained on an anomaly-free supervisory control and data acquisition systems (SCADA) dataset; CNN and deconvolution operations are adopted for better time-series feature extraction and reconstruction performance. A statistical process control chart is applied to determine the alarm threshold. The effectiveness of the CNN-CVAE-based method is validated by datasets collected by SCADA installed in a commercial wind farm in China for impending blade breakage and gearbox failure. Abundant experiments with state-of-the-art deep learning-based CM methods are conducted, which indicate that our proposed method outperforms other methods in robustness, fault detection data sensitivity, fault warning time, and model parameters.

One unintended consequence of the Electronic Health Records (EHR) implementation is the overuse of content-importing technology, such as copy-and-paste, that creates “bloated” notes containing large amounts of textual redundancy. Despite the rising interest in applying machine learning models to learn from real-patient data, it is unclear how the phenomenon of note bloat might affect the Natural Language Processing (NLP) models derived from these notes. Therefore, in this work we examine the impact of redundancy on deep learning-based NLP models, considering four clinical prediction tasks using a publicly available EHR database. We applied two deduplication methods to the hospital notes, identifying large quantities of redundancy, and found that removing the redundancy usually has little negative impact on downstream performances, and can in certain circumstances assist models to achieve significantly better results. We also showed it is possible to attack model predictions by simply adding note duplicates, causing changes of correct predictions made by trained models into wrong predictions. In conclusion, we demonstrated that EHR text redundancy substantively affects NLP models for clinical prediction tasks, showing that the awareness of clinical contexts and robust modeling methods are important to create effective and reliable NLP systems in healthcare contexts.

e19297 Background: Obtaining clinical outcomes for analysis has historically been a critical barrier to cancer genomics research. EHRs could constitute an important data source to bridge this gap, but EHRs rarely capture structured outcomes such as cancer progression. Novel, robust methods are needed to capture clinically relevant outcomes from EHRs. Methods: Among patients with lung adenocarcinoma whose tumors were sequenced via the Dana Farber Cancer Institute/Brigham and Women’s PROFILE study from 2013-2018, imaging reports following first palliative-intent systemic therapy were annotated using natural language processing (NLP) models trained to capture cancer progression according to the structured “PRISSMM” framework. NLP-based cancer progression and imaging report frequency were jointly modeled using inverse-intensity weighted generalized estimated equations, censored at six months, to explore associations between alterations in lung cancer biomarkers (ALK, EGFR, ROS1, BRAF, KRAS, SMARCA4) and progression. Among patients with KRAS mutations who received immunotherapy, we also analyzed the association between STK11 mutations and progression. The novel outcome generated by the model – imaging report-based progression (iPROG) – corresponded to the difference in the mean log odds of progression per inverse-intensity weighted report associated with a given biomarker; it was reported as adjusted mean probability and in exponentiated form as an odds ratio (OR). Results: Among 690 patients with lung adenocarcinoma, associations between tumor mutations and the iPROG outcome are listed in the Table. Conclusions: A deep NLP model applied to EHR data can capture a novel cancer progression outcome, which is associated with known prognostic markers in lung cancer. Application of this method to large “real world” datasets, with attention to interactions between treatment and genomics, could speed biomarker discovery. [Table: see text]

Background Respiratory diseases, including chronic obstructive pulmonary disease (COPD), asthma, and pneumonia, are leading causes of global morbidity and mortality. Early and accurate diagnosis remains challenging, especially in resource-limited settings. Automated respiratory sound classification, supported by machine learning and deep learning, has emerged as a non-invasive and cost-effective diagnostic alternative. This systematic review examines the role of multistage classification frameworks and ensemble learning techniques in improving the accuracy, robustness, and clinical applicability of respiratory sound analysis. Methods A comprehensive literature search was conducted across major scientific databases, focusing on studies employing multistage, machine learning, or ensemble-based approaches for respiratory sound classification. Eligible studies were screened based on predefined inclusion criteria related to pulmonary acoustics, feature extraction, signal processing, and classification models. Extracted data were synthesized to compare methodologies, feature engineering strategies, ensemble algorithms, model architectures, benchmark datasets (ICBHI 2017, PhysioNet), and performance metrics such as accuracy, precision, recall, F1-score, and AUC-ROC. Results Findings indicate that multistage classification pipelines significantly enhance model reliability by integrating sequential preprocessing, feature extraction, dimensionality reduction, and classifier fusion. Ensemble learning methods—including bagging, boosting, random forests, stacking, and deep hybrid models—consistently outperform single-stage classifiers in handling noisy, non-stationary respiratory sounds and imbalanced datasets. Deep learning models such as convolutional neural networks and transformer-based architectures show superior performance when combined with advanced feature representations. However, challenges persist related to dataset variability, limited annotations, lack of standardization, and insufficient real-time or wearable-device compatibility. Explainable AI approaches are increasingly incorporated to support clinical interpretability. Conclusion Multistage and ensemble learning techniques demonstrate strong potential to improve automated respiratory sound classification and support early diagnosis of respiratory diseases. Despite promising results, advancements are needed in dataset standardization, multimodal integration, self-supervised learning, and IoT-enabled real-time monitoring. Future research should prioritize clinically interpretable, scalable, and regulation-ready AI systems to enable widespread deployment in respiratory healthcare.

In the past years, latent variable models have played an important role in various industrial AI systems, among which quality prediction is one of the most representative applications. Inspired by the idea of deep learning, those basic latent variable models have been extended to deep forms, based on which the quality prediction performance has been significantly improved. However, different latent variable models have their own strengths and weaknesses, a model works well under one scenario might not provide satisfactory performance under another. The motivation of this article is based on the viewpoint of information fusion and ensemble learning for heterogeneous latent variable models. Particularly, a collaborative deep learning and model fusion framework is formulated for the purpose of industrial quality prediction. In the first stage of the framework, collaborative layer-by-layer feature extractions are implemented among different latent variable models, through which different patterns of latent variables are identified in different layers of the deep model. Then, in the second stage, an ensemble regression modeling strategy is proposed to fuse the quality prediction results from different latent variable models, which is based on a well-designed data description method. Two real industrial examples are used for performance evaluation of the proposed method, based on which we can observe that information fusions in terms of both collaborative layer-by-layer feature extraction and heterogeneous model ensemble have positive effects in improving prediction accuracy and stability.