Séminaire du DIC

Chris ELIASMITH – 26 janvier 2023

Titre : Spaun 2.0: A large-scale model of biological cognition

Résumé :

The large-scale model of the brain, Spaun, has undergone significant development.  In this talk, I describe how it has more than doubled in size, to 6.3 million neurons, 20 billion connections, and significantly increased in functionality.  New functions include the ability to adapt online to changes in motor dynamics, classification of over 1000 categories of images, and perhaps most importantly the ability to perform simple 'mental gymnastics'. In this talk I describe the semantic pointer architecture (SPA) that is used to construct the model, demonstrate Spaun’s abilities, and discuss future plans for improving on what is currently the world's largest functional brain model.

Bio :

Chris ELIASMITH, Director, Centre for Theoretical Neuroscience (CTN) at the University of Waterloo, is the co-inventor of the Neural Engineering Framework (NEF), the Neural Engineering Objects (Nengo) software environment, and the Semantic Pointer Architecture (SPA), all dedicated to understanding how the brain works. His team has developed the Semantic Pointer Architecture Unified Network (Spaun) which is the most realistic functional brain simulation yet developed. Chris is the author of How to Build a Brain (Oxford University Press) and Neural Engineering (MIT Press).

Références

Duggins, P., & Eliasmith, C. (2022). Constructing functional models from biophysically-detailed neurons. PLOS Computational Biology, 18(9), e1010461.

Jan Gosmann and Chris Eliasmith. CUE: a unified spiking neuron model of short-term and long-term memory. Psychological Review, 128(1):104-124, 01 2021.

Voelker, A. R., Blouw, P., Choo, X., Dumont, N. S. Y., Stewart, T. C., & Eliasmith, C. (2021). Simulating and predicting dynamical systems with spatial semantic pointers. Neural Computation, 33(8), 2033-2067.

Choo, F. X. (2018). Spaun 2.0: Extending the world’s largest functional brain model.

Catherine TESSIER – 19 janvier 2023

Titre : « Éthique de l’intelligence artificielle » : une analyse critique

Résumé :

La profusion de documents et d'organismes qui traitent de l’« éthique de l’intelligence artificielle » nous amène à nous demander pourquoi l'intelligence artificielle est récemment devenue un objet d'attention particulier et quelle éthique est en jeu. À partir d’un examen de quelques documents internationaux, nous mettrons en évidence les problèmes liés au vocabulaire utilisé, aux postulats, et soulignerons des tensions et des paradoxes. À titre d'exemple, nous nous attarderons sur le principe de « contrôle humain ».  Nous conclurons sur les risques de détournement de l'éthique et sur la nécessité d'une véritable réflexion éthique tant au niveau de la recherche que de la conception et de l'utilisation des systèmes d'intelligence artificielle.

Bio :

Catherine Tessier est directrice de recherche à l’ONERA à Toulouse, France, et référente intégrité scientifique et éthique de la recherche de l’ONERA. Elle enseigne à l’ISAE-SUPAERO. Ses recherches portent sur la modélisation de cadres éthiques et sur les questions éthiques liées à l’« autonomie » des robots. Au niveau national français, elle est membre du Comité national pilote d’éthique du numérique et membre du Comité d’éthique de la défense. Elle a fait partie du Groupe d'experts ad hoc de l'UNESCO en vue de l'élaboration de la recommandation relative à l'éthique de l'Intelligence Artificielle.

Références

Tessier, C. (2022). «Autonomie» dans les systèmes d'armes: questionnements sémantiques, techniques, éthiques. Enjeux de l'autonomie des systèmes d'armes létaux.

