AI/ML services
The success of a multi-provider cloud-edge continuum depends on seamlessly integrating AI service capabilities with the underlying infrastructure while fully leveraging its distributed and dynamic properties. To achieve this, algorithms must be designed:
to operate in a distributed manner, ensuring that tasks are effectively shared across the continuum to optimize performance and minimize latency.
be cooperative, enabling seamless interaction and coordination among diverse providers, devices, and nodes to enhance resilience, scalability, and functionality.
be resource-efficient, making optimal use of available processing power, storage, communication bandwidth, and energy across the infrastructure.
This approach not only ensures compliance with the continuum's architectural requirements but also enables secure, efficient, and scalable data processing that aligns with the demands of modern, interconnected systems.
Research directions
Cooperative and collaborative AI algorithms
Our goal is to develop innovative methods for coordinating deep learning agents using complex network approaches. This includes decentralizing AI algorithms, exploring a mixture of agents, and developing socially interpretable reinforcement learning methods. We aim to advance AI by improving deep learning agent coordination, enhancing performance in distributed systems, and fostering human-machine cooperation. Key areas include:
Implementing decentralized AI algorithms for efficient coordination.
Improving system performance with a mixture of agents.
Ensuring transparency with socially interpretable reinforcement learning.
We will explore federated learning techniques, develop cooperative self-rewarding schemes for language models, and design cooperative embodied agents. Finally, we will investigate multi-agent reinforcement learning algorithms to coordinate human-driven and autonomous vehicles, focusing on roundabout scenarios.
Dynamic and resource-aware AI algorithm
We aim to develop AI algorithms that adapt to available computational resources and efficiently use the resources available (data, energy and processing capabilities) by investigating three directions:
The development of methods to reduce compute and data requirements for state-of-the-art models, such as exploring layer-dropping/routing, dynamic quantization, pruning of large multimodal models, tiny transformers, zero-shot & training-free methods for Large Multimodal Models.
The development of hardware-aware deep architectures for multi-modal data processing based on emerging architectures, including state-space models, summary-mixing, tinyCLAP, on-device learning, GenAI and hardware-aware scaling.
The development of AI solutions for low-precision data processing, where low-precision refers to both downsampling for computation efficiency as well as varying operational conditions with noise or corruptions, as: approaches for video-text alignments and video interpretation, test-time adaptation to the operational conditions and the quality of the data of LMM.
AI-enabled edge-cloud continuum
The dynamic and adaptive architectures for AI/ML services developed in the project are by-design “orchestratable” in the edge-cloud framework. The infrastructure will leverage these characteristics to address the challenges of efficient and adaptive service deployment and orchestration in heterogeneous environments.
These methods ensure seamless deployment of AI services in cloud-to-edge environments.
Key aspects include deployment automation, such as resource mapping, selection, and autoscaling, and orchestration automation, leveraging a composition infrastructure (data and control planes) and declarative management through high-level specifications for composition, requirements, constraints, and monitoring.