Skip to main content

Resultados relacionados

Resultados totales (Incluyendo duplicados): 18
Encontrada(s) 1 página(s)

UPCommons. Portal del coneixement obert de la UPC
Publication Contribución de congreso (ConferenceOutput)

Mind the memory gap: unveiling GPU bottlenecks in large-batch LLM inference

  • Garcia Recasens, Pol|||0009-0008-5535-0012
  • Agulló López, Ferran|||0000-0002-7276-2472
  • Zhu, Yue
  • Wang, Changyang
  • Lee, Eun Kyung
  • Tardieu, Olivier
  • Torres Tomás, Juan José
  • Berral García, Josep Lluís|||0000-0003-3037-3580
Large language models have been widely adopted across different tasks, but their auto-regressive generation nature often leads to inefficient resource utilization during inference. While batching is commonly used to increase throughput, per-formance gains plateau beyond a certain batch size, especially with smaller models, a phenomenon that existing literature typically explains as a shift to the compute-bound regime. In this paper, through an in-depth GPU-level analysis, we reveal that large-batch inference remains memory-bound, with most GPU compute capabilities underutilized due to DRAM bandwidth saturation as the primary bottleneck. To address this, we propose a Batching Configuration Advisor (BCA) that optimizes memory allocation, reducing GPU memory requirements with minimal impact on throughput. The freed memory and underutilized GPU compute capabilities can then be leveraged by concurrent workloads. Specifically, we use model replication to improve serving throughput and GPU utilization. Our findings challenge conventional assumptions about LLM inference, offering new in-sights and practical strategies for improving resource utilization, particularly for smaller language models., This work has been partially financed by grant agreement EU-HORIZON GA.101095717 and by the EU-HORIZON MSCA programme under grant agreement EU-HORIZON MSCA GA.101086248. Also, it has been partially financed by Generalitat de Catalunya (AGAUR) under grant agreement 2021-SGR-00478, by Severo Ochoa Center of Excellence CEX-2021-001148-S-20-3, and by the Spanish Ministry of Science (MICINN), the Research State Agency (AEI) and European Regional Development Funds (ERDF/FEDER) under grant agreement PID2021-126248OBI00, MCIN/AEI/10.13039/ 501100011033/ FEDER, UE., Peer Reviewed
UPCommons. Portal del coneixement obert de la UPC
Publication Artículo científico (JournalArticle)

Rethinking trust in synthetic health data: lessons from 7 European research initiatives

  • Declerck, Jens
  • Kalra, Dipak
  • Airola, Antti
  • Ali Amer, Ahmed Youssef
  • Chatzichristos, Christos
  • Mañu Pereira, Maria del Mar
  • de Brito Robalo, Bruno M.
  • Ghini, Francesco
  • Gutiérrez Torre, Alberto|||0000-0002-5548-3359
  • Reidel Lojo, Sara Isabel|||0000-0003-1887-9538
Synthetic data generation (SDG) structured health data is increasingly promoted as a solution to longstanding barriers in health data access. It is offering the promise of privacy-preserving data reuse for research, innovation, and policy. Despite rapid technical advances, the adoption of synthetic health data in real-world settings remains limited. Shaped by challenges around data quality, representativeness, infrastructure readiness, trust, and legal uncertainty, this viewpoint draws on experiences from 7 European research initiatives within the HealthData4EU cluster to reflect on how SDG is being operationalized in practice. It synthesizes cross-project insights to highlight recurring methodological and governance tensions and to examine their implications for trust and responsible use. The analysis argues that trustworthy SDG cannot be achieved through technical optimization alone but requires alignment between evaluation practices, upstream data stewardship, regulatory clarity, and sustained stakeholder engagement. Addressing these conditions is essential for moving synthetic data from experimental pilots toward a credible and sustainable component of European health research ecosystems., This paper was funded by the Innovative Health Initiative through the SYNTHIA project. The study draws on insights from 7 research projects that form part of the HealthData4EU cluster. Projects within the cluster and their funding sources are as follows. SYNTHEMA, PHASE IV AI, SECURED, FLUTE, and AISym4MED are funded by the Horizon Europe program. PHEMS is funded by Horizon Europe and UK Research and Innovation. SYNTHIA is funded through the Innovative Health Initiative., Peer Reviewed, Article signat per 18 autors/es: Jens Declerck 1,2, MSc; Dipak Kalra 1, Prof Dr; Antti Airola 3; Ahmed Youssef Ali Amer 4; Christos Chatzichristos 5; Maria del Mar Mañu Pereira 6; Bruno M. de Brito Robalo 7; Francesco Ghini 8; Alberto Gutierrez-Torre 9; Sem Hoogteijling 10; Susanne Hultsch 11; Jan Ramon 12; Sara Reidel 13,14; Francesco Regazzoni 15,16; Luís Silva 17; Inês Silveira 17; Tsekeridou Sofia 18, PhD; Christophe Maes 1,2 / 1- The European Institute for Innovation through Health Data, Oosterzele, Belgium; 2- Unit of Medical Informatics and Statistics, Department of Public Health and Primary Care, The University of Ghent, Ghent, Belgium; 3- Department of Computing, University of Turku, Turku, Finland; 4- Innovative Medicine, Johnson & Johnson, Beerse, Belgium; 5- Department of Electrical Engineering, KU Leuven, Leuven, Belgium; 6- University Hospital Vall d'Hebron, Barcelona, Spain; 7- Department of Obstetrics and Gynaecology, Erasmus MC, University Medical Center, Rotterdam, The Netherlands; 8- Laboratory of Translational Research, Reggio Emilia, Italy; 9- Barcelona Supercomputing Center, Barcelona, Catalonia, Spain; 10- Department of Neurology and Neurosurgery, University Medical Center Utrecht Brain Center, Utrecht, The Netherlands; 11- VEIL.AI, Helsinki, Finland; 12- Institut national de recherche en sciences et technologies du numérique, Lille, France; 13- Vall d'Hebron Research Institute, University Hospital Vall d'Hebron, Barcelona, Spain; 14- BarcelonaTech (UPC), Universitat Politècnica de Catalunya, Barcelona, Spain; 15- University of Amsterdam, Amsterdam, The Netherlands; 16- Università della Svizzera italiana, Lugano, Ticino, Switzerland; 17- Laboratory for Instrumentation, Biomedical Engineering and Radiation Physics, NOVA School of Science and Technology, Caparica, Portugal; 18- Netcompany BELUX, Athens, Greece
UPCommons. Portal del coneixement obert de la UPC
Publication Artículo científico (JournalArticle)

