Special Issue: Selected papers from the AIxIA 2023 Workshops

The 2023 edition of the AIxIA Conference, held in Rome, Italy, was a collaborative effort organized by the three federated Roman universities (Roma Tre University, Sapienza University of Rome, and University of Rome, Tor Vergata) along with the Consiglio Nazionale delle Ricerche (CNR), celebrating its 100th anniversary. This conference hosted a series of workshops focusing on specific and trending topics in Artificial Intelligence, complementing the main conference program. The workshops were designed to foster collaboration and cross-disciplinary exchange, attracting contributions from a diverse range of disciplines.

The AIxIA 2023 Conference featured 19 workshops, organized by 77 dedicated workshop organizers, which received 248 submissions. Out of these, 16 proceedings were published, primarily through CEUR. Notably, three of these workshops shared a single proceedings volume. The workshops also included 23 invited speakers, four panels, and four parallel sessions spanning four days.

This Special Issue is composed of the extended versions of eleven papers initially presented at these workshops. Each contribution was selected by the workshop organizers and underwent rigorous review by at least two members of the Italian AI community. The diversity of topics reflects the broad and dynamic nature of the AI community in Italy. Below, we briefly introduce the selected papers, summarizing their contributions and highlighting their original versions.

The paper by Carnevale et al. titled “A Human-Centred Approach to Symbiotic AI: Ethical and Conceptual Foundations”, was originally presented at the BEWARE workshop. 1 This extended work, based on their earlier publication [5], argues for a constructivist approach to symbiotic artificial intelligence (SAI). The authors emphasize the importance of restoring human-centeredness in SAI governance, advocating for a flexible, contextual, and relational framework that integrates social and technological factors. They propose a new methodology for evaluating SAI systems, emphasizing human responsibility and ethical considerations.

Building on the theme of enhancing the interpretability and ethical aspects of AI, the paper by Sabbatini and Calegari, “Untying Black Boxes with Clustering-Based Symbolic Knowledge Extraction”, extends their initial work [9] presented as well at the BEWARE workshop. The authors introduce the CReEPy algorithm, which uses explainable clustering to generate human-interpretable Prolog rules from black-box models. They also present the CRASH algorithm for hyper-parameter tuning, and evaluate their methods in real-world applications to enhance the transparency and trustworthiness of machine learning models.

Shifting focus to the creative applications of AI, the paper “Creative Influence Prediction Using Graph Theory”, by Alfieri et al., extends their previous work [2] presented at the CREAI workshop. 2 The authors investigate the social network influence on an artist’s creative process using graph theory. They develop an explainable method to predict artistic influence based on social relationships, demonstrating its effectiveness using a dataset of Jazz musicians. The approach enriches existing knowledge graphs and offers insights into the socio-cognitive aspects of creativity.

Transitioning to the social impact of AI, Olearo et al. in their paper “Facing Multidimensional Poverty in Older Adults: An Artificial Intelligence Approach that Reveals the Variable Relevance” extend their previous study [8], presented at the AIxAS workshop. 3 The authors propose an AI model to predict multidimensional poverty among older adults, leveraging expert knowledge to label data and applying XGBoost to identify key variables affecting poverty. Their flexible approach can be adapted to various scenarios, providing valuable insights for social policy and intervention strategies.

Following the theme of societal applications, the paper “Fostering Artificial Intelligence-based Supports for Informal Caregivers: A Systematic Review of the Literature” by Milella and Bandini, extends their prior study [7], also presented at the AIxAS workshop. This work examines AI-enabled technologies aiding informal caregivers, who provide most long-term care in Europe. The systematic review analyzes recent advancements in AI and assistive technologies, identifying their benefits and gaps. Findings show AI significantly enhances ambient assisted living and care coordination between formal and informal caregivers, with potential for advanced telehealth. The paper underscores the need for more research on assistive technologies like robots and mHealth apps and suggests targeted AI solutions to better address caregiver needs.

