Archivio della ricerca - Fondazione Bruno Kessler
Not a member yet
21227 research outputs found
Sort by
Characterizing US Spatial Connectivity and Implications for Geographical Disease Dynamics and Metapopulation Modeling: Longitudinal Observational Study
Get PDFBackground: Human mobility is expected to be a critical factor in the geographic diffusion of infectious diseases, and this assumption led to the implementation of social distancing policies during the early fight against the COVID-19 emergency in the United States. Yet, because of substantial data gaps in the past, what still eludes our understanding are the following questions: (1) How does mobility contribute to the spread of infection within the United States at local, regional, and national scales? (2) How do seasonality and shifts in behavior affect mobility over time? (3) At what geographic level is mobility homogeneous across the United States? Objective: This study aimed to address the questions that are critical for developing accurate transmission models, predicting the spatial propagation of disease across scales, and understanding the optimal geographical and temporal scale for the implementation of control policies. Methods: We analyzed high-resolution mobility data from mobile app usage from SafeGraph Inc, mapping daily connectivity between the US counties to grasp spatial clustering and temporal stability. Integrating this into a spatially explicit transmission model, we replicated SARS-CoV-2's first wave invasion, assessing mobility's spatiotemporal impact on disease predictions. Results: Analysis from 2019 to 2021 showed that mobility patterns remained stable, except for a decline in April 2020 due to lockdowns, which reduced daily movements from 45 million to approximately 25 million nationwide. Despite this reduction, intercounty connectivity remained seasonally stable, largely unaffected during the early COVID-19 phase, with a median Spearman coefficient of 0.62 (SD 0.01) between daily connectivity and gravity networks., We identified 104 geographic clusters of US counties with strong internal mobility connectivity and weaker links to counties outside these clusters. These clusters were stable over time, largely overlapping state boundaries (normalized mutual information=0.82) and demonstrating high temporal stability (normalized mutual information=0.95). Our findings suggest that intercounty connectivity is relatively static and homogeneous at the substate level. Furthermore, while county-level, daily mobility data best captures disease invasion, static mobility data aggregated to the cluster level also effectively models spatial diffusion. Conclusions: Our work demonstrates that intercounty mobility was negligibly affected outside the lockdown period in April 2020, explaining the broad spatial distribution of COVID-19 outbreaks in the United States during the early phase of the pandemic. Such geographically dispersed outbreaks place a significant strain on national public health resources and necessitate complex metapopulation modeling approaches for predicting disease dynamics and control design. We thus inform the design of such metapopulation models to balance high disease predictability with low data requirements
Scaling and Enhancing LLM-based AVSR: A Sparse Mixture of Projectors Approach
Get PDFAudio-Visual Speech Recognition (AVSR) enhances robustness in noisy environments by integrating visual cues. While recent advances integrate Large Language Models (LLMs) into AVSR, their high computational cost hinders deployment in resourceconstrained settings. To address this, we propose Llama-SMoP, an efficient Multimodal LLM that employs a Sparse Mixture of Projectors (SMoP) module to scale model capacity without increasing inference costs. By incorporating sparsely-gated mixture-of-experts (MoE) projectors, Llama-SMoP enables the use of smaller LLMs while maintaining strong performance. We explore three SMoP configurations and show that LlamaSMoP DEDR (Disjoint-Experts, Disjoint-Routers), which uses modality-specific routers and experts, achieves superior performance on ASR, VSR, and AVSR tasks. Ablation studies conf irm its effectiveness in expert activation, scalability, and noise robustness
"Sì, la montagna può salvarci, ma non magicamente". Intervista a Enrico Camanni
No full textQuesta intervista con Enrico Camanni, alpinista e saggista, ruota attorno al significato delle montagne nella vita moderna. Tra i temi discussi: i significati personali e culturali delle montagne e il loro fascino sui generis; il legame tra amore per la montagna e ribellione giovanile; l’origine storica dell’alpinismo; incanto e disincanto nelle terre alte; la passione per la verticalità e il superomismo; la sacralizzazione e profanazione delle vette; il consumismo e l’arte della resa