Tessier, C. (2021). Éthique et IA: analyse et discussion. In CNIA 2021: Conférence Nationale en Intelligence Artificielle

Tessier, C. (2019). Éthique de la robotique et «robot éthique». Journal Polethis, 2

Xavier HINAUT – 12 janvier 2023

Titre : Sensorimotor Interaction of Language and Symbol Embodiment

Résumé :

Language involves several hierarchical levels of abstraction. Most models focus on a particular level of abstraction, making them unable to model bottom-up and top-down processes. It is not yet known how the brain grounds symbols to perceptions and how these symbols emerge throughout development. Experimental evidence suggests that perception and action shape one another (e.g., motor areas activated during speech perception) but the precise mechanisms involved in this action-perception shaping at various levels of abstraction are still largely unknown. My work includes modelling language comprehension, language acquisition from a robotic perspective, sensorimotor function and extended models of Reservoir Computing. I will also present general results on reservoir computing, and why it is an interesting framework to model cognitive processes, such as working memory.

Bio :

Xavier HINAUT is a Researcher in the Mnemosyne team at Inria in Bordeaux. His work focusses mainly on Recurrent Neural Network modelling (especially prefrontal cortex), language acquisition (applied to Robotics) and the brain codes of bird song syntax. The common thread is the neural coding and the modelling of complex sequence processing, “chunking,” learning and production, for “syntax-based” sequences, to be applied to robotics (for eventual embodiment). He manages the development of a new Reservoir Computing library in Python: https://github.com/reservoirpy/reservoirpy

Xavier HINAUT est chercheur dans l'équipe Mnemosyne à Inria à Bordeaux. Ses travaux portent principalement sur la modélisation des réseaux de neurones récurrents (en particulier le cortex préfrontal), l'acquisition du langage (appliqué à la robotique) et les codes cérébraux de la syntaxe des chants d'oiseaux. Le fil conducteur est le codage neuronal et la modélisation du traitement de séquences complexes, le "chunking", l'apprentissage et la production, pour des séquences "basées sur la syntaxe", à appliquer à la robotique (pour une éventuelle réalisation). Il dirige le développement d’une bibliothèque Python sur le Reservoir Computing: https://github.com/reservoirpy/reservoirpy

Références

Trouvain, N., Rougier, N., & Hinaut, X. (2022). Create Efficient and Complex Reservoir Computing Architectures with ReservoirPy. In International Conference on Simulation of Adaptive Behavior, pp. 91-102.

Pagliarini, S., Leblois, A., & Hinaut, X. (2021). Canary Vocal Sensorimotor Model with RNN Decoder and Low-dimensional GAN Generator. In 2021 IEEE International Conference on Development and Learning (ICDL), pp. 1-8.

Pagliarini, S., Leblois, A., & Hinaut, X. (2020). Vocal imitation in sensorimotor learning models: a comparative review. IEEE Transactions on Cognitive and Developmental Systems, 13(2), 326-342.

Strock, A., Hinaut, X., & Rougier, N. P. (2020). A robust model of gated working memory. Neural Computation, 32(1), 153-181.

Hinaut, X., & Dominey, P. F. (2013). Real-time parallel processing of grammatical structure in the fronto-striatal system: A recurrent network simulation study using reservoir computing. PloS one, 8(2), e52946.

Todd GURECKIS – 15 décembre 2022

Titre : Intuitive Physical Reasoning and Mental Simulation

Résumé :

The ability to reason about the physics of our world (e.g., what arrangements of objects are stable, how things will fall or move under a force) is central to human intelligence.  One influential hypothesis is that this capacity stems from the ability to perform “mental simulations” of physical events (in effect, playing a mental “movie” of the future evolution of a scene according to the laws of physics).  In this talk, I’ll try to pin down several core commitments of the mental simulation approach that must be present for the general theory to be viable.  I then will describe experiments we conducted recently trying to test these commitments.  Along the way, we stumbled into several curious and novel errors and biases in human physical reasoning ability that we believe represent limits to the universality of contemporary simulation theories.  If there is time, I will discuss a related project considering how efficient or optimal people are when they “experiment” in the physical world in order to learn the covert properties of objects such as mass or attractive/repulsive forces like magnetism.

