Séminaire du DIC

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. 

RÉFÉRENCES:

Bevilacqua, M., Pasini, T., Raganato, A., & Navigli, R. (2021). Recent trends in word sense disambiguation: A survey. International Joint Conference on Artificial Intelligence (pp. 4330-4338).

Navigli, R., & Ponzetto, S. P. (2012). BabelNet: The automatic construction, evaluation and application of a wide-coverage multilingual semantic network. Artificial Intelligence, 193, 217-250.

Navigli, R. (2009). Word sense disambiguation: A survey. ACM computing surveys (CSUR), 41(2), 1-69.

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.

Piantadosi, S. T. (2023). Modern language models refute Chomsky’s approach to language. From fieldwork to linguistic theory: A tribute to Dan Everett, 353-414.

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.

RÉFÉRENCES:

Sahoo, Pranab, et al. (2024)  "A systematic survey of prompt engineering in large language models: Techniques and applications." arXiv preprint arXiv:2402.07927 (2024).

Shah, C. (2024). From Prompt Engineering to Prompt Science With Human in the Loop. arXiv preprint arXiv:2401.04122.

White, R. W., & Shah, C. (2025). Information Access in the Era of Generative AI. Springer

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.

RÉFÉRENCES:

White, J., Hays, S., Fu, Q., Spencer-Smith, J., & Schmidt, D. C. (2024). Chatgpt prompt patterns for improving code quality, refactoring, requirements elicitation, and software design. In Generative AI for Effective Software Development (pp. 71-108). Cham: Springer Nature Switzerland.

White, J., Fu, Q., Hays, S., Sandborn, M., Olea, C., Gilbert, H., ... & Schmidt, D. C. (2023). A prompt pattern catalog to enhance prompt engineering with chatgpt. arXiv preprint arXiv:2302.11382.

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.

RÉFÉRENCES:

Benamara, A., Martin, J.-C., Prigent, E., Ravenet, B. (2023) Evaluating a Model of Pathological Affect based on Pedagogical Situations for a Virtual Patient. 23rd  ACM International Conference on Intelligent Virtual Agents (IVA’2023).

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.

Hamet Bagnou J, Prigent E, Martin J-C and Clavel C (2022) Adaptation and validation of two annotation scales for assessing social skills in a corpus of multimodal collaborative interactions. Frontiers in Psychology.

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.

Rei, D., Clavel, C., Martin, J.-C., Ravenet, B. (2024) Adapting goals and motivational messages on smartphones for motivation to walk. Journal Smart Health 32, June 2024.

 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.

RÉFÉRENCES:

Asada, M., & Cangelosi, A. (2024). Reevaluating development and embodiment in robotics. Device, 2(11).

Elman, J. L. (1993). Learning and development in neural networks: The importance of starting small. Cognition, 48(1), 71-99.

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.

Xie, H., Maharjan, R. S., Tavella, F., & Cangelosi, A. (2024). From Concrete to Abstract: A Multimodal Generative Approach to Abstract Concept Learning. arXiv preprint arXiv:2410.02365.

Bing LIU - 27 février 2025 à 10h30 au PK-5115 (201, ave President-Kennedy, 5e étage)

TITRE : Achieving Upper Bound Accuracy in Continual Learning

RÉSUMÉ 

The ability to continuously learn and accumulate knowledge over a lifetime is a hallmark of human intelligence. However, this essential capability is missing in current machine learning paradigms. This talk explores continual learning in machine learning, with a focus on the challenges of catastrophic forgetting and inter-task class separation. These issues have prevented existing methods from reaching the theoretical upper-bound performance, often with a significant gap. Our recent work demonstrates that achieving this upper bound is indeed possible, offering intriguing insights into both cognition and the foundations of AI.

BIOGRAPHIE

Bing LIU is a Distinguished Professor and Peter L. and Deborah K. Wexler Professor of Computing at the University of Illinois Chicago. He earned his Ph.D. from the University of Edinburgh. His current research interests include continual or lifelong learning, continual learning dialogue systems, sentiment analysis, machine learning, and natural language processing. He is the author of several books on these topics and has also received multiple Test-of-Time awards for his research papers. He is a Fellow of ACM, AAAI, and IEEE.

RÉFÉRENCES:

Chen, Z., & Liu, B. (2018). Lifelong machine learning. Morgan & Claypool Publishers.

Ke, Z., Shao, Y., Lin, H., Konishi, T., Kim, G., & Bing Liu. Continual Pre-training of Language Models.  ICLR-2023.

Kim. G., Xiao, C., Konishi, T., Ke, Z., & Liu, B. A Theoretical Study on Solving Continual Learning. NeurIPS-2022.