Progressive Self-Optimization Network: An unsupervised change detection method for VHR optical remote sensing imagery
Get PDFChange detection in very-high-resolution (VHR) remote sensing imagery has consistently been a focus and challenge within the remote sensing community. We present a novel unsupervised method named Progressive Self-Optimization Network. In this method, a new sample strategy is developed based on the initial change detection across three feature domains: spectral, deep, and class signal, to capture the “weak-to-strong” change signals and collect training samples with high accuracy. A new lightweight convolutional neural network is designed by partially replacing traditional convolutional layers with weight-shared partial convolution. To detect changes, a progressive self-optimization pattern is proposed based on the “weak-to-strong” change signals. This pattern detects changes gradually from regions with weak and strong change signals to regions with moderate change signals in a progressive manner. During this progressive process, detection results from each progression are integrated with “weak-to-strong” change signals to reselect training samples for the transfer training in the following progression, thus attempting to optimize the lightweight network. The final change map is generated by fusing all progression results. Five open VHR datasets and fourteen state-of-the-art unsupervised methods validate the proposed method
A Trustworthy Evolutionary Fuzzy Neural Network Framework for Maternal Health Risk Classification
No full textIn medical applications, the demand for explainable AI systems has driven the adoption of symbolic methods, such as Fuzzy Inference Systems, known for their interpretable fuzzy rules and ability to facilitate communication between AI systems and human users. Recently, hybrid approaches that combine sub-symbolic (data-driven) and symbolic methods have gained significant attention due to their capacity to leverage the strengths of both paradigms. For example, Fuzzy Neural Networks integrate the predictive power of neural networks with the interpretability of fuzzy systems. At the same time, Evolutionary Algorithms further improve these models by optimizing parameters, improving performance, and increasing adaptability.
This paper proposes a novel evolutionary fuzzy neural network framework incorporating a genetic algorithm to enhance classification capabilities while preserving model transparency. Our architecture integrates evolutionary optimization into the parameter update process of an existing Fuzzy Neural Network. An extensive validation on the Maternal Health Risk dataset demonstrates the framework’s effectiveness in balancing predictive accuracy and explainability. GitHub: https://github.com/IDA-FBK/NeuroFuzzyProjec
Latent geometry emerging from network-driven processes
No full textUnderstanding network functionality requires integrating structure and dynamics, and emergent latent geometry induced by network-driven processes captures the low-dimensional spaces governing this interplay. In this Perspective, we review generative-model-based approaches, distinguishing two reconstruction classes: fixed-time methods, which infer geometry at specific temporal scales (e.g., equilibrium), and multi-resolution methods, which integrate dynamics across near- and far-from-equilibrium states. Over the past decade, these models have revealed functional organization in biological, social, and technological networks. Hence, we provide a unified overview of these methods, with particular attention to the underlying mathematical constructions. Further, we point to promising extensions of these frameworks, which combine the previously developed methods to other well-established analytical frameworks
A toolbox for volleyball data analytics: a case study on the italian women’s league
Get PDFData analytics plays a central role in volleyball, offering new ways to enhance team performance and inform match strategies. This paper presents an open-source Python-based toolbox that extends the PyDataVolley library, enabling advanced processing, visualization, and analysis of scouting data. The toolbox includes Machine Learning Clustering algorithms, Multi-Criteria Decision Analysis approaches, and Markov Chain models, and is validated using datasets from the 2023–2024 and 2024–2025 seasons of the Italian Women’s Serie A2 Championship. A case study on the 2025 Italian Cup final between Consolini Volley and Trentino Volley highlights the practical impact of the toolbox. Throughout the 2024–2025 season, Consolini Volley consistently showed superior performance metrics compared to Trentino Volley. However, in the final match, Trentino Volley implemented a targeted tactical strategy informed by the insights from the toolbox, which effectively challenged Consolini’s gameplay. As a result, Consolini Volley exhibited a general drop across all main performance indicators during the final