Towards generalizable Federated Learning in medical imaging: a real-world case study on mammography data

  • Tzortzis, Ioannis N.
  • Gutiérrez Torre, Alberto|||0000-0002-5548-3359
  • Sykiotis, Stavros
  • Agulló López, Ferran|||0000-0002-7276-2472
  • Bakalos, Nikolaos
  • Doulamis, Anastasios
  • Doulamis, Nikolaos
  • Berral García, Josep Lluís|||0000-0003-3037-3580
Federated Learning has been rapidly gaining in popularity in medical applications, due to the increased privacy offered, since medical data doesn't need to leave the hospitals' premises for AI model training. However, a direct translation of a classic experiment to a federated one is not always straightforward. In this work, we delve into the intricacies of federated learning for a breast cancer classification tool. We compare classic model training with a federated variant, and highlight the adaptations that need to be taken care of to ensure the equivalence between the two. Specifically, we introduce the Breast Area Detection tool as an essential component of the pre-processing pipeline to enhance the robustness of Federated Learning by offering data harmonization. On top of that, we present an end-to-end Federated Learning framework that is effective for real-world data and scenarios. Among the three real-world hospitals involved in the experimental procedure, the proposed framework significantly improves performance at the first hospital, providing consistent results similar to those achieved in the classic approach. Experimental results demonstrate that the interventions introduced improved model performance by approximately 35%, aligning federated learning and centralized model performance., This paper is funded by the EU funded projects SECURED “Scaling Up Secure Processing, Anonymization And Generation Of Health Data For EU Cross Border Collaborative Research And Innovation” with grant agreement no. 101095717 and INCISIVE “A multimodal AI-based toolbox and an interoperable health imaging repository for the empowerment of imaging analysis related to the diagnosis, prediction and follow-up of cancer” with grant agreement no. 952179. This paper has been partially financed by the Spanish Ministry of Science (MICINN), the Research State Agency (AEI) and European Regional Development Funds (ERDF/FEDER) under grant agreement PID2021-126248OB-I00, Severo Ochoa Center of Excellence CEX-2021-001148-S-20-3, 2021- SGR-00478, UE, also Generalitat de Catalunya (AGAUR) under grant agreement 2021-SGR-00478., Peer Reviewed
UPCommons. Portal del coneixement obert de la UPC
Publication Contribución de congreso (ConferenceOutput)

Beyond parameter averaging in model aggregation

  • Garcia Recasens, Pol|||0009-0008-5535-0012
  • Torres Viñals, Jordi|||0000-0003-1963-7418
  • Berral García, Josep Lluís|||0000-0003-3037-3580
  • Hauberg, Søren
  • Moreno Muñoz, Pablo
The success of foundation models is strongly linked to scale, which has reinforced the interest in federated learning. With the prohibitive cost of training a large language model (LLM) in mind, little attention has been placed on reusing pre-trained models in collaborative training settings. Self-supervision has also played an important role in this success, but its emphasis has been primarily on data. This paper leverages Bayesian principles to bring self-supervision into the model aggregation toolbox. It introduces self-supervised Fisher merging, a framework that successfully merges models in parameter space without re-visiting data, opening a new door in model reusability. Experimental results build the foundation of our method on tractable linear models, and highlight its potential on aggregating neural networks., This work was supported by a research grant (42062) from VILLUM FONDEN. This project has also received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement 757360). Our work was funded in part by the Novo Nordisk Foundation through the Center for Basic Machine Learning Research in Life Sciences (MLSS) with reference number NNF20OC0062606. This work has been partially financed by the EU-H2020 programme under grant agreement EU-H2020 GA.952179 and grant agreement EU-HORIZON GA.101095717. Also, it has been partially financed by Generalitat de Catalunya (AGAUR) under grant agreement 2021-SGR-00478., Peer Reviewed
UPCommons. Portal del coneixement obert de la UPC
Publication Contribución de congreso (ConferenceOutput)

