New approaches for decentralized, federated and sustainable AI data processing (RIA)

Expected Outcome:Project results are expected to contribute to developing new approaches, tools and techniques that overcome the obstacles of today's centralised AI compute techniques: limits in the availability of energy and AI compute capacity in centralised standalone environments, limited availability of types of AI chips, data quality and security and latency in AI data processing. The ultimate objective is to help overcome EU’s AI compute capacity bottlenecks by offering alternative decentralised and sustainable AI compute models that enable exploitation of diverse hardware processing architectures and scaling approaches.Scope:This topic focusses on technologies and techniques that enable AI data processing to leverage distributed compute resources across the cloud and edge computing continuum throughout the whole AI model lifecycle from data collection, training, fine-tuning, and deployment. To overpass today’s state of the art in the area, the considered research areas include:To research on distributed, decentralised, and federated “compute continuum” enabled AI architectures beyond federated learning and integrating model compression tools and new mechanisms to enable AI data processing to scale across multiple and diverse computing infrastructures.Development, deployment, and operation of AI workflows across heterogeneous and distributed infrastructures along the compute continuum (edge, cloud, HPC), including the possibility of incorporating innovative computing paradigms (neuromorphic and quantum computing) and hardware efficiency enhancements ((e.g., including in-memory computing, and hardware and software approximation).Novel methods and techniques to improve data availability and consistency for decentralised AI data processing. These consider tools to ensure data quality (e.g. prevention of data sets imbalance or inconsistency across distributed data sources), volume optimisation for data transfers across environments, and distributed data managemen