Periodic Reporting for period 1 - TALON (Autonomous and Self-organized Artificial Intelligent Orchestrator for a Greener Industry 4.0)
Berichtszeitraum: 2022-10-01 bis 2024-03-31
TALON aims at sculpturing the road towards the next Industrial revolution by developing a fully-automated AI architecture capable of bringing intelligence near the edge in a flexible, adaptable, explainable, energy and data efficient manner. TALON architecture consists of three fundamental pillars: a) an AI orchestrator that coordinates the network and service orchestrators in order to optimise the edge vs cloud relationship, while boosting reusability of datasets, algorithms and models by deciding where each one should be placed; b) a lightweight hierarchical blockchain schemes that introduce new service models and applications under a privacy and security umbrella; and c) new digital-twin empowered transfer learning and visualization approaches that enhance AI trustworthiness and transparency. It combines the benefits of AI, edge and cloud networking, as well as blockchain and DTs, optimized by means of a) new key performance indicators that translate the AI benefits into insightful metrics; b) novel theoretical framework for the characterisation of the AI; c) blockchain used to deliver personalised & perpetual protection based on security, privacy and trust mechanisms; d) AI approaches for automatically and co-optimising edge and cloud resources as well as the AI execution nodes; e) semantic AI to reduce the learning latency and enhance reusability; and f) digital twins that visualize the AI outputs and together with human-in-the-loop approaches. All the technological breakthroughs are demonstrated, validated and evaluated by means of proof-of-concept simulation and four real-world pilots.
1. Within the first reporting period the key result is the architecture design of a microservice-based infrastructure representing a dynamic, efficient, and impactful system. At the heart of our design there's orchestration of computational resources, AI models, and data to optimize system effectiveness and performance.
2. Few-shot Learning models are in the process of being integrated within the TALON platform, that are able to decrease the energy footprint and therefore the computing, processing and training time and cost without sacrificing the models’ performance and accuracy.
3. Several mechanisms are in place to guarantee high-level security and privacy in heterogeneous application environments (e.g. the adoption of containerization technologies, micro segmentation through virtual overlay and programmable networks breaks down networks into smaller, more secure zones, limiting lateral movement of malicious users within the environment, Role-based Access Control (RBAC), Certificate and Identity Management which enable granular access control, a private peer to peer distributed ledger technology is adopted to provide the necessary trustworthiness and security umbrella to protect the E2C network and enhance the federated AI capabilities.)
4. In order to support a set of mechanisms for balancing complex trade-offs (e.g. energy, security, latency), the four primary use cases of the project were analysed, providing comprehensive descriptions, objectives, and identification of both current and desired states.
5. The object store and metadata serving for the AI reusability were deployed, which contains all the trained AI models which are suitable for specific data modalities (e.g. images, time-series and tabular / categorical data) and learning tasks (e.g. classification and regression). Last, this object store will be further enriched in the future with additional AI models coming from the Use Cases to further extend model reusability.
6. TALON’s AI Theoretical Framework has been designed by collecting requirements from the Use Cases and the TALON subsystems using AI models and has adopted the best practices of the MLOps principles.
7. Significant progress has been recorded related with the XAI framework, having developed the support to the input data reliability and feature engineering effectiveness. Various techniques have been adopted to cover time-series and image input data and preprocess.
2. Few-shot Learning models are in the process of being integrated within the TALON platform, that are able to decrease the energy footprint and therefore the computing, processing and training time and cost without sacrificing the models’ performance and accuracy.
3. Several mechanisms are in place to guarantee high-level security and privacy in heterogeneous application environments (e.g. the adoption of containerization technologies, micro segmentation through virtual overlay and programmable networks breaks down networks into smaller, more secure zones, limiting lateral movement of malicious users within the environment, Role-based Access Control (RBAC), Certificate and Identity Management which enable granular access control, a private peer to peer distributed ledger technology is adopted to provide the necessary trustworthiness and security umbrella to protect the E2C network and enhance the federated AI capabilities.)
4. In order to support a set of mechanisms for balancing complex trade-offs (e.g. energy, security, latency), the four primary use cases of the project were analysed, providing comprehensive descriptions, objectives, and identification of both current and desired states.
5. The object store and metadata serving for the AI reusability were deployed, which contains all the trained AI models which are suitable for specific data modalities (e.g. images, time-series and tabular / categorical data) and learning tasks (e.g. classification and regression). Last, this object store will be further enriched in the future with additional AI models coming from the Use Cases to further extend model reusability.
6. TALON’s AI Theoretical Framework has been designed by collecting requirements from the Use Cases and the TALON subsystems using AI models and has adopted the best practices of the MLOps principles.
7. Significant progress has been recorded related with the XAI framework, having developed the support to the input data reliability and feature engineering effectiveness. Various techniques have been adopted to cover time-series and image input data and preprocess.
Innovations beyond the State-of-the-Art (SoTA) envisaged in TALON are summarised as follows:
- Realistic decision-making process modelling and AI fundamental analysis: the TALON’s E2C approach will strive to provide faithful and simple models that outperform the trivial union of rules and are competitive with natively global explanators especially in terms of complexity.