SECURED for health: scaling up privacy to enable the integration of the european health data space

  • Regazzoni, Francesco
  • Acs, Gergely
  • Aszalos, Albert Zoltan
  • Avgerinos, Christos
  • Bakalos, Nikolaos
  • Berral García, Josep Lluís|||0000-0003-3037-3580
  • Bos, Joppe W.
  • Brohet, Marco
  • Castillo Sanz, Andrés G.
  • Gutiérrez Torre, Alberto|||0000-0002-5548-3359
In this paper, we present the SECURED project funded in part by the European Union (EU), Grant Agreement no. 10109571. Views and opinions expressed are those of the authors and do not necessarily reflect those of the EU or the Health and Digital Executive Agency. Neither the EU nor the granting authority are responsible for them., aimed at improving privacy-preserving processing of data in the health domain. The technologies developed in the project will be demonstrated in four health-related use cases and with the involvement of SME's selected through an open funding call., Funded in part by the European Union (EU), Grant Agreement no. 10109571., Peer Reviewed, Article signat per 41 autors/es: Francesco Regazzoni ∗ xviii , Gergely Acs xii , Albert Zoltan Aszalos x, Christos Avgerinos xv , Nikolaos Bakalos xvi , Josep Ll. Berral ††, Joppe W. Bos §, Marco Brohet *, Andres G. Castillo Sanz ‡, Gareth T. Davies §, Stefanos Florescu ∥, Pierre-Elisee Flory ¶, Alberto Gutierrez-Torre ††, Evangelos Haleplidis xi , Alice Heliou ¶, Sotirios Ioannidis †, Alexander Islam El-Kadyxi, 14 , Katarzyna Kapusta ¶, Konstantina Karagianni xi , Pieter Kruizinga ∥, Kyrian Maat ∗, Zoltan Adam Mann ∗, Kalliopi Mastoraki †, SeoJeong Moon §, Maja Nisevic ∗∗, Balazs Pejo xii , Kostas Papagiannopoulos ∗, Vassilis Paliuras xi xix , Paolo Palmieri xiii , Francesca Palumbo ‡‡, Juan Carlos Perez Baun xvii, Peter Pollner x, Eduard Porta-Pardo xiv , Luca Pulina ‡‡, Muhammad Ali Siddiqi ∥, Daniela Spajic ∗∗, Christos Strydis ∥, Georgios Tasopoulos ∗, Vincent Thouvenot ¶, Christos Tselios xi , Apostolos P. Fournaris xi / ∗ University of Amsterdam (NL), † Circular Economy Foundation (BE), ‡ FHUNJ (ES), § NXP (BE), ¶ Thales (FR), ∥ Erasmus Medical Center (NL), ∗∗ KU Leuven (BE), †† Barcelona Supercomputing Center (ES), ‡‡ University of Sassari (IT), x Semmelweis University (HU), xi Industrial Systems Institute/R.C. ATHENA (GR), xii BME (HU), xiii University College Cork (IE), xv Catalink (CY), xvi ICCS (GR), xvii EVIDEN (ES), xiv Josep Carreras Leukaemia Research Institute (ES), xviii Universita della Svizzera italiana (CH), xix University of Patras (GR)
UPCommons. Portal del coneixement obert de la UPC
Publication Artículo científico (JournalArticle)

FRIDA: Free-rider detection using privacy attacks

  • Garcia Recasens, Pol|||0009-0008-5535-0012
  • Horváth, Ádám
  • Gutiérrez Torre, Alberto|||0000-0002-5548-3359
  • Torres Viñals, Jordi|||0000-0003-1963-7418
  • Berral García, Josep Lluís|||0000-0003-3037-3580
  • Pejó, Balázs|||0000-0002-1825-9251
Federated learning is increasingly popular as it enables multiple parties with limited datasets and resources to train a machine learning model collaboratively. However, similar to other collaborative systems, federated learning is vulnerable to free-riders — participants who benefit from the global model without contributing. Free-riders compromise the integrity of the learning process and slow down the convergence of the global model, resulting in increased costs for honest participants. To address this challenge, we propose FRIDA: free-rider detection using privacy attacks. Instead of focusing on implicit effects of free-riding, FRIDA utilizes membership and property inference attacks to directly infer evidence of genuine client training. Our extensive evaluation demonstrates that FRIDA is effective across a wide range of scenarios., Project no. 145832, implemented with the support provided by the Ministry of Innovation and Technology from the NRDI Fund, financed under the PD_23 funding scheme. Also, partially financed by the Spanish Ministry of Science (MICINN), the Research State Agency (AEI), and European Regional Development Funds (ERDF/FEDER) under grant agreement PID2021-126248OB-I00, MCIN/AEI/10.13039/501100011033/FEDER, UE, and Generalitat de Catalunya (AGAUR) under grant 2021-SGR-00478. Lastly, also funded by the European Union (Grant Agreement Nr. 101095717, SECURED Project)., Peer Reviewed
UPCommons. Portal del coneixement obert de la UPC
Publication Artículo científico (JournalArticle)