Moving to biomedical applications, Gabardi et al.’s paper, “DL-based Multi-Artifact EEG Denoising Exploiting Spectral Information” extends their earlier research [6], presented at the AIxHMI workshop. 4 The authors develop a deep learning model to denoise EEG signals, utilizing spectral features to distinguish neural activity from artifacts. Their method, evaluated on the EEGdenoiseNet benchmark, shows superior performance in removing noise, making it applicable for both medical and recreational brain-computer interface applications.

Delving into educational applications, the paper “AI-Assisted Board-Game Based Learning. An Exploratory Study to Assess Reliability and Accuracy of Chat GPT in Game Evaluation” by Tinterri et al., extends their previous study [10] presented at the AIxEDU workshop. 5 This research investigates the use of Chat GPT for selecting educational board games, comparing its evaluations with those of human experts. The study highlights the potential and limitations of AI in instructional design, providing insights into enhancing board game-based learning.

Continuing with advancements in higher education, the paper “Large Language Models for Sustainable Assessment and Feedback in Higher Education: Towards a Pedagogical and Technological Framework” by Agostini and Picasso, extend their prior study [1] also presented at the AIxEDU workshop. This work addresses the challenges in modernizing assessment and feedback practices in higher education. Despite recommendations from the Bologna Process to adopt alternative assessment methods, traditional summative assessments remain prevalent. This study explores the potential of Large Language Models (LLMs) to provide sustainable solutions by processing and summarizing large amounts of text, offering feedback, and assisting instructors with assessments. The authors propose a pedagogical and technological framework to integrate LLMs in educational assessment, aiming to reduce the reliance on traditional methods and enhance the learning experience through innovative, technology-driven practices.

Moving to more technical applications, the paper by Bertagnon and Gavanelli, “Geometric Reasoning on the Euclidean Traveling Salesperson Problem in Answer Set Programming”, extends their previous work [4] presented at the RCRA workshop. 6 The authors explore the applicability of geometric reasoning in solving the Euclidean Traveling Salesperson Problem (TSP) using Answer Set Programming (ASP). They compare the performance of ASP and Constraint Logic Programming (CLP) solvers, demonstrating the benefits of incorporating geometric information to enhance problem-solving efficiency.

Continuing with theoretical advancements, the work by Alviano et al., titled “Weighted Knowledge Bases with Typicality and Defeasible Reasoning in a Gradual Argumentation Semantics”, builds on their earlier research [3], also presented at the RCRA workshop. The paper explores the relationship between weighted knowledge bases and weighted argumentation graphs, proposing methods for defeasible reasoning under a φ-coherent semantics. The authors provide an ASP implementation and present experimental results, highlighting the effectiveness of their approach.

Further expanding on advanced AI techniques, the paper “Conditional Computation in Neural Networks: Principles and Research Trends” by Scardapane et al. expands their presentation at the MLDM workshop, 6 exploring the emerging field of conditional computation in neural networks. This involves dynamically activating or deactivating parts of a neural network based on the input. The authors provide a unified formalism for these techniques and discuss three key implementations: mixture-of-experts (MoEs) networks, token selection mechanisms, and early-exit neural networks. The paper highlights the benefits of these approaches in efficiency, explainability, and transfer learning, with applications in areas such as automated scientific discovery and semantic communication.

These selected papers represent the innovative and diverse research conducted within the Italian AI community, showcasing the potential of AI in addressing complex problems across various domains. Reflecting the interdisciplinary nature of the field, the contributions explore ethical frameworks, enhance model transparency, apply AI to creative and social issues, and advance technical methodologies. This special issue provides valuable insights into the latest advancements in Artificial Intelligence and highlights the dynamic landscape of AI research. The interdisciplinary approach and innovative solutions presented offer a comprehensive view of the current state of AI and its applications, both within Italy and beyond.