Institution profile

University of Catania

Academic institutioneurope · it
Official website
Research library163linked papers
Opportunities0open roles
Selected work

Representative Papers

A Survey on Design Methodologies for Accelerating Deep Learning on Heterogeneous Architectures

Nov 29, 2023arXiv.org

Designing deep learning accelerators for heterogeneous HPC and edge platforms faces key challenges including insufficient parallelism exploitation and excessive data movement overhead. This paper systematically surveys accelerator design methodologies, covering hardware-software co-design, high-level synthesis, domain-specific compilers (e.g., TVM, Halide), design space exploration, and cycle-accurate modeling and simulation. We propose, for the first time, a unified multi-dimensional classification framework that distills two fundamental principles: “minimizing data movement” and “maximizing parallelism.” The survey bridges the gap between architectural overviews and implementation-oriented methodologies, explicitly identifying emerging directions such as approximate computing integrated with reconfigurability. Our work provides both a methodological foundation and practical guidance for developing efficient, scalable AI accelerators—enabling principled design decisions across diverse heterogeneous computing ecosystems.

4 citationsRead paper

Target-aware Bayesian inference via generalized thermodynamic integration

Apr 24, 2023Computational statistics (Zeitschrift)

This paper addresses the high bias and poor robustness of Bayesian posterior expectations and marginal likelihood (Bayesian evidence) estimation under high-dimensional, non-isotropic target distributions. To this end, we propose a target-aware Generalized Thermodynamic Integration (GTI) framework. Our core innovation is a learnable path parameterization that adaptively couples the integration path to the geometric structure of the target distribution, jointly optimizing the path via variational inference and refining sample efficiency through adaptive importance sampling. The GTI framework significantly reduces marginal likelihood estimation bias—achieving an average 38% reduction across multiple models—while improving posterior predictive consistency and computational stability. By explicitly encoding target-distribution geometry into the thermodynamic integration process, GTI establishes a more accurate and robust paradigm for high-dimensional Bayesian inference.

3 citationsRead paper

Experimental Evaluation of a UAV-Mounted LEO Satellite Backhaul for Emergency Connectivity

Jan 07, 2026arXiv.org

This study addresses the critical need for rapid deployment and reliable connectivity in disaster scenarios where terrestrial communication infrastructure is compromised. The authors propose an airborne base station architecture leveraging a rotary-wing unmanned aerial vehicle (UAV) equipped with a commercial Starlink Mini terminal to provide Wi-Fi access in cellular-deprived areas via low Earth orbit (LEO) satellite backhaul. To the best of the authors’ knowledge, this work presents the first real-world flight validation of integrating off-the-shelf LEO satellite terminals with UAVs for emergency communications. System performance is evaluated through a combination of ns-3 simulations and field flight experiments, demonstrating stable uplink throughput of approximately 30 Mbps within a 200-meter coverage radius, with negligible impact on UAV battery life. The results confirm a favorable balance between network performance and energy efficiency under dynamic flight conditions, underscoring the practical viability of the proposed solution in emergency response scenarios.

1 citationsRead paper

TI-PREGO: Chain of Thought and In-Context Learning for Online Mistake Detection in PRocedural EGOcentric Videos

Nov 04, 2024arXiv.org

Real-time online detection of open-set errors (i.e., unknown or novel errors) in first-person procedural videos remains challenging due to their unstructured nature and lack of prior error annotations. Method: We propose a dual-branch online architecture: an action recognition branch performs frame-level, streaming action parsing; an LLM-driven prediction branch conducts chain-of-thought reasoning over action sequences and forecasts subsequent steps, localizing errors via inconsistency between recognition and prediction. The method integrates video action recognition, action token aggregation, and in-context learning from large language models, enabling millisecond-scale inference and dynamic streaming input. Results: Evaluated on two procedural video benchmarks, our approach significantly outperforms state-of-the-art methods in open-set error detection, ultra-low latency (<50 ms), and cross-task generalization. It demonstrates strong robustness and effectiveness without requiring predefined error samples—constituting the first online error detection framework for high-reliability domains such as manufacturing and healthcare.

1 citationsRead paper
Recent publications

Latest Papers