We will explore various ways in which agents—both human and artificial—perceive and make sense of the world around them. A key part of this exploration involves examining how this understanding can be represented through formal ontologies: structured frameworks that allow us to define and organise knowledge in a precise, human-based and machine-readable way. By analysing these processes, we will develop the conceptual tools needed to unpack complex ideas such as (agent-centric) environment, context, and collaboration. This foundation will help us better understand how different types of agents interact—not only with their surroundings but also with each other—paving the way for more effective human-AI partnerships.
Bio
Stefano Borgo is Director of Research and Head of the CNR-ISTC centre in Trento. He studied mathematics (Padua), logic (Bloomington), and informatics (Bozen-Bolzano). His research focuses on applied ontology and knowledge representation with emphasis on interaction, interoperability and model construction. Applications areas span disciplines like engineering, robotics, and urban studies. He co-authored the DOLCE ontology and published about 200 papers in conferences and journals. He is member of the Editorial Board of the Applied Ontology and of the Semantic Web journals, of the Advisory Board of the International Association on Ontology and its Applications (IAOA) and is technical expert for ISO and IEEE working groups, including the Italian UNI. He is well aware that this bio is boring and happy that it stops here.
In recent years, there has been a large wave of research moving towards the usage of large-scale, naturalistic data over small-scale hypothesis driven experimental methods to study human cognition and biology. Large Language Models (LLMs) have already demonstrated a number of effective uses in being able to explain, for example, human language processing in real time or conceptualizing how information flows between natural conversations. LLMs may therefore be incredibly powerful tools for condensing and explaining large, noisy amounts of real-world data. This talk will focus on how embeddings learned by LLMs have been used in these research settings, with an emphasis on how information contained in word-level embeddings may tell us something very important about the distribution of structural, recurrent information learned during the training process. I will specifically focus on two aspects of the embedding space. The first will dive into how LLMs handle frequency in the input and how frequency may be a starting point for understanding how structural patterns are encoded by the model. The second aspect will investigate how word-level relationships change across layers, and what this might mean for current interpretations of word embedding to human language relationships. The overall scope of my work is to provide researchers in non-machine learning fields a better working understanding of how readily available, light-weight models (like GPT-2) may be applicable for extracting meaningful interpretations from naturalistic data. Therefore, the work presented in this talk will focus on these smaller scale LLMs, with the hope to generate discussion on the implications of this work to other LLM usage cases and applications.
Bio
I am currently a post-doctoral researcher working at the Laboratory for Autism and Neurodevelopmental Disorders in Rovereto, Italy. I have a PhD in Theoretical and Applied Linguistics from the University of Cambridge (UK), a combined MSc in Education Neuroscience from UCL & Birkbeck College London (UK) and a BA in Linguistics and Phonetics from the University of Leeds (UK). My interests can be boiled down to looking for novel methodological approaches across disciplines that allow us to improve our ability to draw concrete inferences from naturalistic, ‘uncontrollable’ data. During my PhD I primarily worked on understanding multilingual diversity in autism, using previous applications of information theory on multilingual data, and my most recent work has been in investigating interpersonal similarity in naturalistic gaze data and LLM interpretability for naturalistic discourse. In a few months, I will be continuing my research journey as a post-doc at the University of Cambridge as part of the Molecules to Health Research Program working on methodological innovations in large scale clinical imaging and genomics data.
Link: https://land.iit.it/people-details/-/people/sarah-crockford