Liu, B. (2023). Grounding for Artificial Intelligence. arXiv preprint arXiv:2312.09532.

Momeni, S., Mazumder, S., & Liu, B. Continual Learning Using a Kernel-Based Method Over Foundation Models, AAAI-2025, 2025.

Herbert ROITBLAT - 20 février 2025 à 10h30 au PK-5115 (201, ave President-Kennedy, 5e étage)

TITRE : Theories of Artificial Intelligence

RÉSUMÉ 

GenAI models are computationally complex, but conceptually simple.  They are trained to fill in the blanks.  Model semantics is limited to word distribution patterns (Harris, 1956), yet many claim that GenAI models are capable of deep cognitive processes (such as reasoning and understanding). These assertions imply that cognitive processes can spontaneously emerge from these behavioral patterns—that a theory of cognition can be constructed at the purely behavioral level of word use patterns.  We have seen that movie before, but Chomsky (1959/1967) and a good deal of research with humans and animals, peaking in the 1980s, have demonstrated that a purely behavioral theory of cognition is not viable.  Those same research methods could be applied to the analysis of the latest forms of artificial intelligence, but their relevance is rarely recognized.  Instead, much of what passes for theoretical analysis of GenAI models is based on the logical fallacy of “affirming the consequent.” The models behave as if they had underlying cognitive processes, but their proponents fail to consider whether other explanations (e.g., stochastically parroting training data) could also explain the observations.  I will discuss the internal structure of GenAI models and how to understand them.  I will also offer some theoretical suggestions for approaching understanding and artificial general intelligence.

BIOGRAPHIE

Herbert Roitblat, is lead data scientist for Egnyte’s research and development in artificial intelligence. Formerly professor of Psychology, Marine Biology, and Second Language Acquisition, University of Hawaii, Roitblat’s work on how dolphins recognize targets underwater with biosonar led to significant contributions to early neural network research and a patent on a binaural sonar. His more recent work is on artificial intelligence: “Algorithms Are Not Enough: Creating General Artificial Intelligence” (MIT Press, 2020) argues that algorithms and neural network models cannot fully capture the complexity of human cognition or animal intelligence and suggests what is needed to achieve artificial general intelligence.  

RÉFÉRENCES:

Bender, E. M., et al. (2021) On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM. 610-623.

Chomsky, N. (1967) A Review of B. F. Skinner’s Verbal Behavior. 142-143.

Harris, Z. (1954). Distributional structure. Word, 10(23): 146-162.

Huang K., & Chang, J.-C. (2023) Towards Reasoning in Large Language Models: A Survey. Findings of the Association for Computational Linguistics: ACL 2023, pages 1049–1065.

Roitblat, H. L. (2024). An Essay concerning machine understanding. arXiv preprint arXiv:2405.01840.

Roitblat, H. L. (2020). Algorithms are not enough: Creating general artificial intelligence. Mit Press.

Roitblat, H. L. (2017). Animal cognition. In: Bechtel & Graham, eds, A Companion to Cognitive Science, Wiley.

Olivier GEORGEON - 13 février 2025 à 10h30 au PK-5115 (201, ave President-Kennedy, 5e étage)

TITRE : Robotique développementale ancrée dans l'expérience moto-sensorielle

RÉSUMÉ 

La recherche en intelligence artificielle développementale puise certaines inspirations dans les théories du comportement animal de von Uexküll, l'épistémologie génétique et la psychologie développementale de Piaget, et l'épistémologie constructiviste de von Glasersfeld. Depuis la première implémentation d'un "schema mechanism" Piagétien par Drescher en 1991, ces recherches continuent à investiguer les questions clés associées : la motivation intrinsèque, la construction de sens, l'autonomie constitutive, le passage à l'échelle en monde ouvert, tout en tentant de les appliquer à la robotique autonome. Cette communication présente notre contribution à ces questions avec la plateforme de robotique libre PetitCat. Le robot PetitCat implémente l'hypothèse de l'ancrage de la connaissance dans des boucles d'interaction moto-sensorielles situées dans le temps et l'espace physique. Nous utilisons "moto-sensoriel" plutôt que "sensorimoteur" pour souligner le rôle clé de l'action, et l'exploitation de certains signaux sensoriels comme des feedbacks pas nécessairement représentationnels--idées déjà présentes chez Piaget et Wiener, puis approfondies par la théorie de l'enaction.

