Auteur : Dagenais, Mylène

Yair LAKRETZ - 2 avril 2026 à 10h30 au PK-5115 (201, ave President-Kennedy, 5e étage)

TITRE :  Linguistic theory and deep language models

RÉSUMÉ 

Linguistic theory suggests that human language is organized by a latent hierarchical structure, which allows us to convey complex meanings and to seamlessly interpret sentences we have never encountered before. A central debate in the field is whether the brain encodes this abstract structure independently from the specific semantic content of words. While probing the human brain remains a challenge, modern neural AI models provide new means to test these theories by allowing for more granular analysis of how syntactic rules and meanings are represented within their architectures. In this talk, I will discuss evidence suggesting that these models can learn to decouple structure from content, aligning with the functional modularity proposed in linguistic theory. I will describe various methods for studying the underlying neural representations and mechanisms in these models, ranging from behavioral tasks to internal weight analysis. Finally, I will show how neuroscientific experiments can be designed to test specific predictions derived from these models, exploring how the human brain acquires and processes the structures of language.

BIOGRAPHIE

Yair LAKRETZ is a CNRS Research Scientist at the Laboratoire de Sciences Cognitives et Psycholinguistique (LSCP) in Paris, where he heads the Neuro-Linguae-AI team. His research focuses the neural mechanisms underlying language processing, particularly how the human brain and artificial intelligence systems encode and compute complex linguistic structures. He investigates these questions by combining the analysis of neural models (AI) with empirical studies on humans, utilizing neuroimaging techniques including fMRI, E/MEG, and intracranial recordings.

RÉFÉRENCES

Lakretz, Y., Hupkes, D., Vergallito, A., Marelli, M., Baroni, M., & Dehaene, S. (2021). Mechanisms for handling nested dependencies in neural-network language models and humans. Cognition, 213, 104699.

Evanson, L., Lakretz, Y., & King, J. R. (2023). Language acquisition: do children and language models follow similar learning stages?. In Findings of the Association for Computational Linguistics: ACL 2023 (pp. 12205-12218).

Lakretz, Y., Desbordes, T., Hupkes, D., & Dehaene, S. (2022). Can transformers process recursive nested constructions, like humans?. In Proceedings of the 29th International Conference on Computational Linguistics (pp. 3226-3232).

Diego Simon, P. J., d'Ascoli, S., Chemla, E., Lakretz, Y., & King, J. R. (2024). A polar coordinate system represents syntax in large language models. Advances in Neural Information Processing Systems, 37, 105375-105396. Rambaud, V., Mascarenhas, S., & Lakretz, Y. (2025). MapFormer: Self-Supervised Learning of Cognitive Maps with Input-Dependent Positional Embeddings. arXiv preprint arXiv:2511.19279.

Paul KANTOR - 26 mars 2026 à 10h30 au PK-5115 (201, ave President-Kennedy, 5e étage)

TITRE : Turing Meets Darwin: A challenge for LLMs

RÉSUMÉ 

The Turing Test asks whether a machine can imitate human conversation. A Darwinian perspective asks a prior question: how did humans evolve language at all? On this view, language is an adaptation for coordination, enabling individuals to pool information and solve problems that exceed the capacities of a single mind. This talk proposes a complementary challenge for LLMs: not whether they can simulate dialogue, but whether multiple artificial agents can discover that communication is instrumentally necessary for joint success. Drawing on comparative work on human and machine rule learning, and examples from animal intelligence, Kantor proposes a search for experimental paradigms in which communication must arise under task pressure. Such paradigms shift the focus from imitation to adaptive function, probing whether artificial systems can converge on communication for reasons analogous to those that shaped human language. 

BIOGRAPHIE

Paul KANTOR, Distinguished Professor (emeritus) of Information Science at Rutgers University, was affiliated with RUTCOR, DIMACS, and the Department of Computer Science. His research is on information storage and retrieval, rigorous evaluation of system effectiveness, and collaborative and large-scale data systems, with applications from scientific recommendation systems to homeland security analytics. He is currently Honorary Associate, Department of Industrial and Systems Engineering, University of Wisconsin–Madison. Educated in physics and mathematics at Columbia and Princeton, he is a Fellow of the AAAS and a recipient of the ASIST Research Award.

RÉFÉRENCES

Range, F., Kassis, A., Taborsky, M., Boada, M., Marshall-Pescini, S., 2019. Wolves and dogs recruit human partners in the cooperative string-pulling task. Sci Rep 9, 17591.

Jean-Rémy KING - 19 mars 2026 à 10h30 au PK-5115 (201, ave President-Kennedy, 5e étage)

TITRE : Language: in search of neural code

RÉSUMÉ 

Deep learning algorithms offer new methods to understand and model how language is processed in the human brain. Using both encoding (representation -> brain) and decoding (brain -> representations), we show that comparing modern speech and language models can account for brain responses to natural speech as recorded with EEG, MEG, iEEG and fMRI, including in children between 2 and 12 years old. This provides an operational foundation for modelling language in the adult and developing brain, and a new path to understanding the neural and computational bases of this human-specific ability.

