Research Pillars
Human-Centric AI
The Human-Centric AI pillar focuses on the development and evaluation of methods and algorithms to improve human-AI interaction in decision support, user experience, explainability, and interpretability, while also addressing societal and ethical impacts. It aims to enhance transparency and human oversight in AI-based decision-making.
Project: Constraints-Abiding Explainable Reinforcement Learning
Third parties involved: George Vouros, George Papadopoulos, Piyabhum Chaysri – University of Piraeus
This research addresses the limited attention given to reinforcement learning (RL) methods that provide transparency regarding operational constraints—domain-specific requirements that must be maintained during operation. Transparency about these constraints is essential, since automation must preserve humans’ awareness of them and their ability to inspect whether they are upheld, thereby assuring operational safety. The project aims to devise an inherently interpretable safe RL method that offers clear visibility into operational constraints. It studies symbolic representations for constrained RL policy models and designs, implements, and validates an interpretable safe RL–based solution in constrained settings. The project delivers symbolic models that enable an inherently interpretable safe RL method under operational constraints.
Project: Explainability-driven decision support systems for small clinical datasets
Third parties involved: Prof. Dr. Darian M. Onchiș, Dr. Codruța Istin – West University of Timisoara
The project targets two ENFIELD challenges: human-oriented explanations for AI-assisted medical diagnostics and metrics for evaluating explainability/interpretability. It focuses on post-hoc, model-agnostic surrogate methods to make deep learning “black boxes” clinically transparent. Aim 1 is to use explanation insights as feedback to improve models, with emphasis on small medical datasets. Aim 2 is to address instability in techniques like LIME by proposing complementary metrics for stability, clinical usefulness, and interpretability. The outcome is actionable, human-centered explanations underpinned by rigorous evaluation.
Project: Multimodal Analysis of Sleep Dynamics with Explainable Transformers
Third parties involved: Ander Cejudo – Vicomtech
This project leverages transformer-based models with explainable AI to analyze sleep dynamics and physiological signals related to schizophrenia relapse using wearable devices. Expected outcomes include pre-trained models for sleep analysis, new explainability techniques, enhanced libraries for signal analysis with deep learning, and improved relapse prediction to advance AI-driven healthcare in Europe.
Project: XAI for Advancing Hydrological Forecasting and Process Understanding
Third parties involved: Rafael Francisco – Higher Technical Institute of the University of Lisbon
The project applies Explainable AI to hydrological forecasting, addressing the growing complexity of water management. It uses the Temporal Fusion Transformer, a deep learning model for time-series data, to improve prediction accuracy and identify the main drivers of hydrological processes through attention mechanisms. By making forecasts more transparent and interpretable, the project builds trust in model outputs and supports better-informed water management decisions.
Project: Explainable AI for Solar Cell Design via Electrochemical Impedance Spectra
Third parties involved: Ricardo Grau-Crespo – University of Reading
The project develops a fast and interpretable machine learning framework for analysing EIS data from solar cell materials. Training models on synthetic EIS datasets enables prediction of key physical parameters such as carrier mobility, ionic conductivity, recombination, and interfacial transport properties. Explainable AI methods identify the spectral features behind each prediction, improving trust and physical insight. Validated through experimental case studies, the framework accelerates EIS interpretation from hours or days to seconds and supports more efficient solar cell materials optimisation.
Project: Uncertainty-aware feature attribution for temporal predictive models
Third parties involved: Niklas Koenen – University of Bremen
The project investigates post-hoc explanation methods for probabilistic models with multivariate temporal outputs. It introduces a hierarchy of entropy-based Shapley games that decompose predictive uncertainty across three levels — marginal, sequential, and joint — enabling attribution of both per-component uncertainty and cross-component dependence structure. By grounding attribution in information-theoretic quantities, the framework identifies which input features drive forecast uncertainty and how, going beyond standard component-wise methods that miss dependence effects.