Sustainable Artificial Intelligence (SAI)

Ico_CTTC

Investigating the sustainability of AI and AI for sustainability

Sustainable AI research unit targets a twofold objective: we aim to address the sustainability of developing and using AI systems and, we direct AI usage towards the sustainable development goals (i.e., AI4Good). In other words, for us, it is of upmost importance to design AI systems for sustainable development as well as explicitly targeting the sustainability of AI training and usage.

Our RU is, therefore, multi-disciplinary in nature and spans several research areas including data science, computer science, network science, information engineering, wireless communications, energy engineering, environmental engineering and remote sensing.

Energy-aware AI solutions for sustainable development

Our research activities devise energy-aware, high-accuracy and explainable Machine Learning (ML) models to support data processing from distributed sources such as sensors, wearable devices, mobile phones, base stations and other edge network entities, for system control and optimization.

Our proposals target a holistic approach to learning systems; meaning that our algorithms will be designed to solve multiple learning tasks such as prediction, classification, anomaly detection and automatic control. The discovery of elusive governing dynamics in physical processes is also considered in our analysis.

Decentralized learning is our main area of research. Such approach allows to distribute computing power among several network entities without completely relying on energy-hungry cloud data centres. Our studies focus on the orchestration of such distributed process via Federated learning, Gossip learning as well as Multi-Agent Reinforcement learning. We also consider methods to improve the efficiency  of our ML models via Knowledge Transfer learning, Meta-learning and Continual learning.

Sustainable Artificial Intelligence | Research units | CTTC

Given the distributed nature of our methodology, communications and networking among learning nodes play a key role. For this, ML-based semantic communication is studied to maximize the communication system capacity by capitalizing on the largely untapped intrinsic attributes of the conveyed information. Moreover, Blockchain technology is analysed and tailored into sustainable learning systems to achieve high reliability and security during the learning process iterations.

An additional goal of our methodology is to provide explainable insight and interpretability into a system’s behaviour  by introducing appropriate observational, inductive, or learning biases that steer the learning process towards identifying physically consistent solutions.

Research lines

In this research line, we study how to efficiently use AI algorithms on Edge services, as it is the case with the widely and increasingly popular federated learning. Of relevance here is the fact that Edge AI needs to consider how to run AI models on constrained devices and how to integrate synergies from end-user and devices. Our methodology follows Green AI principles, therefore accuracy and energy efficiency are always assessed in our design.

To this respect, our effort will concentrate on the following topics:

  • Investigating computational and communication efficient solutions for performing distributed model training and inference on the Edge.
  • Providing security, trustworthiness and resiliency against attacks and faults, using decentralized anomaly detection methods. Distributed ledger technology (blockchain) is also considered in our design.
  • Solutions for increasing the computational efficiency of Pervasive AI, such as Knowledge Transfer Learning, Continual Learning, Meta Learning.
  • Exploration of novel promising cutting-edge AI solutions for further increasing the computation and energy efficiency, such as the Quantum Machine Learning (ML) and neuromorphic architectures.

 

In this research line, we design control frameworks for cyber-physical systems (CPS), which sense different underlying dynamic environments such as mobile traffic demands, urban mobility patterns, industrial IoT, assisted and autonomous driving, etc. In our methodology, we state the problem as a multi-objective optimization, considering the several and possibly conflicting optimization targets (e.g., accuracy and energy consumption). We mainly conceive data-driven methods, which are capable to capture the non-linear, high-dimensional, multi-scale characteristics of the networked CPSs.

Our work is spanning several research topics and, in particular:

  • Decentralized and collaborative learning applied to CPSs.
  • Efficient ML through semantic information and data compression.
  • Energy-aware network optimization via intelligent decision-making,
  • Explainable insights and interpretability of the proposed ML-based solutions.

Here, our effort goes in the direction of using AI capabilities to support United Nation (UN) Sustainable Development Goals (SDGs), mitigating the climate crisis and fostering effective environmental governance. To achieve such goals, we design digital twins of the natural environment to capture the elusive and hidden dynamics of a highly complex system like our planet.  We use data collected from multi-temporal and multi-sensor platforms (e.g., satellite remote sensing, instrumentation of the Earth’s surface or sub-surface), integrate them to extract knowledge and support environmental science needs.

Our target scenarios are:

  • Sustainable cities and communities, enabling safe, efficient and clean transportation systems.
  • ML for Earth observation, enabling conservation and sustainable use of the terrestrial ecosystem and oceans.