Human-Centered Artificial Intelligence

The goal of this activity is to contribute to the creation of new methodological approaches for designing and developing Human-Centered Artificial Intelligence (HCAI) technologies. HCAI is an emerging research field that aims to design and develop intelligent systems, which are not only technologically advanced, useful, and effective, but also ethically aligned with universal human values and goals. Therefore, this particular research field focuses on creating Artificial Intelligence systems that can predict and adapt to human needs and preferences, while also being transparent, responsible, accountable and explainable, helping end users to easily understand their reasoning and results. The Laboratory's activities in this area include research and methodological proposals to explore how human-centered design can be pursued and applied at the core of Artificial Intelligence development, putting humans at the center of the development cycle. The results of the above research are used for the development of Artificial Intelligence technologies in various domains, such as security, industrial environments, big data visualization systems, etc.


Indicative Outcomes


Human-Centered Design framework for Artificial Intelligence (2021): This framework focuses on describing how the human-centered design (HCD) process can be revisited and expanded in an artificial intelligence (AI) context, proposing a methodological approach for putting the human in the loop. It explores how exactly the ‘human-in-the loop’ paradigm can be pursued and relevant core AI concepts and approaches.

Real-Time Adaptation of Context-Aware Intelligent User Interfaces, for Enhanced Situational Awareness (2021): A novel computational approach for the dynamic adaptation of User Interfaces (UIs) is proposed, which aims at enhancing the Situational Awareness (SA) of users by leveraging the current context and providing the most useful information, in an optimal and efficient manner. By combining Ontology modeling and reasoning with Combinatorial Optimization, the system decides what information to present, when to present it, where to visualize it in the display - and how , taking into consideration contextual factors as well as placement constraints.

Real-Time Stress Level Feedback from Raw ECG Signals for Personalised, Context-Aware Applications Using Lightweight Convolutional Neural Network Architectures (2021): Convolutional Neural Network architectures for the stress detection and 3-level (low, moderate, high) stress classification tasks, using ultra short-term raw ECG signals (3 s). One architecture is suitable for running in wearable edge-computing nodes, and the other is able to learn more complex features, having more trainable parameters. The evaluation demonstrated high accuracy both on the 3- and 2-level stress classification task using, superseding state-of-the-art in the field, reporting an accuracy of 83.55% and 98.77% respectively.

Methods and tools, based on no-reference Machine Learning (ML) approaches, for the real-time assessment of the Quality of Experience of interactive media (2021): An automatic QoE estimation framework, focusing on the case of video streaming. To achieve this objective, a novel architecture is introduced for encoding both visual and network-related information and validate the merits of the proposed scheme on an appropriately defined large-scale dataset. Experimental results indicate that considering all types of available information leads to the best performance in terms of prediction accuracy.