Bio :

Todd M. Gureckis, Professor of Psychology, New York University, studies how people actively explore their world in order to learn, including everyday reasoning capacities for the physical and social world. His research combines methods of computational modeling, developmental psychology, cognitive neuroscience, and online data collection. He is the founder and a lead developer of the psiTurk package, a tool for facilitating online experiments used in hundreds of research labs. His work has been recognized by the NSF CAREER award, the Presidential Early Career Award (PECASE) from the Office of Science and Technology Policy at the White House, the James S. McDonnell Foundation Scholar award, and several paper and conferences awards with his students including the Marr Prize from the Cognitive Science Society, the Clifford T. Morgan Prize from the Psychonomic Society. He has variously served an Associate Editor for Cognitive Science, Topics in Cognitive Science, and Computational Brain and Behavior.

References

https://gureckislab.org/ :

https://gureckislab.org/papers/#/ref/ludwin2021limits

https://gureckislab.org/papers/#/ref/ludwinpeery2020broken

https://gureckislab.org/papers/#/ref/bramley2018intuitive

Titre : Active inference and artificial curiosity

Résumé :

This talk offers a formal account of insight and learning in terms of active (Bayesian) inference. It deals with the dual problem of inferring states of the world and learning its statistical structure. In contrast to current trends in machine learning (e.g., deep learning), we focus on how agents learn from a small number of ambiguous outcomes to form insight. I will use simulations of abstract rule-learning and approximate Bayesian inference to show that minimising (expected) free energy leads to active sampling of novel contingencies. This epistemic, curiosity-directed behaviour closes `explanatory gaps' in knowledge about the causal structure of the world, thereby reducing ignorance, in addition to resolving uncertainty about states of the known world. We then move from inference to model selection or structure learning to show how abductive processes emerge when agents test plausible hypotheses about symmetries in their generative models of the world. The ensuing Bayesian model reduction evokes mechanisms associated with sleep and has all the hallmarks of aha moments.

Bio :

Karl Friston is a theoretical neuroscientist and authority on brain imaging. He invented statistical parametric mapping (SPM), voxel-based morphometry (VBM) and dynamic causal modelling (DCM). Mathematical contributions include variational Laplacian procedures and generalized filtering for hierarchical Bayesian model inversion. Friston currently works on models of functional integration in the human brain and the principles that underlie neuronal interactions. His main contribution to theoretical neurobiology is a free-energy principle for action and perception (active inference). Friston received the first Young Investigators Award in Human Brain Mapping (1996), the Weldon Memorial prize and Medal in 2013 for contributions to mathematical biology, the 2016 recipient of the Charles Branch Award for unparalleled breakthroughs in Brain Research and the Glass Brain Award, a lifetime achievement award in the field of human brain mapping. He holds Honorary Doctorates from the University of Zurich and Radboud University.

A repository of active inference papers: GitHub - BerenMillidge/FEP_Active_Inference_Papers: A repository for major/influential FEP and active inference papers. Theoretical lecture on the physics behind active inference: I am therefore I think by Karl Friston - YouTube

Titre : Atlas of Forecasts: Modeling and Mapping Desirable Futures

Résumé :

Envisioning and implementing desirable futures requires a deep understanding of developments in science and technology as well as the ability to both simulate and communicate the likely impact of alternative actions. At a time when our relationship to a vulnerable planet Earth is especially important, such a profound awareness of complex, interlinked systems is needed more than ever. Atlas of Forecasts uses advanced data visualizations to introduce different types of computational models and demonstrates how model results can be used to inform effective decision-making. The models aim to capture the structure and dynamics of developments in education and the job market, progress in science and technology, and the impact of government policies—all from the micro to the macro levels. Model results can help us decide which human skills are needed in an artificial intelligence–empowered economy; which courses and degrees are most effective in upskilling and reskilling the current and future workforce; what progress in science and technology is likely to happen; and how policymakers can future-proof regions or nations.