BIOGRAPHIE

Jean-Rémy KING is a CNRS researcher at École Normale Supérieure currently seconded to Meta AI, where he leads the Brain & AI team, which aims to identify the cerebral and computational bases of human intelligence. The focus is on language, developing deep learning algorithms to decode and model brain activity recorded with MEG, EEG, electrophysiology and fMRI.

RÉFÉRENCES:

Evanson, L., Bulteau, C., Chipaux, M., Dorfmüller, G., Ferrand-Sorbets, S., Raffo, E., ... & King, J. R. (2025). Emergence of language in the developing brain. arXiv.

Lévy, J., Zhang, M., Pinet, S., Rapin, J., Banville, H., d'Ascoli, S., & King, J. R. (2025). Brain-to-text decoding: A non-invasive approach via typing. arXiv preprint arXiv:2502.17480.

Banville, H., Benchetrit, Y., d'Ascoli, S., Rapin, J., & King, J. R. (2025). Scaling laws for decoding images from brain activity. arXiv preprint arXiv:2501.15322.

Jacob FELDMAN - 5 mars 2026 à 10h30 au PK-5115 (201, ave President-Kennedy, 5e étage)

TITRE :  Human vs Machine in the Game of Hidden Rules

RÉSUMÉ 

Comparisons of human and machine intelligence are often grounded in supposition, unencumbered by empirical data about human performance. In this talk I'll present results comparing human and machine performance in on a common platform, the "Game of Hidden Rules" (GOHR). The GOHR is a simple rule-discovery game in which a player---human or AI---tries to classify objects into categories based on an unknown rule that they must infer by trial and error. Human players solve such problems about two orders of magnitude faster than (blank slate) AI models. In general, human and AI performance are almost completely uncorrelated, suggesting that contemporary AI does not yet effectively reflect the way that humans learn.

BIOGRAPHIE

Jacob FELDMAN is Professor of Psychology and Cognitive Science at Rutgers University, where he directs the Visual Cognition Lab. His research focuses on computational models of human visual perception and concept learning, particularly perceptual organization, shape representation, and categorization. Feldman has worked on the simplicity principle in human concept learning and Boolean complexity minimization, as well as on Bayesian models of perception and learning.

RÉFÉRENCES

Feldman, J. (2025). Simplicity and complexity of probabilistically-defined concepts. Psychological Review, in press.

Feldman, J. (2024). Probabilistic origins of compositional mental representations. Psychological Review, 131(3), 599-624.

Destler, N., Singh, M., & Feldman, J. (2023). Skeleton-based shape similarity. Psychological Review, 130(6), 1653-1671.

Feldman, J. (2021). Information-theoretic signal detection theory. Psychological Review, 128(5), 976-987.

Gary LUPYAN - 19 février 2026 à 10h30 au PK-5115 (201, ave President-Kennedy, 5e étage)

TITRE :  The unreasonable effectiveness of pattern matching

RÉSUMÉ 

  Large language models (LLMs) can often recover the meaning of texts in which most or all content words have been replaced by nonsense strings—so-called “Jabberwockified” language. In this talk, I argue that this ability reveals something important, and underappreciated, about both artificial and human cognition: the extraordinary power of large-scale pattern matching and constraint satisfaction. Drawing on new demonstrations showing that LLMs can reconstruct the gist—and sometimes surprisingly specific content—of highly degraded texts, I suggest that these systems are not best understood as parrots, databases, or blurry copies of the web. Instead, they exploit deeply learned structural regularities spanning syntax, discourse, and genre. I connect these findings to construction-grammar approaches to language, classic work on relational cognition, and evidence that human reasoning itself is graded, probabilistic, and pattern-based rather than strictly rule-governed. The broader implication is that pattern matching is not an alternative to “real” intelligence, but a central ingredient of it—one that helps clarify both the power and the limits of current language models.

BIOGRAPHIE

Gary LUPYAN is Professor of Psychology at the University of Wisconsin–Madison. His research examines how language shapes perception, thought, and learning, and how linguistic systems adapt to the needs of their users and learners. He has worked extensively on the role of language in human cognition, category learning, conceptual structure, and the evolution of communicative systems, with recent work exploring what large language models reveal about the nature of meaning and generalization.

RÉFÉRENCES

Lupyan, G., & Agüera y Arcas, B. (2026). The unreasonable effectiveness of pattern matching. arXiv preprint.

Lupyan, G. (2025). Large language models have learned to use language. arXiv preprint arXiv:2512.12447.

