Today, Internet of Things (IoT) sensors are being extensively used for monitoring processes/phenomena in smart cities. The data samples generated by these IoT sensors are wirelessly transmitted to servers at the network edge where compute-intensive Machine Learning (ML) models, specifically Deep Neural Networks (DNNs), are used for providing inference. However, a large percentage of data samples are redundant because they do not (significantly) improve inference. This leads to an excessive and unjustified carbon footprint of these systems as each redundant data sample will contribute to the Total System Energy (TSE) consumption. However, there is a lack of research on the design of these systems to reduce the TSE by considering the redundancy in the data. In DIME, we explore the TSE energy savings in a distributed inference setup by envisaging the deployment of the emerging small DNN models on the IoT sensors. My objective is to maximize TSE energy savings by answering two key questions: 1) when should an IoT sensor sample the process (to reduce redundant samples) and 2) where to do the inference on the sample, on the IoT sensor or at the edge server (to reduce TSE)? I will develop a general modelling framework and subsequently design and validate scheduling algorithms and sampling techniques that minimize the TSE by reducing the redundant data and maximize accuracy in ML-based monitoring systems. To achieve the objective, I will leverage my theoretical research experience on modelling and design and analysis of algorithms and the expertise of IMDEA Networks in applied machine learning and systems research. DIME directly contributes to reducing the carbon footprint of monitoring in smart cities, which is in line with the goal of Horizon Europe to achieve 100 climate-neutral smart cities by 2030.

To fully exploit the 6G potential and ensure near-zero latency, infinite capacity, and 100% reliability and availability communications, Mobile Network Operators (MNOs) are expected to deploy Zero-touch Network and Service Management (ZSM) solutions that completely automate the resource orchestration and work at a very fast timescale, and Artificial Intelligence (AI) is regarded as the primary enabler for proactive decision-making algorithms that will underpin ZSM. However, the robustness and trustworthiness of AI predictors are critical aspects and represent one of the major barriers presently withholding MNOs from trusting ZSM technologies. In fact, all existing studies on mobile traffic forecasting work under assumption of stationary network, but, in the case of mobile network Key Performance Indicators (KPIs) prediction, user demands and network configurations are time-varying (non-stationary) in operational mobile networks. For example, update of antenna configurations or shifts in popularity of mobile applications can happen over time. Thus, this project is termed "practical AI for Time-vAryiNG network traffic fOrecasting in 6G" (6G-AI-TANGO) and aims to: (i) assess and quantify the severity of temporal changes (due to non-stationarity) in 6G network KPI forecasting, (ii) design AI models tailored to mobile KPI forecasting that are resilient to the long-timescale temporal variations regularly encountered in real-world user demands and network configurations, and (iii) verify the robustness of proposed predictors on real-word data and in production-grade 6G mobile network during non-academic placement at Telefónica.

Stroke is a first-order medical problem (about 600,000 strokes occurred in the EU in 2015), in which rehabilitation is critical. Currently, there are no reliable systems to monitor the patient adherence to this rehabilitation, nor its effectiveness. Combining the ER experience on biosensors and gamification, the expertise on outlier detection and machine learning of IMDEA Networks, and the knowledge on deep learning applied to medicine of the AI Lab at Brown University, in MAESTRO, we will develop algorithms capable of determining rehabilitation adherence and effectiveness by using wearables. This will optimize rehabilitation and forecast recovery by providing information to neurologists and feedback to patients and caregivers. MAESTRO aligns with the H2020 goals in Area III (digitization, research and innovation) as well as health, demographic change, and wellbeing. MAESTRO aims at recruiting 50 patients from Rhode Island Hospital for 4 months in the first of three development cycles. Mobile applications, IoT devices and questionnaires will be used in the first of the three cycles. This is viable since we will use the infrastructure and connections of an existing stroke project on-site. The innovation in MAESTRO lays in the development of software solutions to monitor the rehabilitation of post-stroke patients remotely and passively using off-the-shelf hardware and gamification. The methods employed in MAESTRO, particularly deep learning, permit the automated classification of extremely complex data, allowing scientists to extract important information from data sets that would be unmanageable otherwise. MAESTRO is a unique scientific advance because it will provide doctors, patients and caregivers, group-specific levels of feedback. In addition, the algorithms specifically developed within the project can be the bases of novel developments with different goals, for example translation to clinical practice, or expansion to other neurodegenerative diseases.