Séminaire du DIC

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  Large language models (LLMs) can often recover the meaning of texts in which most or all content words have been replaced by nonsense strings—so-called “Jabberwockified” language.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.

NOTER qu'un enregistrement de la conférence sera mis en ligne après le jour du séminaire

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.

Sylvain CALINON - 11 décembre 2025 à 10h30 au PK-5115 (201, ave President-Kennedy, 5e étage)

TITRE:  Robot Learning from Demonstration

RÉSUMÉ

This talk explores how robots can efficiently acquire complex manipulation skills from minimal human demonstrations, addressing one of the fundamental challenges in modern robotics. I will present approaches that exploit the inherent structure and geometry of demonstration data to enable few-shot learning, moving beyond traditional imitation learning that requires extensive datasets. The discussion will cover representations for manipulation skills that can capture task variations and coordination patterns, optimal control techniques that bridge learning and control, and intuitive interfaces for meaningful human-robot interaction. Key topics include learning on Riemannian manifolds to handle orientation and manipulability constraints, tensor methods for exploiting multidimensional sensorimotor data, and bidirectional interaction strategies that allow robots to actively collect better demonstration data. I will demonstrate applications ranging from industrial manipulation tasks to assistive robotics, showing how robots can adapt learned skills to new situations and perturbations. The talk will address both the theoretical foundations of demonstration-based learning and practical considerations for deploying such systems in real-world scenarios.

BIOGRAPHIE

Sylvain CALINON is Senior Research Scientist at the Idiap Research Institute in Martigny, Switzerland, and Lecturer at the École Polytechnique Fédérale de Lausanne (EPFL). He heads the Robot Learning & Interaction group at Idiap, with expertise in human-robot collaboration, robot learning from demonstration, and model-based optimization. From 2009 to 2014, he was Team Leader at the Department of Advanced Robotics, Italian Institute of Technology (IIT). He holds a PhD from EPFL (2007), awarded the Robotdalen Scientific Award, ABB Award, and EPFL-Press Distinction. His work focuses on human-centered robotics applications where robots acquire new skills from few demonstrations, developing models that exploit data structure and geometry efficiently. He has received Best Paper Awards in Intelligent Service Robotics (2017) and IEEE RO-MAN (2007), and he currently serves as TC Chair on Model-based optimization for robotics for IEEE RAS.

RÉFÉRENCES

Li, Y., Chi, X., Razmjoo, A., & Calinon, S. (2024). Configuration Space Distance Fields for Manipulation Planning. Robotics: Science and Systems (RSS) - Outstanding Paper Award Finalist.
Shetty, S., Lembono, T., Löw, T., & Calinon, S. (2023). Tensor Train for Global Optimization Problems in Robotics. IEEE RAS Best Paper Award.

Jaquier, N., Rozo, L., Calinon, S., & Buerger, M. (2019). Bayesian Optimization Meets Riemannian Manifolds in Robot Learning. Conference on Robot Learning (CoRL) - Best Presentation Award.

Emmanuel DUPOUX - 4 décembre 2025 à 10h30 au PK-5115 (201, ave President-Kennedy, 5e étage)

TITRE: Is it really easier to build a child AI than an adult AI?

RÉSUMÉ

This talk reexamines Turing’s proposal to achieve machine intelligence by building an artificial child. With language acquisition as a testbed, I examine whether recent advances in self-supervised learning and large language models applied to child-centered audio or audio/video data take into account early phonetic and lexical developmental landmarks in real children. Focussing on the issues of robustness and data efficiency in child language learning, I will recast the long-standing controversy between statistical learning, social approaches and nativist hypotheses as an investigation of inductive biases in AI models in the light of ecologically realistic data. 

BIOGRAPHIE

Emmanuel DUPOUX is Professor at the École des Hautes Études en Sciences Sociales (EHESS) and directs the Cognitive Machine Learning team at the Laboratoire de Sciences Cognitives et Psycholinguistique (LSCP, ENS/CNRS/EHESS). He is also a part-time scientist at Meta AI Research. His research focuses on the mechanisms underlying cognitive and linguistic development in infants, combining experimental psychology, brain imaging, and machine learning. He holds a PhD in Cognitive Science (EHESS), an MA in Computer Science, and a BA in Applied Mathematics (ENS). He is recipient of an ERC Advanced Grant and organizer of the Zero Resource Speech Challenge, developing computational approaches to understanding how children learn language from their environment.

