Invited Speakers

Albert Gatt

Albert Gatt is professor at the Department of Information and Computing Sciences, where he holds the chair in Natural Language Generation and leads the NLP group. His research focuses on the generation of natural language and on the interface between vision and language. In his research, he relies on machine learning methods such as neural networks, but also on experimental psycholinguistic methods. Here are some of the topics he explores:

Data-to-text generation, that is, the automatic summarisation of non-linguistic information, in forms that are understandable by people;
The vision-language interface, especially image captioning and grounded inference in multimodal, neural models;
The production and generation of referring expressions, especially the question of what to include in object descriptions in visual scenes;
Evaluation, that is, how we can determine the quality of the outputs of NLP models (especially generation models).

Apart from these, another long-standing interest is the development of tools and resources for under-resourced languages.

Title: What models "see" and what they "say" – Explaining multimodal models' visual grounding abilities

Abstract: The "grounding problem" is one of the foundational philosophical issues in Artificial Intelligence and NLP: how can an artificial system learn to link its symbolic representations to perception and experience? Multimodal deep neural networks suggest a possible way to address this, by training large-scale models on data from different modalities. Vision-Language Models are an important example of this endeavour, aiming to link linguistic expressions to visual data in images or video. But what is the quality of the representations they learn? Are their grounding abilities the same for all classes of linguistic expressions?

In this talk, I will synthesise research carried out over the past few years on probing the grounding abilities of multimodal Vision-Language models and developing carefully designed benchmarks for this purpose. Complementary to this research program is an exploration of model interpretability and explainability techniques which addresses the specific challenges of explaining model performance on generative and understanding tasks with multimodal input.

Konstantia Zarkogianni

Konstantia Zarkogianni received a MEng in Electrical and Computer Engineering (2003) from the Aristotle University of Thessaloniki, a MSc Degree in Electronic and Computer Engineering (2005) from the Technical University of Crete, and a PhD degree (2011) from the National Technical University of Athens (NTUA), Greece. In October 2017, she was appointed as a permanent laboratory teaching staff member at the School of Electrical and Computer Engineering of the NTUA. In 2023, she was appointed as Associate Professor of Human-Centered AI at the Department of Advanced Computing Sciences, Maastricht University. Her research is mainly focused on AI (e.g., ML, XAI, knowledge bases & reasoning, and recommender systems) towards the development of innovative decision support systems and intelligent user interfaces. She has authored or co-authored 17 papers in refereed international journals, 3 invited editorials and reviews, one chapter in book, and more than 30 papers in international conference proceedings. She has participated as research associate and PI in greek (ENDORSE, smarty4covid) and EU funded projects (FP7 - MOSAIC, VOXReality). She is a member of the Institute of Electrical and Electronics Engineers (ΙΕΕΕ).

Title: Harnessing Explainable AI for Informed Decision Making in Healthcare

Abstract: In this talk, cutting-edge technologies that enhance human oversight in decision-making processes will be presented. Emphasis will be placed on deep learning, explainable AI techniques, and intelligent recommender systems that elevate user interfaces. Methods for generating explainable predictions, creating explanation rules, formulating counterfactual explanations, and delivering personalized messages that provide valuable insights for patients and healthcare professionals will be analysed. The integration of multimodal data sources will also be discussed, highlighting how combining various types of data can enhance the accuracy and reliability of AI-driven insights. Additionally, the critical need to identify and mitigate data biases as a fundamental step towards achieving responsible AI will be underscored. Addressing biases in data inputs is crucial to ensure that AI systems deliver fair and accurate insights, especially in the sensitive domain of healthcare. Specific examples in chronic and infectious disease management will be presented to demonstrate how these technologies can effectively improve patient outcomes and optimize healthcare delivery.