Chris POTTS - 25 septembre 2025 à 10h30 au PK-5115(201, ave President-Kennedy, 5e étage)
TITRE : Meaning in Large Language Models: Bridging Formal Semantics, Pragmatics, and Learned Representations
RÉSUMÉ
In its modern form, semantics (the study of the conventionalized aspects of linguistic meaning) is firmly rooted in symbolic logic. Such logics are also a cornerstone of pragmatics (the study of how people create meaning together in interaction). We can trace this methodological orientation to the roots of these fields in mathematical logic and the philosophy of language. This origin story has profoundly shaped both semantics and pragmatics at every level. How would these fields have looked had they instead been rooted in connectionism? They would have been radically different: the distinction between semantics and pragmatics would fall away, the range of relevant empirical phenomena would expand, and the theories themselves would have greater predictive force. This is not to say that there would be no role for symbolic logic in this hypothetical connectionist “semprag.” Large language models do learn solutions that reflect existing symbolic theories of meaning, and this is key to their success. This points to a future in which the fields of semantics and pragmatics embrace much more of what is happening in AI – without, however giving up their roots in symbolic logic.
BIOGRAPHIE
Christopher POTTS is Professor of Linguistics and, by courtesy, of Computer Science at Stanford, and a faculty member in the Stanford NLP Group and the Stanford AI Lab. His research group uses computational methods to explore topics in context-dependent language use, systematicity and compositionality, model interpretability, information retrieval, and foundation model programming. This research combines methods from linguistics, cognitive psychology, and computer science, in the service of both scientific discovery and technology development. Chris is also Co-Founder and Chief Scientist at Bigspin AI, a start-up focused on collaborative development of AI systems.
Roger LEVY - 18 septembre 2025 à 10h30 au PK-5115(201, ave President-Kennedy, 5e étage)
TITRE : Behavioral evaluation of language models as models of human sentence processing
RÉSUMÉ
Cette conférence examine comment les grands modèles de langage peuvent servir de modèles computationnels du traitement humain des phrases, en se concentrant sur les méthodes d'évaluation comportementale qui comparent les prédictions des modèles avec les données psycholinguistiques humaines. Je discuterai de travaux récents montrant que les mesures de probabilité directes des modèles de langage fournissent souvent de meilleures perspectives sur les connaissances linguistiques que les évaluations basées sur les invites. La conférence couvrira les considérations méthodologiques pour utiliser la théorie de la surprise et d'autres mesures théoriques de l'information pour valider les LLMs comme modèles cognitifs, examinant à la fois les promesses et les limites des modèles de langage neuraux actuels dans la capture des mécanismes de traitement des phrases humaines. Je présenterai des preuves sur la façon dont les mesures dérivées des modèles de difficulté de traitement s'alignent avec les données de temps de lecture humaines et discuterai des implications pour la science cognitive et le traitement du langage naturel.
This talk examines how large language models can serve as computational models of human sentence processing, focusing on behavioral evaluation methods that compare model predictions with human psycholinguistic data. I will discuss recent work showing that direct probability measurements from language models often provide better insights into linguistic knowledge than prompting-based evaluations. The talk will cover methodological considerations for using surprisal theory and other information-theoretic measures to validate LLMs as cognitive models, examining both the promises and limitations of current neural language models in capturing human sentence processing mechanisms. I will present evidence on how model-derived measures of processing difficulty align with human reading time data and discuss implications for both cognitive science and natural language processing.
BIOGRAPHIE
Roger LEVYest professeur de sciences du cerveau et cognitives au MIT, où il dirige le laboratoire de psycholinguistique computationnelle. Ses recherches portent sur des questions théoriques et appliquées dans le traitement et l'acquisition du langage naturel, étudiant comment la communication linguistique résout l'incertitude sur des signaux et significations potentiellement illimités. Il combine la modélisation computationnelle, l'expérimentation psycholinguistique et l'analyse de grands ensembles de données linguistiques naturelles pour comprendre les fondements cognitifs du traitement du langage et aider à concevoir de meilleurs systèmes de traitement automatique du langage. Avant de rejoindre le MIT en 2016, il a fondé un laboratoire de psycholinguistique computationnelle à UC San Diego. Il est actuellement président de la Cognitive Science Society (2024--2025).
Roger LEVY is Professor of Brain and Cognitive Sciences at MIT, where he heads the Computational Psycholinguistics Laboratory. His research focuses on theoretical and applied questions in the processing and acquisition of natural language, investigating how linguistic communication resolves uncertainty over potentially unbounded signals and meanings. He combines computational modeling, psycholinguistic experimentation, and analysis of large naturalistic language datasets to understand cognitive underpinnings of language processing and to help design better machine language processing systems. Before joining MIT in 2016, he founded a Computational Psycholinguistics Laboratory at UC San Diego. He currently serves as President of the Cognitive Science Society (2024--2025).
