Our vision is that Artificial Intelligent Systems are irreversibly set on the evolutionary path of every future vertical as well as of every object and service we humans will interact with in the near future. This trend is motivated by the need to support elastic and demanding real-world use cases such as automated and cooperative mobility, e-health, gaming, entertainment, etc. For this reason, in AI@EDGE we leveraged the concept of reusable, secure, and trustworthy AI for network automation to achieve an EU-wide impact on industry-relevant aspects in multi-stakeholders’ environments.
AI@EDGE approach to answer the above-mentioned challenges has two lines of action. First, we designed, prototyped, and validated a network and service automation platform capable of supporting flexible and programmable pipelines for the creation, utilization, and adaptation of secure and privacy-aware AI/ML models. Second, we used this network and service automation platform to orchestrate AI-enabled end-to-end applications. Here, we introduced the novel concept of Artificial Intelligence Functions (AIFs) to refer to the AI-enabled end-to-end applications sub-components that can be deployed across the AI@EDGE platform.
Six are the AI@EDGE project main breakthroughs:
AI/ML for closed loop automation
The AI@EDGE Network and Service Automation Plaftorm (NSAP) is a framework for automated network management that provides an environment for data-driven, intelligent, methods that support decision making. The NSAP makes use of the concept of closed-loop network intelligence, where the goal is to relieve the human operator as far as possible from the need to take manual action to operate the system. Closed-loop control is described by several steps that form the loop, including sensing, monitoring, aggregation, training/inferring, decision making, and taking action (orchestrating/actuating). To support these steps, a data collecting pipeline and an AI model pipeline have been designed, prototyped and evaluated in the project as part of the NSAP framework. Three classes of AI/ML-based automation methods have been developed: (i) methods for optimising the deployment of distributed AI, (ii) methods for predicting or forecasting future performance, and (iii) methods for monitoring and preparing data.
Privacy preserving, machine learning for multi-stakeholder environments
The AI@EDGE environment with the NSAP and the connect-compute platform (CCP) providing edge computing for application services is in general a multi-stakeholder environment for data-driven AI/ML-based functions and services. The data collecting pipeline is a scalable framework for providing platform data in a scalable and controlled way to services, with the possibility for data filtering and transformation including for privacy preserving purposes. The environment furthermore supports federated learning methods, including training, which enable local processing without the need to collect privacy-sensitive data to central locations, e.g., for training, or for its operation. Methods for orchestrating and deploying such distributed methods have been developed using an example service in the form of distributed anomaly detection.
Distributed and decentralized connect-compute platform
A novel MEC Orchestrator (MEO) and a Multi-Tier Orchestrator (MTO) are the core of the Connect-Compute Platform (CCP). The MEC orchestrator handles the management of distributed MEC systems and the lifecycle of MEC applications, while the MTO enables the inclusion of multiple edge and cloud orchestrators. Additionally, an innovative interface aligned with ETSI MEC specifications facilitates application migration across MEC Systems. These CCP innovations empower the MEOs to operate independently of the MTO in case of failures, thereby enhancing the platform’s resilience. Furthermore, the CCP’s modular design facilitates seamless integration with other components like IARM and intelligent orchestrator components. This technology provides efficient management of distributed MEC applications, seamlessly integrates with existing container-based technologies, and enhances system robustness, shaping the landscape of distributed computing.
Provisioning of AI-enabled applications
To support provisioning and life-cycle management of AI-enabled applications over connect-compute platform new methods and tools for the provisioning and life-cycle management of the AI-enabled applications over connect-compute platform. Building upon an extended representation of AI functions, including acceleration capabilities, characteristics of the underlying ML model and data, the techniques engage Serverless technologies and AI-related solutions to deploy, monitor, reconfigure / retrain, and update the AI-intensive apps. Within AI@EDGE project AIF specification is used by the CCP components for AIF provisioning, while the specific techniques and tools have been provided and integrated together with the Serverless framework in compliance with the CCP implementation. In this way these methods and tools support zero-touch automation and continuous management of the life-cycle activities necessary for AI-intensive applications at the edge.
Hardware-accelerated serverless platform for AI/ML
CCP platform adopts and uses multiple heterogeneous programmable accelerators with diverse HW microarchitectures and SW tools, which enable increased parallelization and customization of the AI functions towards enhancing the throughput/latency/energy of CCP. The technology includes a framework for mapping high-level TensorFlow code to vendor-specific executable, as well as a SW infrastructure for Intelligent Accelerators Resource Management (IARM) accommodating dynamic adaptation to abrupt changes. This approach has been Integrated in the CCP via installation of HW card/SoC’s and OS drivers, containerization technology, and automated management via custom IARM transactions. In this way an order of magnitude improvement of CCP throughput/latency, with less than a second overhead during AIF assignment has been achieved and demonstrated.
Cross-layer, multi-connectivity and disaggregated radio access
A novel transport-layer multi-connectivity scheduling technique to jointly use multiple radio-access-technologies. The proposed technique handles at the device and intermediate proxy level the scheduling of IP packets over multiple interfaces, where at least one interface corresponds to the 5G radio access interface, and at least one interface is WiFi. The intermediate proxy can sit in between the 5G router (UPF) and the application server, if the latter has no multipath transport support (namely using MPTCP). Other interfaces can moreover be LiFi and Ethernet interfaces type. This technique has been engaged in UC4 to aggregate WiFi and 5G interfaces. In in-cabin communications, 5G, Ethernet, WiFi and LiFi interfaces can be made available at seat screens and/or bring-your-own-device nodes so that inefficiency of one access interface can be automatically compensated by the multi-connectivity scheduler in the choice of the downlink path to the user equipment. The technique allows for restoration time upon RAT failure below 50ms and increased throughput up to 100%.