Bio :

Katy Börner’s research focuses on the development of data analysis and visualization techniques for information access, understanding, and management. She is particularly interested in the formalization, measurement, and systematic improvement of people’s data visualization literacy; the study of the structure and evolution of scientific disciplines; the construction and usage of a Human Reference Atlas; and the development of cyberinfrastructures for large-scale scientific collaboration and computation.

References

Börner, Katy. 2021. Atlas of Forecasts: Modeling and Mapping Desirable Futures. Cambridge, MA: The MIT Press.

Börner, Katy, Andreas Bueckle, and Michael Ginda. 2019. Data visualization literacy: Definitions, conceptual frameworks, exercises, and assessments. PNAS, 116 (6) 1857-1864.

Börner, Katy. 2015. Atlas of Knowledge: Anyone Can Map. Cambridge, MA: The MIT Press.

Börner, Katy. 2010. Atlas of Science: Visualizing What We Know. Cambridge, MA: The MIT Press.

Here are links you might enjoy:

https://scimaps.org https://scimaps.org/books   <- Atlas Trilogy  https://www.pnas.org/doi/10.1073/pnas.1807180116  https://www.pnas.org/doi/pdf/10.1073/pnas.1804247115  https://www.nature.com/articles/s42254-021-00374-7   https://visanalytics.cns.iu.edu

Christian KEYSERS – 24 novembre 2022

Titre : Neural Basis of Empathy and Prosociality Across Species

Résumé :

:  How does our brain make us feel what others feel? How does it motivate us to help others? In humans, the somatosensory, insular and cingulate cortices are activated both when feeling pain and while witnessing others feeling pain. Altering brain activity in these brain regions alters emotional contagion and prosociality.

In humans, activity in the somatosensory cortex of observers predicts helping; perturbing that activity perturbs helping. Single cell recordings in rats show that neurons involved in an animal’s own pain become reactivated while the animal witnesses another animal in pain. This occurs in area 24, the rodent homologue of the anterior cingulate cortex in which humans show activation while witnessing the pain of others.

This region plays a causal role in sharing the emotions of others. The data show the existence of an evolutionarily conserved mechanism that maps the pain of others onto an observer’s own pain circuitry and triggers emotional contagion. When a rat can choose between a lever that produces food for herself, and one that produces food for herself but triggers a foot-shock to another rat, she learns to avoid the shock-lever. Deactivating area 24 abolishes this harm aversion, suggesting a causal link between emotional contagion and helping.

These experiments suggest that emotion-sharing is an evolutionarily conserved mechanism that allows humans and other animals to better prepare for unseen dangers by tuning into the state of those that have already detected them. This selfishly beneficial mechanism can promote prosociality, but it does so in fewer animals and situations than does the emotional contagion itself.

I will close with evidence that humans can voluntarily regulate how strongly they recruit their empathy, allowing us to leverage this ability when it is most helpful, and to downregulate it when it would be harmful.

Bio :

Christian Keysers studied how mirror neurons process the actions of others with Giacomo Rizzolatti. With Valeria Gazzola, he built the Social Brain Lab, Groningen, where their human fMRI work showed that participants activate their own actions, emotions and sensations while they witness those of others; this neural marker of empathy is reduced in patients with psychopathy. Since 2010, he leads the comparative social neuroscience effort in the Social Brain Lab, Netherlands Institute for Neuroscience, Amsterdam where he investigates the neural basis of empathy and prosociality across species: 

Keysers, C. (2011). The empathic brain https://doi.org/10.1016/j.tics.2022.05.005

Keysers, C., & Gazzola, V. (2014). Hebbian learning and predictive mirror neurons for actions, sensations and emotions. Philosophical Transactions of the Royal Society B: Biological Sciences, 369(1644), 20130175. https://www.ncbi.nlm.nih.gov/pmc/articles/PMC4006178/

Tom ZIEMKE – 17 novembre 2022

Titre : The observer’s grounding problem in human-robot interaction

Résumé :