Megan PETERS - 11 septembre 2025 à 10h30 au PK-5115(201, ave President-Kennedy, 5e étage)
TITRE : Confidence, Metacognition, and the "Hard Problem" of Consciousness
RÉSUMÉ
Conscious feeling---the "hard problem"---remains a central challenge for cognitive science. This talk explores whether computational models of metacognitive confidence and uncertainty, grounded in introspective psychophysics and higher-order representations, can illuminate phenomenal experience. I will discuss recent work showing how the brain encodes uncertainty, how confidence can be modeled computationally, and how subjective reports can be formalized. These approaches point toward canonical computations linking perception, decision-making, and metacognition, offering a possible path to studying subjective experience---and perhaps a way to rethink the hard problem itself.
BIOGRAPHIE
Megan PETERS is Associate Professor in the Department of Cognitive Sciences at UC Irvine, with an affiliation in Logic and Philosophy of Science, and is a Fellow of the CIFAR Brain, Mind & Consciousness program. Her research investigates how the brain represents and uses uncertainty, how these computations support metacognitive evaluations of perceptual decisions, and how they may relate to subjective experience in both humans and artificial systems. She uses human neuroimaging (fMRI, EEG), computational modeling, machine learning, and psychophysics to study these questions. She is also co-founder and President of Neuromatch, which develops globally accessible education programs in computational neuroscience and related fields.
Roberto NAVIGLI - 10 avril 2025 à 10h30 au PK-5115(201, ave President-Kennedy, 5e étage)
TITRE : Can Large Language Models Prove Their Understanding of Language?
RÉSUMÉ
I will discuss some developments in multilingual lexical semantics and word sense disambiguation (WSD). BabelNet (Navigli & Ponzetto, 2012) introduced a large-scale, automatically constructed multilingual semantic network, integrating structured and unstructured lexical resources to support cross-lingual applications. A 2009 survey provided a comprehensive analysis of WSD methodologies, highlighting the challenges of ambiguity resolution and the evolution of knowledge-based and statistical approaches. A more recent survey (Bevilacqua et al., 2021) tracks developments in WSD, emphasizing neural architectures and data-driven improvements. These works have helped shape the understanding of semantic representation and disambiguation in Natural Language Processing.
BIOGRAPHIE
Roberto NAVIGLI, is Professor in the Department of Computer, Control and Management Engineering at Sapienza University of Rome, where he leads the Sapienza Natural Language Processing (NLP) Group. His research focuses on multilingual NLP, computational semantics, and knowledge representation. He developed BabelNet, a multilingual lexical-semantic knowledge graph that integrates resources like WordNet, Wikipedia, and Wiktionary and has contributed to word sense disambiguation, creating large-scale, automatically extracted training sets. In semantic role labeling, he has highlighted the need for improved models to handle diverse non-verb predicate types, such as nouns and adjectives, and he has contributed to multilingual semantic parsing techniques for creating language-independent semantic representations.
Steven T. PIANTADOSI - 3 avril 2025 à 10h30 au PK-5115(201, ave President-Kennedy, 5e étage)
TITRE : Rules vs. neurons and what may be next
RÉSUMÉ
I will discuss the relationship between large language models and Chomskyan theories of linguistics in the context of the broader debate between rule-based and neural approaches to cognitive modeling. While language models provide a working implementation that surpasses symbolic theories in many respects, I will also present work based in early computer science that seeks to formalize what latent structures must be present in a system in order to generate its observed behavior. This approach is the topic of a forthcoming open textbook, and the approach holds promise for understanding if any grammar-like structures are necessarily present in, for instance, statistical language models. This line of work also points to ways we can rigorously connect neuroscience to behavior.
BIOGRAPHIE
Steven T. PIANTADOSI, Professor in the Psychology Department and Helen Wills Neuroscience Institute, University of California, Berkeley, leads the Computation and Language Lab (CoLaLa). His computational and behavioral research is on the learning of language and concepts, the evolution of human-like cognition, and how ambiguity can serve communicative functions. In his critique of Noam Chomsky’s theory of Universal Grammar, Piantadosi argues that the success of large language models (LLMs) challenges most assumptions of standard linguistic theories.
RÉFÉRENCES:
Piantadosi, S. T., Muller, D. C., Rule, J. S., Kaushik, K., Gorenstein, M., Leib, E. R., & Sanford, E. (2024). Why concepts are (probably) vectors. Trends in Cognitive Sciences, 28(9), 844-856.