Uni-Mate: A Retrieval-Augmented Generation System to Provide High School Students with Accurate Academic Guidance
Get PDFThis paper introduces the development and evaluation of a Retrieval-Augmented Generation (RAG) system designed to assist prospective students in navigating university options. The system provides accurate academic guidance by retrieving and synthesizing information on undergraduate and single-cycle master’s degree programs, as well as library resources, from the University of Trento and the University of Verona. The RAG pipeline utilizes a streamlined toolchain, incorporating a Markdown parser for efficient data handling and the Llama3-8b-8192 Large Language Model (LLM) for query processing. The system’s performance was assessed through both automated evaluation, using the Llama3-70b LLM as a reference, and blinded human evaluation. The results demonstrate the system’s potential for providing relevant and accurate information to students. The evaluation also highlighted areas for further development, including enhanced retrieval mechanisms and expanded LLM testing. Future work aims to broaden the system’s scope to include more degree levels and universities, ultimately creating a comprehensive platform to support students in their academic decision-making journey
A frequency selective surface (FSS) based on a reconfigurable MEMS switch for GNSS L1-band
No full textThis paper presents a tunable Frequency Selective Surface (FSS) for the L1-band of navigation frequencies that utilises a MEMS switch. The reconfigurable frequency-selective electromagnetic filter, achieved by combining hard magnetic materials with microelectromechanical systems (MEMS), provides a novel approach to reconfigurable frequency-selective surfaces (FSS). By incorporating magnetically actuated dipole components that can tilt away from the base surface, we can adjust the operating frequency of the FSS without physically modifying the size of the dipole components. The 9 × 9 array, measuring 365 mm, consists of plates made from Rogers RO3003 material, each with dimensions of 531.2 × 531.2 × 8.768 mm, layered with a 0.03 mm-thick copper conductor (Cu). The proposed system features a cross dipole printed on a Rogers-RO3003 substrate, with a MEMS switch placed between one of the dipole arms to adjust its length. The MEMS switch facilitates frequency tuning by altering the length of a rectangular dipole. This phase modulation technique enables the steering of reflected waves, thereby enhancing beam resolution and coverage, while allowing the intelligent reflecting surface (IRS) to control the reflection of reflection. The presented reconfigurable FSS design has effectively demonstrated the ability to tune its resonant frequency for the L1-band without physically changing its dimensions. The design was assessed using the commercial simulation software CST, and the numerical results corroborate the findings, thereby confirming its effectiveness
System-level simulation-based verification of Autonomous Driving Systems with the VIVAS framework and CARLA simulator
No full textEnsuring the safety and reliability of increasingly complex Autonomous Driving Systems (ADS) poses significant challenges, particularly when these systems rely on AI components for perception and control. In the ESA-funded project VIVAS, we developed a comprehensive framework for system-level, simulation-based Verification and Validation (V&V) of autonomous systems. This framework integrates a simulation model of the system, an abstract model describing system behavior symbolically, and formal methods for scenario generation and verification of simulation executions. The automated scenario generation process is guided by diverse coverage criteria.
In this paper, we present the application of the VIVAS framework to ADS by integrating it with CARLA, a widely-used driving simulator, and its ScenarioRunner tool. This integration facilitates the creation of diverse and complex driving scenarios to validate different state-of-the-art AI-based ADS agents shared by the CARLA community through its Autonomous Driving Challenge. We detail the development of a symbolic ADS model and the formulation of a coverage criterion focused on the behaviors of vehicles surrounding the ADS. Using the VIVAS framework, we generate and execute various highway-driving scenarios, evaluating the capabilities of the AI components. The results demonstrate the effectiveness of VIVAS in automating scenario generation for different off-the-shelf AI solutions