SynthVal: A framework for validating synthetic medical images

  • Guidotti, Dario
  • Pandolfo, Laura
  • Gutiérrez Torre, Alberto|||0000-0002-5548-3359
  • López Rúbio, Omar
  • Pulina, Luca
Synthetic data is increasingly used in medical imaging to overcome data scarcity and privacy constraints. However, assessing the fidelity of synthetic images remains a critical challenge for ensuring their safe and effective use in clinical and AI applications. We present SynthVal, a Python-based framework for validating the quality of synthetic medical images through statistical comparisons in deep feature space. SynthVal extracts semantic image embeddings using transformer-based models and computes similarity metrics – including Fréchet Distance, Wasserstein Distance, and Kullback-Leibler Divergence – between real and synthetic data distributions. The framework is designed for modularity, scalability, and ease of integration into existing workflows via pip installation. We evaluate SynthVal using real images from the CSAW-CC dataset, and synthetic images produced by a model developed by the Barcelona Supercomputing Center. Our experiments systematically benchmark the influence of different feature extraction models and similarity metrics, providing practical insights for selecting validation strategies in medical image synthesis. SynthVal offers a reproducible and extensible solution for quality control in synthetic data pipelines., The work of Dario Guidotti, Laura Pandolfo, and Luca Pulina was supported by EU Project SECURED, ‘‘Scaling Up Secure Processing, Anonymization and Generation of Health Data for EU Cross-Border Collaborative Research and Innovation,’’ under Agreement 101095717. The work of Alberto Gutierrez-Torre and Omar Lopez-Rubio was supported by SECURED under Agreement 101095717; in part by Spanish Ministry of Science (MICINN), the State Research Agency (AEI), and European Regional Development Funds (ERDF/FEDER) under Grant PID2021-126248OB-I00; in part by MCIN/AEI/10.13039/501100011033/FEDER, UE; and in part by the Generalitat de Catalunya (AGAUR) under Grant 2021-SGR-00478., Peer Reviewed
UPCommons. Portal del coneixement obert de la UPC
Publication Contribución de congreso (ConferenceOutput)

HealthMesh: An architectural framework for federated healthcare data management

  • Bisquert Parés, Aniol
  • Hmimou Ham Man, Achraf|||0009-0004-0448-3406
  • Berral García, Josep Lluís|||0000-0003-3037-3580
  • Gutiérrez Torre, Alberto|||0000-0002-5548-3359
  • Romero Moral, Óscar|||0000-0001-6350-8328
Recently, significant milestones have been achieved in the field of healthcare data analysis. However, alongside these accomplishments, substantial data-related challenges have emerged in the domain of big data management. Modern healthcare projects are no more dealing with a single data repository but many heterogeneous ones and must overcome data variety, privacy and governance issues. Yet, current solutions face a privacy-decentralization trade-off. To address this dual challenge, we introduce HealthMesh, a novel layered architectural framework based on the Data Mesh principles, providing a domaindecentralised paradigm. In addition, the framework incorporates a Semantic Data Model which establishes robust governance, enables interoperability and guarantees policy compliance for all the data assets. To demonstrate the capabilities of the proposed approach, we provide an illustrative example inspired by the use case of the INCISIVE project for breast cancer analytics. Overall, this work makes a significant contribution on collecting key challenges, identifying actors and providing a set of components and guidelines for establishing a holistic framework for the complex field of healthcare data management., The work reported here was supported by the EU-H2020 programme under GA.952179 INCISIVE, Horizon Europe Programme under GA.101095717 (SECURED) and GA.101135513 (CYCLOPS), the Spanish Research State Agency (MICINN/AEI, ERDF/FEDER) under GA.MCIN/AEI/10.13039/501100011033/FEDER (UE DALEST), the Spanish Ministerio de Ciencia e Innovación under project PID2020-117191RB-I00 / AEI/10.13039/501100011033 (DOGO4ML), and the Generalitat de Catalunya (AGAUR) 2021-SGR-00478 (CROMAI)., Peer Reviewed
UPCommons. Portal del coneixement obert de la UPC
Publication Contribución de congreso (ConferenceOutput)

Towards Pareto optimal throughput in small language model serving

  • Garcia Recasens, Pol|||0009-0008-5535-0012
  • Zhu, Yue
  • Wang, Chen
  • Lee, Eun Kyung
  • Tardieu, Olivier
  • Youssef, Alaa
  • Torres Viñals, Jordi|||0000-0003-1963-7418
  • Berral García, Josep Lluís|||0000-0003-3037-3580
Large language models (LLMs) have revolutionized the state-of-the-art of many different natural language processing tasks. Although serving LLMs is computationally and memory demanding, the rise of Small Language Models (SLMs) offers new opportunities for resource-constrained users, who now are able to serve small models with cutting-edge performance. In this paper, we present a set of experiments designed to benchmark SLM inference at performance and energy levels. Our analysis provides a new perspective in serving, highlighting that the small memory footprint of SLMs allows for reaching the Pareto-optimal throughput within the resource capacity of a single accelerator. In this regard, we present an initial set of findings demonstrating how model replication can effectively improve resource utilization for serving SLMs., This work has been partially financed by grant agreement EUHORIZON GA.101095717 and by the EU-HORIZON MSCA programme under grant agreement EU-HORIZON MSCA GA.101086248. Also, it has been partially financed by Generalitat de Catalunya (AGAUR) under grant agreement 2021- SGR-00478, and by the Spanish Ministry of Science (MICINN), the Research State Agency (AEI) and European Regional Development Funds (ERDF/FEDER) under grant agreement PID2021-126248OB-I00, MCIN/AEI/ 10.13039/ 501100011033/ FEDER, UE., Peer Reviewed
Recercat. Dipósit de la Recerca de Catalunya
Publication Artículo científico (JournalArticle)