Chirag SHAH - 27 mars 2025 à 10h30 au PK-5115(201, ave President-Kennedy, 5e étage)
TITRE : Optimizing LLM Prompts for Scientific Use
RÉSUMÉ
As large language models (LLMs) increasingly permeate scientific research, their use for generating or analyzing data often relies on ad-hoc decisions, raising concerns about transparency, objectivity, and rigor. This talk introduces a methodology inspired by qualitative codebook construction to systematize prompt engineering. By integrating humans in the loop and a multi-phase verification process, this approach enhances replicability and trustworthiness in using LLMs for data analysis. Practical examples will illustrate how rigorous labeling, deliberation, and documentation can reduce subjectivity and ensure more robust and generalizable research outcomes.
BIOGRAPHIE
Chirag Shah is Professor in the Information School at the University of Washington, where he conducts research at the intersection of information retrieval, human-computer interaction, and artificial intelligence. His recent work focuses on prompt engineering for optimizing interactions with large language models, exploring both theoretical underpinnings and practical applications. Dr. Shah has authored numerous publications and is actively involved in advancing the understanding of how humans and AI systems can collaborate more effectively.
Jules WHITE - 20 mars 2025 à 10h30 au PK-5115(201, ave President-Kennedy, 5e étage)
TITRE : Generative AI will Reshape Computing and Innovation
RÉSUMÉ
Generative AI introduces a new computing paradigm, serving as an interface that translates human goals into computational actions. This talk examines three core aspects: generating computation from prompts, optimizing code, and coordinating API integration. A key shift is the emergence of prompt engineering, which enables users to specify complex tasks in natural language, differing from traditional programming. These new abstractions redefine how computation is conceived and executed. Generative AI also reshapes software development, empowering domain experts while computer scientists focus on scalable frameworks. This shift fosters collaboration between human creativity and machine intelligence, expanding computational accessibility.
BIOGRAPHIE
Jules WHITE, Professor of Computer Science at Vanderbilt University and Senior Advisor to the Chancellor for Generative AI, directs Vanderbilt’s Initiative on the Future of Learning & Generative AI. His research spans cybersecurity, mobile/cloud computing, and AI, with over 170 publications and multiple Best Paper Awards. He is a National Science Foundation CAREER Award recipient. He created one of the first online classes for Prompt Engineering.
Présentation publique – Projet de thèse de la personne étudiante Safwen NAIMI
Date : 18 mars 2025 Heure : 9h00 Lien zoom: https://uqam.zoom.us/j/89678175580(il est important d'être présent à l'heure et d'inscrire votre nom en entier lors de la connexion) Programme : Doctorat en informatique cognitive, UQAM
TITRE :Vidéosurveillance intelligente pour la détection automatique des comportements suicidaires
RÉSUMÉ : Le suicide est une problématique complexe et un défi majeur de santé publique à l’échelle mondiale, représentant une cause significative de mortalité prématurée. Chaque année, des milliers de personnes se retrouvent en situation de crise dans des espaces publics, comme les stations de métro, où la détection proactive des comportements à risque pourrait prévenir des tragédies. Les approches traditionnelles de vidéosurveillance reposent sur l’intervention humaine, et ont de ce fait des limites importantes, telles que la fatigue des opérateurs et la difficulté d’interpréter des signaux subtils dans des environnements encombrés et très fréquentés. Ces défis rendent nécessaire le développement de solutions innovantes intégrant des outils de vision par ordinateur et d’intelligence artificielle. Cette proposition de recherche propose une approche interdisciplinaire pour détecter et prévenir les comportements pré-suicidaires en s’appuyant sur des technologies de pointe. Le premier volet de notre projet inclut la création d’une base de données exclusive basée sur des vidéos de surveillance, ainsi que l’analyse approfondie des comportements associés aux risques suicidaires. Cette base de données servira de fondement au développement d’algorithmes avancés pour la détection des comportements pré-suicidaires. Les travaux du deuxième volet de notre projet s’inscrivent dans le domaine de la vidéosurveillance intelligente. Trois approches principales seront étudiées : (1) Le développement de méthodes de reconnaissance d’action humaine (RAH). Ces méthodes seront basées sur des données squelettiques offrant une robustesse dans des environnements complexes en minimisant les effets des variations d'éclairage, des occlusions et des arrière-plans encombrés, tout en mettant en évidence les relations spatio-temporelles des mouvements humains pour une identification précise des comportements à risque. (2) Le développement d’algorithmes de segmentation sémantique pour identifier les zones critiques des stations, telles que les quais et les lignes jaunes, permettant une intégration contextuelle dans l’analyse des trajectoires des passagers. (3) L’analyse des trajectoires des passagers, visant à détecter les mouvements inhabituels ou répétitifs, tels que le va-et-vient prolongé ou le stationnement prolongée près de la ligne jaune, qui pourraient indiquer des comportements à risque nécessitant une intervention proactive. Outre ses contributions scientifiques, ce projet vise à réduire les pertes humaines dans les stations de métro tout en ouvrant de nouvelles perspectives pour les systèmes de surveillance intelligente. Nous contribuons à renforcer la capacité des systèmes de surveillance à intervenir de manière proactive.