- New AI-specific KPIs definition: TALON aims in defining a dynamic and independent evaluation pipeline coupled with self-optimizing and maintaining methodologies to further minimize human intervention in AI-powered systems, making them more robust and able to hold heavier duties.
- Novel digital-twin empowered transfer learning: TALON architecture provides an advanced transparency in the operation of the AI since they actively partici-pate in the design of the symbolic AI component increasing the trust and the transparency to the system.
- Experimentally-driven semantic AI: TALON adopts an XAI solution with semantics segmentation and visualisation capabilities, considering 5 Trust Levels (TrL): (a) raw data (TrL1), (b) feature engineering & labelling (TrL2), (c) learning methods/algorithms (TrL3), (d) concrete ML/DL models and (e) evaluation & user expectation (TrL5).
- E2C AI orchestrator: the E2C AI orchestrator is responsible for controlling the data flow between different nodes that execute AI and AI-related tasks. This is achieved by solving another challenging problem, i.e. to minimize data sharing, while ensuring that each node has the appropriate amount and type of data in order to output accurate decisions.
- Edge AI with real-time learning capabilities: new data analytics approaches will allow to execute federated and/or reinforcement learning in the edge provide a viable solution that offers scalability and reusability.
- Security and privacy: TALON will realize an innovative AI-based architecture, which will incorporate lightweight hierarchical blockchain functionalities to support e2e security, and with the combination of AI-powered solutions, new, security-enhanced services will be introduced.
- SDN-enabled EC for autonomous self-healing, self-recovery, and self-configuration: TALON develops a two-layers distributed federated learning framework (2L-FLF), in which the model training is distributed across a number of edge devices that are selected as a solution of a well-defined optimization problem with constraints the energy autonomy of each device and its computational capabilities.
- Novel resource mobility manager: TALON will deliver an intelligent resource manager that is able to identify the computational resources required by a specific task and make the appropriate offloading decisions.
- New AI visualization and monitoring approaches: TALON aims in boosting the explainability and transparency of the AI system, by adopting DT methodologies in the Visualization and Reporting Layer, which will provide the ability to visualize AI outputs and the AI’s decisionmaking processes to the human factor, in a comprehensible manner
- Smart pricing: pricing policies will focus on providing computational offloading incentives. These incentives will assist TALON to create an efficient E2C framework, in which the reward process will motivate collaboration among different trusted entities by leveraging smart contracts’ selfexecution nature.
- Realistic decision-making process modelling and AI fundamental analysis: the TALON’s E2C approach will strive to provide faithful and simple models that outperform the trivial union of rules and are competitive with natively global explanators especially in terms of complexity.
- New AI-specific KPIs definition: TALON aims in defining a dynamic and independent evaluation pipeline coupled with self-optimizing and maintaining methodologies to further minimize human intervention in AI-powered systems, making them more robust and able to hold heavier duties.
- Novel digital-twin empowered transfer learning: TALON architecture provides an advanced transparency in the operation of the AI since they actively partici-pate in the design of the symbolic AI component increasing the trust and the transparency to the system.
- Experimentally-driven semantic AI: TALON adopts an XAI solution with semantics segmentation and visualisation capabilities, considering 5 Trust Levels (TrL): (a) raw data (TrL1), (b) feature engineering & labelling (TrL2), (c) learning methods/algorithms (TrL3), (d) concrete ML/DL models and (e) evaluation & user expectation (TrL5).
- E2C AI orchestrator: the E2C AI orchestrator is responsible for controlling the data flow between different nodes that execute AI and AI-related tasks. This is achieved by solving another challenging problem, i.e. to minimize data sharing, while ensuring that each node has the appropriate amount and type of data in order to output accurate decisions.
- Edge AI with real-time learning capabilities: new data analytics approaches will allow to execute federated and/or reinforcement learning in the edge provide a viable solution that offers scalability and reusability.
- Security and privacy: TALON will realize an innovative AI-based architecture, which will incorporate lightweight hierarchical blockchain functionalities to support e2e security, and with the combination of AI-powered solutions, new, security-enhanced services will be introduced.
- SDN-enabled EC for autonomous self-healing, self-recovery, and self-configuration: TALON develops a two-layers distributed federated learning framework (2L-FLF), in which the model training is distributed across a number of edge devices that are selected as a solution of a well-defined optimization problem with constraints the energy autonomy of each device and its computational capabilities.
- Novel resource mobility manager: TALON will deliver an intelligent resource manager that is able to identify the computational resources required by a specific task and make the appropriate offloading decisions.
- New AI visualization and monitoring approaches: TALON aims in boosting the explainability and transparency of the AI system, by adopting DT methodologies in the Visualization and Reporting Layer, which will provide the ability to visualize AI outputs and the AI’s decisionmaking processes to the human factor, in a comprehensible manner
- Smart pricing: pricing policies will focus on providing computational offloading incentives. These incentives will assist TALON to create an efficient E2C framework, in which the reward process will motivate collaboration among different trusted entities by leveraging smart contracts’ selfexecution nature.