Wigner, E. (1960). The Unreasonable Effectiveness of Mathematics in the Natural Sciences. Communications on pure and applied mathematics, 12, 1–14.

Thomas SERRE - 12 février 2026 à 10h30 au PK-5115 (201, ave President-Kennedy, 5e étage)

TITRE : Cortical feedback mechanisms in visual reasoning: From perceptual grouping to abstract compositional reasoning

RÉSUMÉ 

This talk examines how cortical feedback facilitates compositional visual reasoning, distinguishing biological from artificial vision. Same-different judgments—fundamental symbolic operations that newborn ducklings master from single examples yet challenge state-of-the-art feedforward networks—illustrate this gap. Our computational studies suggest that structured, object-centered representations enable efficient learning of such abstract relations. These tasks require attention and working memory—inherently recurrent processes—as confirmed through human neurophysiological recordings. How are such structured representations built? Compositional understanding—from feature binding to relational and abstract reasoning—depends on cortical feedback. Brain-inspired recurrent models tuse feedback for iterative refinement of compositional representations. These solve challenging visual reasoning tasks—from curve tracing to object tracking—that remain difficult for feedforward architectures, including transformers. Recent evidence that primates use visual mental simulation when solving complex problems further supports the computational role of feedback in maintaining internal representations during reasoning. Together, these findings provide computational evidence that cortical feedback contributes essential mechanisms for the compositional reasoning capabilities that connect perception and abstract thought.

BIOGRAPHIE

Thomas SERRE is the Thomas J. Watson, Sr. Professor of Science and Professor of Cognitive & Psychological Sciences and Computer Science at Brown University. He is the Faculty Director of the Center for Computation and Visualization and Associate Director of the Center for Computational Brain Science. He also holds an International Chair in AI in the Artificial and Natural Intelligence Toulouse Institute (France). His research focuses on understanding neural computations supporting visual perception, with particular emphasis on the role of recurrent and feedback processes in visual reasoning.

RÉFÉRENCES

S. Muzellec, D. Linsley, A.K. Ashok, E. Mingolla, G. Malik, R. VanRullen & T. Serre. Tracking objects that change in appearance with phase synchrony. International Conference on Learning Representations, 2025

A. Ahuja, N.Y. Rodriguez, A.K. Ashok, T. Serre, T. Desrochers & D. Sheinberg. Monkeys engage in visual simulation to solve complex problems. Current Biology, 2024

A. Alamia, C. Luo, M. Ricci, J. Kim, T. Serre & R. VanRullen. Differential involvement of EEG oscillatory components in sameness vs. spatial-relation visual reasoning tasks, eNeuro, 2020

M. Ricci, R. Cadene & T. Serre. Same-different conceptualization:  A machine vision perspective. Current Opinion in Behavioral Sciences, 2020

Rajesh P. N. RAO - 5 février 2026 à 10h30 au PK-5115 (201, ave President-Kennedy, 5e étage)

TITRE : Predictive coding and generative models in natural and artificial intelligence

RÉSUMÉ 

This talk explores how predictive coding principles illuminate the computational foundations of both natural and artificial intelligence. Rao will examine his recent work on Dynamic Predictive Coding and Active Predictive Coding (APC) models, which proposes that the brain uses hierarchical generative models to predict sensory inputs and motor consequences. The discussion will cover how these models enable compositionality, hierarchical learning, and efficient planning by combining perception and action in a unified framework. Neuroscience evidence and AI applications suggests how predictive coding can help  us understand biological intelligence and develop more capable artificial systems that learn hierarchical world models for perception, action, and cognition.

BIOGRAPHIE

Rajesh P. N. RAO is the CJ and Elizabeth Hwang Professor of Computer Science & Engineering and Electrical & Computer Engineering at the University of Washington, Seattle. He is co-Director of the Center for Neurotechnology and directs the Neural Systems Laboratory. Rao received his PhD from University of Rochester (1998) and was a Sloan Postdoctoral Fellow at the Salk Institute. His research spans computational neuroscience, brain-computer interfaces, and artificial intelligence. He co-proposed the predictive coding model of brain function with Dana Ballard in 1999. His awards include a Guggenheim Fellowship, IEEE Fellow award, Fulbright Scholar award, NSF CAREER award, ONR Young Investigator Award, Sloan Faculty Fellowship, and Packard Fellowship.

RÉFÉRENCES:

Jiang, L. P., & Rao, R. P. N. (2024). Dynamic predictive coding: A model of hierarchical sequence learning and prediction in the neocortex. PLOS Computational Biology, 20(2), e1011801.

Rao, R. P. N. (2024). Active Predictive Coding: A Unifying Neural Model for Active Perception, Compositional Learning, and Hierarchical Planning. Neural Computation, 36(1), 1-58.
Gklezakos, D. C., & Rao, R. P. N. (2024). A sensory-motor theory of the neocortex based on active predictive coding. Nature Neuroscience.