BIOGRAPHIE

Olivier GEORGEON est maître de conférences à l’Université Catholique de Lyon. Ses recherches portent sur l’apprentissage artificiel développemental, notamment les approches constructivistes appliquées à la robotique autonome. Il a développé des algorithmes pour des agents autonomes motivés intrinsèquement et a exploré la construction autonome de la mémoire spatiale et l’inférence énactive dans des environnements tridimensionnels. Georgeon a reçu une formation interdisciplinaire avec un diplôme d’ingénieur en informatique de l’École Centrale de Marseille et un doctorat en psychologie cognitive de l’Université Lumière Lyon

RÉFÉRENCES:

Georgeon, O., Lurie, D., and Robertson, P. (2024). Artificial Enactive Inference in Three-Dimensional World. Cognitive Systems Research  84: 101234.

Schneider H. Georgeon O. (2024). Grounding artificial general intelligence with robotics: The PetitCat project.  Proceedings of Brain inspired Cognitive Architectures (BICA). August 17 2024. Seattle.

Vidéo du séminaire à venir.

Michael ARBIB - 6 février 2025 à 10h30 au PK-5115

TITRE : The Evolution of the Construction-Ready Brain

RÉSUMÉ 

Humans, like animals, have perceptions, actions, and thoughts that they cannot put into words. The challenge is to understand how humans gained the ability to put so much into words, describing not only what is but also what is not and what could possibly be, and in this way came to construct a diversity of physical and symbolic worlds (Arbib et al., 2023, 2024). Building on modeling brain mechanisms for sensorimotor interaction with the world, we have explored both “how the brain got language” (Arbib, 2012) and what happens “when brains meet buildings” (Arbib, 2021). The little explored relation between these two studies is rooted in the notion of “construction” as used in a physical sense in architecture and in a symbolic sense as the tool for assembling utterances hierarchically. This talk offers hypotheses that address the question of how biological evolution yielded humans with the “construction-ready brains” and bodies that made us capable of the cultural evolution that created the diversity of our mental and physical constructs that we know today. The framework for all this is EvoDevoSocio – the idea that biological evolution yields biological mechanisms for both development and adult function of members of a species, but that social interaction is an important part of that environment, and that in humans cultural evolution has played the crucial role in changing the social, physical and increasingly symbolic and technological environments in which most humans now develop. The bridge between the two forms of construction is provided by the rooting of pantomime in manual action.

BIOGRAPHIE

Michael ARBIB, University Professor Emeritus and Professor Emeritus of Computer Science, Biomedical Engineering, Biological Sciences, and Psychology at the University of Southern California (USC). He is currently also an Adjunct Professor of Psychology at UCSD. Arbib’s research bridges neuroscience, computer science, and cognitive science, with a focus on the coordination of perception and action in frogs, rats, monkeys and humans. He applies schema theory and neural network analysis to study brain function, robotics, and machine vision. Known for the Mirror System Hypothesis, he explores language evolution through neural mechanisms for action understanding. More recently, he has explored the neuroscience of the experience of buildings, the design of buildings, and what it might mean for buildings to have “brains.

RÉFÉRENCES:

Arbib, M. A. (1964). Brains, machines, and mathematics. McGraw-Hill.

Arbib, M. A., & Caplan, D. (1979). Neurolinguistics must be Computational. Behavioral and Brain Sciences, 2, 449-483. [My thanks to Stevan Harnad for founding this important journal.]

Arbib, M. A. (Ed.). (2003). The handbook of brain theory and neural networks, Second Edition. MIT press.

Arbib, M. A. (2005). From monkey-like action recognition to human language: An evolutionary framework for neurolinguistics. Behavioral and Brain Sciences, 28(2), 105-124.

Arbib, M. A. (2012). How the brain got language: The mirror system hypothesis. Oxford University Press

Arbib, M.A. (2021). When Brains Meet Buildings: A Conversation Between Neuroscience and Architecture, Oxford University Press.

Arbib, M. A. (2023). Pantomime within and Beyond the Evolution of Language. pp. 16-57. In: Żywiczyński, P., Wacewicz, S., Boruta-Żywiczyńska, M., & Blomberg, J. (Eds.) Perspectives on Pantomime: Evolution, Development, Interaction. Benjamins.

Arbib, M. A., Fragaszy, D. M., Healy, S. D., & Stout, D. (2023). Tooling and construction: From nut-cracking and stone-tool making to bird nests and language. Current Research in Behavioral Sciences, 5, 100121

Arbib, M.A., Barham, L., Braun, R., Calder, B., Fox, M., Healy, S., Memmott, P., Smith, M., Stiphany, K., and Watkins, T. (2024) How Humans Came to Construct Their Worlds. A CARTA Symposium. Abstracts and videos of the talks are at https://carta.anthropogeny.org/events/how-humans-came-construct-their-worlds.