Jury : Roger Villemaire (président du jury) Johanne Saint-Charles Mohamed Bouguessa Wassim Bouachir (direction de recherche) Brian L. Mishara et Guillaume-Alexandre Bilodeau (codirection de recherche)
Jean-Claude MARTIN - 13 mars 2025 à 10h30 au PK-5115(201, ave President-Kennedy, 5e étage)
TITRE :Interaction Sociale Motivationnelle en Psychologie et en Interaction Humain-Machine
RÉSUMÉ
La motivation et l’interaction sociale sont deux concepts fondamentaux en psychologie qui sont rarement considérés ensemble. Je présenterai des recherches interdisciplinaires visant à concevoir des interactions homme-machine qui, soit motivent les utilisateurs à interagir avec autrui (par exemple, des interactions tangibles et virtuelles pour les utilisateurs autistes), soit soutiennent des interactions destinées à motiver les utilisateurs (par exemple, des technologies mobiles motivationnelles pour adopter de meilleurs modes de vie). J’expliquerai les différences individuelles que nous avons observées et comment nous nous appuyons sur ces différences pour mieux comprendre les personnes et leur offrir des interactions personnalisées, motivationnelles et sociales.
BIOGRAPHIE
Jean-Claude MARTIN est professeur à l’Université Paris-Saclay, en France, dans le domaine de l’interaction homme-machine. Il dirige l’équipe de recherche « Cognition, Perception et Usages » au Laboratoire Interdisciplinaire des Sciences du Numérique. Ses recherches portent sur l’adaptation et la combinaison des théories psychologiques avec des approches de conception centrée sur l’utilisateur afin de concevoir des interactions homme-machine pour la formation aux compétences sociales et la motivation à l’activité physique.
Florian Debackere, Céline Clavel, Alexandra Roren, Viet-Thi Tran, Yosra Messai, François Rannou, Christelle Nguyen, and Jean-Caude Martin. 2023. Design framework for the development of tailored behavior change technologies. In Adjunct Proceedings of the 31st ACM Conference on User Modeling, Adaptation and Personalization (UMAP '23 Adjunct). 140–146.
Mohammed (Ehsan) Hoque, Matthieu Courgeon, Jean-Claude Martin, Bilge Mutlu, and Rosalind W. Picard. 2013. MACH: my automated conversation coach. In Proceedings of the 2013 ACM international joint conference on Pervasive and ubiquitous computing (UbiComp '13). 697–706.
Angelo CANGELOSI - 6 mars 2025 à 10h30 au PK-5115(201, ave President-Kennedy, 5e étage)
TITRE :The importance of starting small: Developmental robotics for language grounding
RÉSUMÉ
Cognitive robotics aims to develop robots capable of human-like learning, interaction, and behavior by grounding abstract concepts in sensorimotor experiences and social interactions. This talk explores how principles like “starting small” and “super-embodiment” can address the limitations of AI tools, such as large language models (LLMs), which rely heavily on large datasets and static learning protocols. By integrating incremental, multimodal learning and redefining embodiment to encompass physical, mental, and social processes, we can enable robots to better understand and utilize abstract concepts. These advancements hold promise for applications in caretaking, education, and beyond, while advancing the intersection of AI, grounded intelligence, and human development.
BIOGRAPHIE
Angelo Cangelosi, Professor of Machine Learning and Robotics at the University of Manchester and co-directs the Manchester Centre for Robotics and AI. His research focuses on cognitive and developmental robotics, neural networks, language grounding, human-robot interaction, and robot companions for health and social care. He is the author of Developmental Robotics: From Babies to Robots (MIT Press, 2015) and Cognitive Robotics (MIT Press, 2022), co-edited with Minoru Asada and Editor-in-Chief of Interaction Studies and IET Cognitive Computation and Systems.
Marchetti, A., Di Dio, C., Cangelosi, A., Manzi, F., & Massaro, D. (2023). Developing ChatGPT’s theory of mind. Frontiers in Robotics and AI, 10, 1189525.