Rethinking trust in synthetic health data: lessons from 7 European research initiatives

  • Declerck, Jens
  • Kalra, Dipak
  • Airola, Antti
  • Ali Amer, Ahmed Youssef
  • Chatzichristos, Christos
  • Mañu Pereira, Maria del Mar
  • de Brito Robalo, Bruno M.
  • Ghini, Francesco
  • Gutiérrez Torre, Alberto
  • Reidel Lojo, Sara Isabel
Synthetic data generation (SDG) structured health data is increasingly promoted as a solution to longstanding barriers in health data access. It is offering the promise of privacy-preserving data reuse for research, innovation, and policy. Despite rapid technical advances, the adoption of synthetic health data in real-world settings remains limited. Shaped by challenges around data quality, representativeness, infrastructure readiness, trust, and legal uncertainty, this viewpoint draws on experiences from 7 European research initiatives within the HealthData4EU cluster to reflect on how SDG is being operationalized in practice. It synthesizes cross-project insights to highlight recurring methodological and governance tensions and to examine their implications for trust and responsible use. The analysis argues that trustworthy SDG cannot be achieved through technical optimization alone but requires alignment between evaluation practices, upstream data stewardship, regulatory clarity, and sustained stakeholder engagement. Addressing these conditions is essential for moving synthetic data from experimental pilots toward a credible and sustainable component of European health research ecosystems., This paper was funded by the Innovative Health Initiative through the SYNTHIA project. The study draws on insights from 7 research projects that form part of the HealthData4EU cluster. Projects within the cluster and their funding sources are as follows. SYNTHEMA, PHASE IV AI, SECURED, FLUTE, and AISym4MED are funded by the Horizon Europe program. PHEMS is funded by Horizon Europe and UK Research and Innovation. SYNTHIA is funded through the Innovative Health Initiative., Peer Reviewed, Article signat per 18 autors/es: Jens Declerck 1,2, MSc; Dipak Kalra 1, Prof Dr; Antti Airola 3; Ahmed Youssef Ali Amer 4; Christos Chatzichristos 5; Maria del Mar Mañu Pereira 6; Bruno M. de Brito Robalo 7; Francesco Ghini 8; Alberto Gutierrez-Torre 9; Sem Hoogteijling 10; Susanne Hultsch 11; Jan Ramon 12; Sara Reidel 13,14; Francesco Regazzoni 15,16; Luís Silva 17; Inês Silveira 17; Tsekeridou Sofia 18, PhD; Christophe Maes 1,2 / 1- The European Institute for Innovation through Health Data, Oosterzele, Belgium; 2- Unit of Medical Informatics and Statistics, Department of Public Health and Primary Care, The University of Ghent, Ghent, Belgium; 3- Department of Computing, University of Turku, Turku, Finland; 4- Innovative Medicine, Johnson & Johnson, Beerse, Belgium; 5- Department of Electrical Engineering, KU Leuven, Leuven, Belgium; 6- University Hospital Vall d'Hebron, Barcelona, Spain; 7- Department of Obstetrics and Gynaecology, Erasmus MC, University Medical Center, Rotterdam, The Netherlands; 8- Laboratory of Translational Research, Reggio Emilia, Italy; 9- Barcelona Supercomputing Center, Barcelona, Catalonia, Spain; 10- Department of Neurology and Neurosurgery, University Medical Center Utrecht Brain Center, Utrecht, The Netherlands; 11- VEIL.AI, Helsinki, Finland; 12- Institut national de recherche en sciences et technologies du numérique, Lille, France; 13- Vall d'Hebron Research Institute, University Hospital Vall d'Hebron, Barcelona, Spain; 14- BarcelonaTech (UPC), Universitat Politècnica de Catalunya, Barcelona, Spain; 15- University of Amsterdam, Amsterdam, The Netherlands; 16- Università della Svizzera italiana, Lugano, Ticino, Switzerland; 17- Laboratory for Instrumentation, Biomedical Engineering and Radiation Physics, NOVA School of Science and Technology, Caparica, Portugal; 18- Netcompany BELUX, Athens, Greece, Postprint (published version)
Recercat. Dipósit de la Recerca de Catalunya
Publication Contribución de congreso (ConferenceOutput)

SECURED for health: scaling up privacy to enable the integration of the european health data space

