Extreme Food Risk Analytics

Projektinformationen

EFRA

ID Finanzhilfevereinbarung: 101093026

DOI

10.3030/101093026

Projekt abgeschlossen

EK-Unterschriftsdatum 17 November 2022

Startdatum 1 Januar 2023

Enddatum 31 Dezember 2025

Finanziert unter

• Digital, Industry and Space

Gesamtkosten

€ 4 833 797,50

EU-Beitrag

€ 4 833 797,50

Koordiniert durch

AGROKNOW IKE
Greece

Periodic Reporting for period 1 - EFRA (Extreme Food Risk Analytics)

Berichtszeitraum: 2023-01-01 bis 2024-06-30

Zusammenfassung vom Kontext und den Gesamtzielen des Projekts

EFRA operates within a complex social and economic context characterized by the intricate interplay of various drivers, including economic, socio-economic, environmental, and climate factors that can impact food safety risks. The need for a proactive approach to food safety is imperative, shifting from reactive measures to predictive analytics to prevent incidents before they occur. Data sharing and interoperability are in their infancy, hindered by the sensitivity of food safety information and legal constraints. However, integrating high-quality, multilingual, and heterogeneous data from public and private sources is crucial for accurate risk prediction. The project also aligns with the growing emphasis on sustainability, aiming to reduce energy consumption and improve AI efficiency. By leveraging federated learning, NLP, and other AI technologies, EFRA seeks to enhance data privacy and security, ensuring compliance with regulations like GDPR and the AI Act. EFRA aims to create the first secure-by-design green data space for AI-driven food risk prevention, utilizing advanced data technologies and AI models. The integrated platform collects, processes, and analyzes diverse food safety data. Intelligent crawlers and federated learning frameworks ensure comprehensive, privacy-aware data handling.
Objective 1: Improve Searching, Mining, and Processing of Dispersed, Multilingual, Heterogeneous, Deep/Hidden Food Safety-Related Data Sources
Main Results: The Food Safety Scraping Platform was deployed, integrating new sources. An AI-assisted web spider using fine-tuned RoBERTa achieved high accuracy in classifying web pages. NLP models for classifying food incident reports and tracking emotions were developed, along with a benchmark dataset and advancements in data summarization and video analysis.
Objective 2: Foster the Adoption of AI-Enabled Food Risk Predictions through Novel Methods for Privacy-Preserving AI Training and Explainability
Main Results: A Federated Learning architecture using the Flower framework was established.. AI models for outlier detection and time series analysis were developed, including a Timeseries Predictions Engine. Explainability techniques were implemented, and an international summit was conducted.
Objective 3: Explore Novel Techniques for Green AI and Their Connection to Green HPC Infrastructure Operations
Main Results: Sustainable training frameworks combining pruning and binarization were proposed. An energy-efficient summarization pipeline for regulatory texts was developed. Contributions included early-exit inference strategies and a framework for k-nearest neighbor search on FPGA architectures.
Objective 4: Deliver an Analytics-Capable FAIR EU Data Space for Food Safety Data and Predictions
Main Results: The Data Hub and Analytics Powerhouse REST APIs were defined, and core functionalities were implemented. A Kubernetes cluster was configured, and the components were deployed as microservices. The EFRA API Gateway and Data & Analytics Marketplace requirements were specified.
Objective 5: Validate, Evaluate, and Demonstrate the Developed Tools in Real-World Use-Cases
Main Results: EFRA's tools were validated in several use-cases. AI models for outlier detection in poultry, pest prediction, and real-time regulatory decision-making were implemented. A use-case on mycotoxin presence in animal feed was introduced, showcasing EFRA’s innovations' practical impact.

Arbeit, die ab Beginn des Projekts bis zum Ende des durch den Bericht erfassten Berichtszeitraums geleistet wurde, und die wichtigsten bis dahin erzielten Ergebnisse

Outcome 1. Deployment of the Food Safety Scraping Platform
Progress: Enhanced the platform with an AI-assisted web spider based on fine-tuned RoBERTa. Developed NLP models for classifying food incident reports, a benchmark dataset, and advancements in emotion detection and video analysis.
Outcome 2. Development of Novel NLP Models
Progress: Developed advanced NLP models for food safety data, including web page classification, a benchmark dataset, emotion detection in political speech, and a novel NLP method using conformal prediction. Prepared a dataset for regulatory text summarization and designed an energy-efficient summarization pipeline. Developed a video analysis pipeline for food safety authorities.
Outcome 3. Establishment of Federated Learning Architecture
Progress: Established a Federated Learning architecture using the Flower framework for privacy-preserving AI model training. Initiated an exemplary use-case on mycotoxins in poultry.
Outcome 4. Explainable AI Models & Time Series Forecasting
Progress: Developed the Timeseries Predictions Engine, enhancing EFRA's predictive capabilities. Created AI models for outlier detection and predicting extreme events. Implemented Explainable AI techniques, including SHAP, relevance propagation, and TF-IDF-based approaches, enhancing model transparency.
Outcome 5. Green AI Techniques
Progress: Designed compression techniques and energy-efficient architectures, focusing on pruning, binarization, and FPGA-based k-nearest neighbor search.
Outcome 6. Validation of AI Models through Use-Cases
Progress: Validated AI models through real-world use-cases, including risk prevention in poultry, pest prediction, and real-time regulatory decision-making. Introduced a use-case on mycotoxin presence in animal feed.

Fortschritte, die über den aktuellen Stand der Technik hinausgehen und voraussichtliche potenzielle Auswirkungen (einschließlich der bis dato erzielten sozioökonomischen Auswirkungen und weiter gefassten gesellschaftlichen Auswirkungen des Projekts)

1. Enhanced AI-Driven Algorithms for Multilingual and Heterogeneous Food Safety Data Mining and Processing
Progress: Developed an AI platform for integrating and processing multilingual, heterogeneous data. Innovations include an AI-assisted web spider and sophisticated NLP models for classifying and categorizing food incident reports.
Potential Impacts: Enhanced food safety monitoring and incident prevention, reducing foodborne illnesses, healthcare costs, and improving consumer confidence.
2. Sustainable AI Model Training and Exploitation
Progress: Pioneered sustainable AI model training techniques, combining pruning and binarization to compress neural networks. Developed energy-efficient summarization pipelines for regulatory texts.
Potential Impacts: Reduced computational and environmental footprint of AI models.
3. Early Outlier Detection in Poultry Farm Operations (Mortality and End-Cycle Weights)
Progress: Developed AI models for early outlier detection in poultry operations, targeting mortality rates and end-cycle weights using advanced time series analysis and survival analysis techniques.
Potential Impacts: Enhanced animal welfare, reduced mortality rates, optimized production processes.
4. Implementation of Explainable AI Techniques
Progress: Implemented cutting-edge Explainable AI techniques, including SHAP, backpropagation-based methods, and TF-IDF-based approaches to enhance transparency and interpretability of AI models.
Potential Impacts: Improved decision-making for regulatory bodies and industry stakeholders, greater transparency and accountability.

efra.jpg