RÉFÉRENCES

Lavechin, M., de Seyssel, M., Métais, M., Metze, F., Mohamed, A., Bredin, H., Dupoux, E., & Cristia, A. (2024). Modeling early phonetic acquisition from child-centered audio data. Cognition, 245, 105734.

Rita, M., Strub, F., Chaabouni, R., Michel, P., Dupoux, E., & Pietquin, O. (2024). Countering Reward Over-optimization in LLM with Demonstration-Guided Reinforcement Learning. ACL Findings 2024.

Rita, M, Michel, P., Chaabouni, R., Pietquin, O., Dupoux, E., Strub, F. (2025). Language Evolution with Deep Learning. Chapter to appear in the Oxford Handbook of Approaches to Language Evolution
Benchekroun, Y., Dervishi, M., Ibrahim, M., Gaya, J.-B., Martinet, X., Mialon, G., Scialom, T., Dupoux, E., Hupkes, D., & Vincent, P. (2023). WorldSense: A Synthetic Benchmark for Grounded Reasoning in Large Language Models. arXiv:2311.15930.

Poli, M., Schatz, T., Dupoux, E., & Lavechin, M. (2025). Modeling the initial state of early phonetic learning in infants. Language Development Research, 5(1).

Chloe CLAVEL - 27 novembre 2025 à 10h30 au PK-5115 (201, ave President-Kennedy, 5e étage)

TITRE: Affective Computing and Emotional Understanding: Beyond the Cold Logic of the Turing Test

RÉSUMÉ

This talk examines how affective computing can transcend the traditional boundaries of the Turing Test by incorporating emotional understanding and socio-affective intelligence into AI systems. While the classic Turing Test evaluates a machine's ability to exhibit intelligent behavior indistinguishable from humans through purely linguistic exchanges, I will argue for expanding this paradigm to include emotional and social competencies. Drawing from recent advances in multimodal emotion recognition, social signal processing, and human-agent interaction, I will present computational models that capture not just what people say, but how they feel and the social context of their interactions. The discussion will cover our work on modeling socio-emotional behaviors including trust, engagement, and social stances in conversational AI systems, demonstrating how machines can be designed to recognize, understand, and appropriately respond to human emotions. I will address the challenges of moving beyond "cold logic" to develop socially intelligent systems that can navigate the nuanced landscape of human emotional expression, ultimately arguing that true artificial intelligence must incorporate affective understanding to be genuinely useful and acceptable to human users.

BIOGRAPHIE

Chloe CLAVEL is Senior Researcher (Directrice de recherche) at INRIA Paris in the ALMAnaCH team, focusing on Affective Computing and Artificial Intelligence. Until October 2023, she was Professor of Affective Computing at LTCI, Telecom-Paris, Institut Polytechnique de Paris, where she coordinated the Social Computing team. Her research lies at the intersection of multiple disciplines including speech and natural language processing, machine learning, and social robotics. Clavel studies computational models of socio-emotional behaviors--sentiments, social stances, engagement, trust--in both human-human and human-agent interactions. Her work spans multimodal emotion recognition, opinion analysis, social signal processing, and conversational AI systems. She is motivated by applications in health and education where affective computing can empower people and improve quality of life. Clavel has contributed to numerous European and national collaborative projects and serves as program chair for major AI conferences.

RÉFÉRENCES

Chenain, L., Bachoud-Lévi, A.-C., & Clavel, C. (2024). Acoustic Characterization of Huntington's Disease Emotional Expression: An Explainable AI Approach. ACIIW 2024.

Clavel, C., Labeau, M., & Cassell, J. (2022). Socio-conversational systems: Three challenges at the crossroads of fields. Frontiers in Robotics and AI, 9, 737173.

Guo, Y., Suchanek, F., & Clavel, C. (2024). The Curious Decline of Linguistic Diversity: Training Language Models on Synthetic Text. Findings of NAACL.

Guibon, G., Labeau, M., Flamein, H., Lefeuvre, L., & Clavel, C. (2021). Few-Shot Emotion Recognition in Conversation with Sequential Prototypical Networks. EMNLP.