  • Regazzoni, Francesco
  • Acs, Gergely
  • Aszalos, Albert Zoltan
  • Avgerinos, Christos
  • Bakalos, Nikolaos
  • Berral García, Josep Lluís
  • Bos, Joppe W.
  • Brohet, Marco
  • Castillo Sanz, Andrés G.
  • Gutiérrez Torre, Alberto
In this paper, we present the SECURED project funded in part by the European Union (EU), Grant Agreement no. 10109571. Views and opinions expressed are those of the authors and do not necessarily reflect those of the EU or the Health and Digital Executive Agency. Neither the EU nor the granting authority are responsible for them., aimed at improving privacy-preserving processing of data in the health domain. The technologies developed in the project will be demonstrated in four health-related use cases and with the involvement of SME's selected through an open funding call., Funded in part by the European Union (EU), Grant Agreement no. 10109571., Peer Reviewed, Article signat per 41 autors/es: Francesco Regazzoni ∗ xviii , Gergely Acs xii , Albert Zoltan Aszalos x, Christos Avgerinos xv , Nikolaos Bakalos xvi , Josep Ll. Berral ††, Joppe W. Bos §, Marco Brohet *, Andres G. Castillo Sanz ‡, Gareth T. Davies §, Stefanos Florescu ∥, Pierre-Elisee Flory ¶, Alberto Gutierrez-Torre ††, Evangelos Haleplidis xi , Alice Heliou ¶, Sotirios Ioannidis †, Alexander Islam El-Kadyxi, 14 , Katarzyna Kapusta ¶, Konstantina Karagianni xi , Pieter Kruizinga ∥, Kyrian Maat ∗, Zoltan Adam Mann ∗, Kalliopi Mastoraki †, SeoJeong Moon §, Maja Nisevic ∗∗, Balazs Pejo xii , Kostas Papagiannopoulos ∗, Vassilis Paliuras xi xix , Paolo Palmieri xiii , Francesca Palumbo ‡‡, Juan Carlos Perez Baun xvii, Peter Pollner x, Eduard Porta-Pardo xiv , Luca Pulina ‡‡, Muhammad Ali Siddiqi ∥, Daniela Spajic ∗∗, Christos Strydis ∥, Georgios Tasopoulos ∗, Vincent Thouvenot ¶, Christos Tselios xi , Apostolos P. Fournaris xi / ∗ University of Amsterdam (NL), † Circular Economy Foundation (BE), ‡ FHUNJ (ES), § NXP (BE), ¶ Thales (FR), ∥ Erasmus Medical Center (NL), ∗∗ KU Leuven (BE), †† Barcelona Supercomputing Center (ES), ‡‡ University of Sassari (IT), x Semmelweis University (HU), xi Industrial Systems Institute/R.C. ATHENA (GR), xii BME (HU), xiii University College Cork (IE), xv Catalink (CY), xvi ICCS (GR), xvii EVIDEN (ES), xiv Josep Carreras Leukaemia Research Institute (ES), xviii Universita della Svizzera italiana (CH), xix University of Patras (GR), Postprint (author's final draft)
Recercat. Dipósit de la Recerca de Catalunya
Publication Contribución de congreso (ConferenceOutput)

Beyond parameter averaging in model aggregation

  • Garcia Recasens, Pol
  • Torres Viñals, Jordi
  • Berral García, Josep Lluís
  • Hauberg, Søren
  • Moreno Muñoz, Pablo
The success of foundation models is strongly linked to scale, which has reinforced the interest in federated learning. With the prohibitive cost of training a large language model (LLM) in mind, little attention has been placed on reusing pre-trained models in collaborative training settings. Self-supervision has also played an important role in this success, but its emphasis has been primarily on data. This paper leverages Bayesian principles to bring self-supervision into the model aggregation toolbox. It introduces self-supervised Fisher merging, a framework that successfully merges models in parameter space without re-visiting data, opening a new door in model reusability. Experimental results build the foundation of our method on tractable linear models, and highlight its potential on aggregating neural networks., This work was supported by a research grant (42062) from VILLUM FONDEN. This project has also received funding from the European Research Council (ERC) under the European Union’s Horizon 2020 research and innovation programme (grant agreement 757360). Our work was funded in part by the Novo Nordisk Foundation through the Center for Basic Machine Learning Research in Life Sciences (MLSS) with reference number NNF20OC0062606. This work has been partially financed by the EU-H2020 programme under grant agreement EU-H2020 GA.952179 and grant agreement EU-HORIZON GA.101095717. Also, it has been partially financed by Generalitat de Catalunya (AGAUR) under grant agreement 2021-SGR-00478., Peer Reviewed, Postprint (published version)
Recercat. Dipósit de la Recerca de Catalunya
Publication Artículo científico (JournalArticle)

FRIDA: Free-rider detection using privacy attacks

  • Garcia Recasens, Pol
  • Horváth, Ádám
  • Gutiérrez Torre, Alberto
  • Torres Viñals, Jordi
  • Berral García, Josep Lluís
  • Pejó, Balázs
Federated learning is increasingly popular as it enables multiple parties with limited datasets and resources to train a machine learning model collaboratively. However, similar to other collaborative systems, federated learning is vulnerable to free-riders — participants who benefit from the global model without contributing. Free-riders compromise the integrity of the learning process and slow down the convergence of the global model, resulting in increased costs for honest participants. To address this challenge, we propose FRIDA: free-rider detection using privacy attacks. Instead of focusing on implicit effects of free-riding, FRIDA utilizes membership and property inference attacks to directly infer evidence of genuine client training. Our extensive evaluation demonstrates that FRIDA is effective across a wide range of scenarios., Project no. 145832, implemented with the support provided by the Ministry of Innovation and Technology from the NRDI Fund, financed under the PD_23 funding scheme. Also, partially financed by the Spanish Ministry of Science (MICINN), the Research State Agency (AEI), and European Regional Development Funds (ERDF/FEDER) under grant agreement PID2021-126248OB-I00, MCIN/AEI/10.13039/501100011033/FEDER, UE, and Generalitat de Catalunya (AGAUR) under grant 2021-SGR-00478. Lastly, also funded by the European Union (Grant Agreement Nr. 101095717, SECURED Project)., Peer Reviewed, Postprint (published version)
Recercat. Dipósit de la Recerca de Catalunya
Publication Artículo científico (JournalArticle)