People commonly attribute intentional mental states, such as beliefs and goals, to robots (Thellman et al., 2022; Ziemke, 2020). In a recent paper we formulated the perceptual belief attribution problem (Thellman & Ziemke, 2021): How can people interacting with robots understand what they know about the shared physical environment without knowing much about those robots’ sensors, perception, memory, etc.? In this talk I’ll focus on the observer’s grounding problem, which is the other side of the same coin, i.e., the fact that in interaction with a robot people tend to make anthropomorphic, folk-psychological attributions, based on their own grounding rather than the robot’s

Bio :

Tom ZIEMKE is Professor of Cognitive Systems at Linkoping University, Sweden. His main research interests are in situated/embodied cognition and social interaction, with a current focus on people’s interaction with different types of autonomous technologies, ranging from social robots to automated vehicles. A long-standing research interest is the relation between cognition and computation – and the resulting (mis-) conceptions of AI among both researchers and the general public

References:

Understanding robots https://www.science.org/doi/10.1126/scirobotics.abe2987  

Explainability in Social Robotics https://doi.org/10.1145/3461781  

Mental State Attribution to Robots https://doi.org/10.1145/3526112

Lorenzo NATALE – 10 novembre 2022

Titre : AI/robotics  and active visual and tactile perception

Résumé :

Modern AI algorithms provide exceptional performance but require long training time and large datasets that are expensive to annotate. On the other hand, robots can actively interact with the environment and humans using their sensory system to learn on-line how to perceive and interact with objects. To extract structured information, however, the robot needs to be endowed with appropriate sensors, fast learning algorithms, and exploratory behavior that guide the interaction with the world.

In this talk I will introduce the sensory system we developed for the iCub humanoid robot, and in particular the tactile sensing technology. I will then review work in which we studied how to use visual and tactile feedback to explore unknown objects and to control the interaction between the hand and the objects for shape modelling, object discrimination and tracking. Finally, I will present recent work in which we developed fast learning algorithms for object segmentation that leverage on the interaction with a teacher and active learning for adaptation to new contexts.

Bio :

Lorenzo NATALE, Senior Researcher at the Italian Institute of Technology and coordinator of the Center for Robotics and Intelligent Systems, was one of the main contributors to the design and development of the iCub humanoid robot. His research interests span artificial vision, tactile perception and software architectures for robotics.

References:

Ceola, F., Maiettini, E., Pasquale, G., Meanti, G., Rosasco, L., and Natale, L., Learn Fast, Segment Well: Fast Object Segmentation Learning on the iCub Robot, IEEE Transactions on robotics, 2022.

Maiettini, E., Tikhanoff, V., and Natale, L., Weakly-Supervised Object Detection Learning through Human-Robot Interaction, in Proc. International Conference on Humanoid Robotics, Munich, Germany, 2021

Vezzani, G., Pattacini, U., Battistelli, G., Chisci, L., and Natale, L., Memory Unscented Particle Filter for 6-DOF Tactile Localization, in IEEE Transactions on Robotics, vol. 33, no. 5, pp. 1139-1155, 2017

Baptiste CARAMIAUX – 3 novembre 2022

Titre : Interactive Machine Learning: Principles and Applications

Résumé :

Machine learning algorithms are present in many of the applications and services we use every day. These technologies are often designed in isolation from their users, leading to a standardisation of their uses and a centralised control of their capabilities. Creating learning technologies that are closer to people and their context of use opens up the possibility of more responsive, appropriable and inclusive interactions. In this talk, I will present the context and the research community working on these themes at the intersection between HCI and AI. Then I will focus on my work in this field. I will show examples of research where the artistic approach is sometimes seen as a tool to reflect on technologies as cultural actors, and sometimes seen as a tool to inspire the design of rich and expressive interactions. Finally, I will present concrete ways to design interactions with machine learning algorithms through the concept of Machine Teaching.

Bio :

Baptiste CARAMIAUX is a CNRS researcher at ISIR, Sorbonne Université in Paris, in the HCI Sorbonne group. He conducts research in human-computer interaction (HCI), studying and designing interactions with machine learning algorithms in the context of performing arts, health and pedagogy. engineering.