Ari HOLTZMAN- 20 novembre 2025 à 10h30 au PK-5115 (201, ave President-Kennedy, 5e étage)

TITRE:  Articulating the Ineffable: The Analytic Turn in Generative AI

RÉSUMÉ

Generative AI has taken an analytic turn: we now cultivate models from objectives and data, then try to understand what we’ve grown. Current approaches to studying LLMs—focused on engineering progress or mechanistic explanations at the implementation level-—are insufficient for grasping their emergent behaviors. I will discuss what it means for interpretability approaches to be predictive rather than mechanistic, the changing landscape of machine communication, and efforts to identify fundamental laws that govern LLM behavior. I will argue that developing precise behavioral vocabulary and conceptual frameworks is the only way to turn the ‘fieldwork’ of finding surface regularities in LLMs into a science of LLMs. The guiding questions are basic, empirical, and exploratory: what do models consistently do, what do they reliably miss, and how do they incorporate and store new information? Along the way we’ll discover that AI has been given a new mandate—to articulate the ineffable, by describing aspects of communication and computation that we previously had no words for because they were stuck to deep inside human cognition to be easily referenced.

BIOGRAPHIE

Ari HOLTZMAN is Assistant Professor of Computer Science and Data Science at the University of Chicago, where he directs the Conceptualization Lab. His research is on developing new conceptual frameworks for understanding generative models, treating them as complex systems rather than traditional engineering artifacts. He introduced nucleus sampling, a text generation algorithm used in deployed systems including the OpenAI API.

RÉFÉRENCES

Holtzman, A., et al. (2023). Generative Models as a Complex Systems Science. arXiv:2308.00189.

Holtzman, A., Buys, J., Du, L., Forbes, M., & Choi, Y. (2019). The curious case of neural text degeneration. International Conference on Learning Representations (ICLR).

West, P., Holtzman, A., Hessel, J., Chandu, K., & Choi, Y. (2021). Symbolic knowledge distillation: from general language models to commonsense models. Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics.

Holtzman, A., West, P., Shwartz, V., Choi, Y., & Zettlemoyer, L. (2021). Surface form competition: Why the highest probability answer isn't always right. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing.

Rufin VANRULLEN - 13 novembre 2025 à 10h30 au PK-5115 (201, ave President-Kennedy, 5e étage)

TITRE: Global Workspace Theory Meets Deep Learning: Consciousness as Computational Architecture or Biological Phenomenon?

RÉSUMÉ

This talk examines the convergence of Global Workspace Theory and deep learning architectures in the quest to understand and potentially implement consciousness in artificial systems. I will present our recent work on Global Latent Workspace (GLW) models that bridge computational implementations of consciousness theories with state-of-the-art machine learning. The discussion will explore how these architectures integrate multimodal information processing through a central latent hub, enabling cross-modal translation and globally accessible representations. I will address the fundamental question of whether consciousness emerges from specific computational architectures or requires biological substrates, drawing on evidence from our implementations that combine global workspace dynamics with sensorimotor contingency theory. The talk will also examine the implications for AI consciousness assessment, discussing indicator properties derived from neuroscientific theories and their application to current AI systems. Finally, I will consider the ethical dimensions of potentially conscious AI and the importance of rigorous empirical approaches to machine consciousness research.

BIOGRAPHIE

Rufin ANRULLEN is CNRS Research Director in neuroscience and artificial intelligence at the Centre de Recherche Cerveau et Cognition (CerCo) and holds a research chair at the Artificial and Natural Intelligence Toulouse Institute (ANITI). His research focuses on brain-inspired AI architectures, visual perception, attention, and consciousness. Following mathematics and computer science studies, he completed his PhD in cognitive science with Simon Thorpe, then conducted postdoctoral research at Caltech with Christof Koch on visual attention mechanisms. He received the CNRS Bronze Medal in 2007 and was awarded a 2022 ERC Advanced Grant for his project “GLoW – The Global Latent Workspace.” VanRullen has authored over 200 scientific papers and is a leading researcher in computational approaches to consciousness and neural oscillations in perception.

RÉFÉRENCES

Kuske, N., & VanRullen, R. (2024). Consciousness in Artificial Systems: Bridging Global Workspace and Sensorimotor Theory in In-Silico Models. arXiv preprint.

Devillers, B., Maytié, L., & VanRullen, R. (2024). Semi-Supervised Multimodal Representation Learning Through a Global Workspace. IEEE Transactions on Neural Networks and Learning Systems.

Butlin, P., Long, R., Elmoznino, E., et al. [including VanRullen, R.] (2023). Consciousness in Artificial Intelligence: Insights from the Science of Consciousness. arXiv preprint.

VanRullen, R., & Kanai, R. (2021). Deep learning and the global workspace theory. Trends in Neurosciences, 44(9), 692-704.