SynthVal: A framework for validating synthetic medical images

  • Guidotti, Dario
  • Pandolfo, Laura
  • Gutiérrez Torre, Alberto
  • López Rúbio, Omar
  • Pulina, Luca
Synthetic data is increasingly used in medical imaging to overcome data scarcity and privacy constraints. However, assessing the fidelity of synthetic images remains a critical challenge for ensuring their safe and effective use in clinical and AI applications. We present SynthVal, a Python-based framework for validating the quality of synthetic medical images through statistical comparisons in deep feature space. SynthVal extracts semantic image embeddings using transformer-based models and computes similarity metrics – including Fréchet Distance, Wasserstein Distance, and Kullback-Leibler Divergence – between real and synthetic data distributions. The framework is designed for modularity, scalability, and ease of integration into existing workflows via pip installation. We evaluate SynthVal using real images from the CSAW-CC dataset, and synthetic images produced by a model developed by the Barcelona Supercomputing Center. Our experiments systematically benchmark the influence of different feature extraction models and similarity metrics, providing practical insights for selecting validation strategies in medical image synthesis. SynthVal offers a reproducible and extensible solution for quality control in synthetic data pipelines., The work of Dario Guidotti, Laura Pandolfo, and Luca Pulina was supported by EU Project SECURED, ‘‘Scaling Up Secure Processing, Anonymization and Generation of Health Data for EU Cross-Border Collaborative Research and Innovation,’’ under Agreement 101095717. The work of Alberto Gutierrez-Torre and Omar Lopez-Rubio was supported by SECURED under Agreement 101095717; in part by Spanish Ministry of Science (MICINN), the State Research Agency (AEI), and European Regional Development Funds (ERDF/FEDER) under Grant PID2021-126248OB-I00; in part by MCIN/AEI/10.13039/501100011033/FEDER, UE; and in part by the Generalitat de Catalunya (AGAUR) under Grant 2021-SGR-00478., Peer Reviewed, Postprint (published version)
Recercat. Dipósit de la Recerca de Catalunya
Publication Contribución de congreso (ConferenceOutput)

Towards Pareto optimal throughput in small language model serving

  • Garcia Recasens, Pol
  • Zhu, Yue
  • Wang, Chen
  • Lee, Eun Kyung
  • Tardieu, Olivier
  • Youssef, Alaa
  • Torres Viñals, Jordi
  • Berral García, Josep Lluís
Large language models (LLMs) have revolutionized the state-of-the-art of many different natural language processing tasks. Although serving LLMs is computationally and memory demanding, the rise of Small Language Models (SLMs) offers new opportunities for resource-constrained users, who now are able to serve small models with cutting-edge performance. In this paper, we present a set of experiments designed to benchmark SLM inference at performance and energy levels. Our analysis provides a new perspective in serving, highlighting that the small memory footprint of SLMs allows for reaching the Pareto-optimal throughput within the resource capacity of a single accelerator. In this regard, we present an initial set of findings demonstrating how model replication can effectively improve resource utilization for serving SLMs., This work has been partially financed by grant agreement EUHORIZON GA.101095717 and by the EU-HORIZON MSCA programme under grant agreement EU-HORIZON MSCA GA.101086248. Also, it has been partially financed by Generalitat de Catalunya (AGAUR) under grant agreement 2021- SGR-00478, and by the Spanish Ministry of Science (MICINN), the Research State Agency (AEI) and European Regional Development Funds (ERDF/FEDER) under grant agreement PID2021-126248OB-I00, MCIN/AEI/ 10.13039/ 501100011033/ FEDER, UE., Peer Reviewed, Postprint (author's final draft)
Recercat. Dipósit de la Recerca de Catalunya
Publication Contribución de congreso (ConferenceOutput)