Rao, R. P. N., & Ballard, D. H. (1999). Predictive coding in the visual cortex: A functional interpretation of some extra-classical receptive-field effects. Nature Neuroscience, 2(1), 79-87.

Jacob ANDREAS - 22 janvier 2026 à 10h30 au PK-5115 (201, ave. President-Kennedy, 5e étage)

TITRE: Systematicity in language models' knowledge and self-knowledge

RÉSUMÉ

Current language models (LMs) can converse knowledgeably, and in remarkable depth, about a wide range of topics. But these same LMs often generate confident-but-incorrect outputs, contradict themselves, and generally behave in ways that appear surprising and unnatural to human users. Increasingly, researchers attribute these failures not to surface-level statistical errors, but instead to mistakes and inconsistencies in LMs' "knowledge" or "beliefs" about the outside world. To what extent should we understand LMs as possessing beliefs at all? How should this understanding influence the procedures we use to train them? This talk will describe a family of training objectives that optimize language models for *internal systematicity* rather than predictive accuracy on some external dataset, showing that such objectives can improve models' linguistic and factual generalization, as well as the reliability of their explanations of their own behavior.

BIOGRAPHIE

Jacob ANDREAS is Associate Professor in the Department of Electrical Engineering and Computer Science at MIT and a member of CSAIL, where he directs the Language & Intelligence Group. His research focuses on understanding computational foundations of language learning and building intelligent systems that communicate effectively with humans. Andreas earned his PhD from UC Berkeley, MPhil from Cambridge as a Churchill Scholar, and BS from Columbia. He has received the Samsung AI Researcher of the Year award, MIT's Kolokotrones teaching award, and paper awards at NAACL and ICML. His work bridges machine learning and natural language processing, with particular expertise in compositional generalization, neural module networks, and systematic reasoning in language models.

RÉFÉRENCES

Akyürek, A. F., Akyürek, E., Choshen, L., Wijaya, D. T., & Andreas, J. (2024, August). Deductive closure training of language models for coherence, accuracy, and updatability. In Findings of the Association for Computational Linguistics: ACL 2024 (pp. 9802-9818).

Li, B. Z., Guo, Z. C., Huang, V., Steinhardt, J., & Andreas, J. (2025). Training Language Models to Explain Their Own Computations. arXiv preprint arXiv:2511.08579.

Damani, M., Puri, I., Slocum, S., Shenfeld, I., Choshen, L., Kim, Y., & Andreas, J. (2025). Beyond binary rewards: Training lms to reason about their uncertainty.

TITRE : The symbol grounding problem 75 years after Turing's Test (why computational success still leaves meaning unexplained)

RÉSUMÉ 

This talk examines the enduring symbol grounding problem 75 years after Turing's seminal 1950 paper, questioning why computational success in language tasks fails to resolve fundamental questions about meaning. Strohmaier will explore how large language models' impressive linguistic capabilities paradoxically highlight rather than solve the challenge of connecting computational representations to real-world meanings. Drawing from recent work in computational lexical semantics and philosophical analysis of machine understanding, the discussion will address why statistical pattern learning, despite its empirical success, leaves core questions about semantic grounding unanswered. The talk bridges computational linguistics and philosophy of mind, examining whether computational approaches can ever truly capture the referential and intentional aspects of human meaning-making, or whether meaning remains fundamentally unexplained by computational success alone.

BIOGRAPHIE

David STROHMAIER is a Research Associate in the Natural Language and Information Processing (NLIP) group at the University of Cambridge, supported by the Institute for Automated Language Teaching and Assessment (ALTA). His research applies machine learning and deep learning to lexical semantic acquisition, investigating how neural models learn word meanings and their relationship to human learning processes. Strohmaier has a highly interdisciplinary background, holding a PhD in Philosophy from the University of Sheffield (2018) and an MPhil in Advanced Computer Science from Cambridge. His work bridges computational linguistics and philosophy, with publications spanning word sense disambiguation, computational semantics, social ontology, and decision theory. He is co-author of "Preference Change" (Cambridge University Press, 2024).

RÉFÉRENCES:

Strohmaier, D., & Messerli, M. (2024). Preference Change. Cambridge University Press.
Strohmaier, D., & Wimmer, S. (2023). Contrafactives and Learnability: An Experiment with Propositional Constants. In Post-Proceedings of Logic and Engineering of Natural Language Semantics 19, 67-82.
Strohmaier, D., & Tyen, G. (2022). A Category Theory Framework for Sense Systems. In GLOBALEX 2022 @ LREC.
Strohmaier, D. (2021). Organisations as Computing Systems. Journal of Social Ontology, 7(1), 1-25.