Mind the memory gap: unveiling GPU bottlenecks in large-batch LLM inference

  • Garcia Recasens, Pol
  • Agulló López, Ferran
  • Zhu, Yue
  • Wang, Changyang
  • Lee, Eun Kyung
  • Tardieu, Olivier
  • Torres Tomás, Juan José
  • Berral García, Josep Lluís
Large language models have been widely adopted across different tasks, but their auto-regressive generation nature often leads to inefficient resource utilization during inference. While batching is commonly used to increase throughput, per-formance gains plateau beyond a certain batch size, especially with smaller models, a phenomenon that existing literature typically explains as a shift to the compute-bound regime. In this paper, through an in-depth GPU-level analysis, we reveal that large-batch inference remains memory-bound, with most GPU compute capabilities underutilized due to DRAM bandwidth saturation as the primary bottleneck. To address this, we propose a Batching Configuration Advisor (BCA) that optimizes memory allocation, reducing GPU memory requirements with minimal impact on throughput. The freed memory and underutilized GPU compute capabilities can then be leveraged by concurrent workloads. Specifically, we use model replication to improve serving throughput and GPU utilization. Our findings challenge conventional assumptions about LLM inference, offering new in-sights and practical strategies for improving resource utilization, particularly for smaller language models., This work has been partially financed by grant agreement EU-HORIZON GA.101095717 and by the EU-HORIZON MSCA programme under grant agreement EU-HORIZON MSCA GA.101086248. Also, it has been partially financed by Generalitat de Catalunya (AGAUR) under grant agreement 2021-SGR-00478, by Severo Ochoa Center of Excellence CEX-2021-001148-S-20-3, and by the Spanish Ministry of Science (MICINN), the Research State Agency (AEI) and European Regional Development Funds (ERDF/FEDER) under grant agreement PID2021-126248OBI00, MCIN/AEI/10.13039/ 501100011033/ FEDER, UE., Peer Reviewed, Postprint (published version)
Recercat. Dipósit de la Recerca de Catalunya
Publication Contribución de congreso (ConferenceOutput)

HealthMesh: An architectural framework for federated healthcare data management

  • Bisquert Parés, Aniol
  • Hmimou Ham Man, Achraf
  • Berral García, Josep Lluís
  • Gutiérrez Torre, Alberto
  • Romero Moral, Óscar
Recently, significant milestones have been achieved in the field of healthcare data analysis. However, alongside these accomplishments, substantial data-related challenges have emerged in the domain of big data management. Modern healthcare projects are no more dealing with a single data repository but many heterogeneous ones and must overcome data variety, privacy and governance issues. Yet, current solutions face a privacy-decentralization trade-off. To address this dual challenge, we introduce HealthMesh, a novel layered architectural framework based on the Data Mesh principles, providing a domaindecentralised paradigm. In addition, the framework incorporates a Semantic Data Model which establishes robust governance, enables interoperability and guarantees policy compliance for all the data assets. To demonstrate the capabilities of the proposed approach, we provide an illustrative example inspired by the use case of the INCISIVE project for breast cancer analytics. Overall, this work makes a significant contribution on collecting key challenges, identifying actors and providing a set of components and guidelines for establishing a holistic framework for the complex field of healthcare data management., The work reported here was supported by the EU-H2020 programme under GA.952179 INCISIVE, Horizon Europe Programme under GA.101095717 (SECURED) and GA.101135513 (CYCLOPS), the Spanish Research State Agency (MICINN/AEI, ERDF/FEDER) under GA.MCIN/AEI/10.13039/501100011033/FEDER (UE DALEST), the Spanish Ministerio de Ciencia e Innovación under project PID2020-117191RB-I00 / AEI/10.13039/501100011033 (DOGO4ML), and the Generalitat de Catalunya (AGAUR) 2021-SGR-00478 (CROMAI)., Peer Reviewed, Postprint (published version)
Recercat. Dipósit de la Recerca de Catalunya
Publication Artículo científico (JournalArticle)

Towards generalizable Federated Learning in medical imaging: a real-world case study on mammography data

  • Tzortzis, Ioannis N.
  • Gutiérrez Torre, Alberto
  • Sykiotis, Stavros
  • Agulló López, Ferran
  • Bakalos, Nikolaos
  • Doulamis, Anastasios
  • Doulamis, Nikolaos
  • Berral García, Josep Lluís
Federated Learning has been rapidly gaining in popularity in medical applications, due to the increased privacy offered, since medical data doesn't need to leave the hospitals' premises for AI model training. However, a direct translation of a classic experiment to a federated one is not always straightforward. In this work, we delve into the intricacies of federated learning for a breast cancer classification tool. We compare classic model training with a federated variant, and highlight the adaptations that need to be taken care of to ensure the equivalence between the two. Specifically, we introduce the Breast Area Detection tool as an essential component of the pre-processing pipeline to enhance the robustness of Federated Learning by offering data harmonization. On top of that, we present an end-to-end Federated Learning framework that is effective for real-world data and scenarios. Among the three real-world hospitals involved in the experimental procedure, the proposed framework significantly improves performance at the first hospital, providing consistent results similar to those achieved in the classic approach. Experimental results demonstrate that the interventions introduced improved model performance by approximately 35%, aligning federated learning and centralized model performance., This paper is funded by the EU funded projects SECURED “Scaling Up Secure Processing, Anonymization And Generation Of Health Data For EU Cross Border Collaborative Research And Innovation” with grant agreement no. 101095717 and INCISIVE “A multimodal AI-based toolbox and an interoperable health imaging repository for the empowerment of imaging analysis related to the diagnosis, prediction and follow-up of cancer” with grant agreement no. 952179. This paper has been partially financed by the Spanish Ministry of Science (MICINN), the Research State Agency (AEI) and European Regional Development Funds (ERDF/FEDER) under grant agreement PID2021-126248OB-I00, Severo Ochoa Center of Excellence CEX-2021-001148-S-20-3, 2021- SGR-00478, UE, also Generalitat de Catalunya (AGAUR) under grant agreement 2021-SGR-00478., Peer Reviewed, Postprint (published version)