Le cerveau sensoriel: visions bayésiennes et au-delà / The Bayesian sensory brain and beyond

Séminaire FSER organisé par Pierre Marie Lledo, Mariana Alonso, Gabriel Lepousez et Antoine Nissant (Institut Pasteur, Paris, France) du 17 au 22 avril 2017


Mariana Alonso, Brice Bathellier, Hugues Bersini, Sophie Deneve, Alfredo Fontanini, Patrick Haggard, Florent Haiss, Michael Halassa, Daniel Huber, Georg Keller, Adam Kepecs, Takaki Komiyama, Gabriel Lepousez, Pierre-Marie Lledo, Jean-Baptiste Masson, Vankatesh Murthy, Antoine Nissant, Mathias Pessiglione, Virginie van Wassenhove


Nos représentations du monde extérieur naissent d’un assemblage complexe de connaissances construites à partir d’informations multi-sensorielles parfois bruitées et souvent dynamiques. Pour donner du sens à ces informations sensorielles et agir en conséquence, notre cerveau doit effectuer tout une série d’opérations perceptuelles complexes qui sont nécessaires pour faire émerger la perception et qui sont parfois altérées dans certaines conditions pathologiques.

Dans ce symposium intitulé “Le cerveau sensoriel : Visions bayésiennes et au-delà”, nous souhaitions réunir à la fois des théoriciens et des expérimentateurs dans le contexte informel et stimulant que propose la Fondation Les Treilles pour partager autour des dernières avancées sur les mécanismes cérébraux de la perception et en particulier sur l’apport des approches statistiques dites Bayésiennes à tous les niveaux d’analyse de la perception sensorielle. En mettant autour de la table des statisticiens, des neurophysiologistes et des psycho-physiciens, travaillant sur des systèmes sensoriels différents (vision, audition, toucher, olfaction, gustation) à la fois chez l’Homme et l’animal, les diverses contributions ont permis d’éclaircir plusieurs points à la lumière des principes bayésiens incluant :

1) de nouvelles méthodes pour analyser et modéliser des données neurobiologiques,

2) les mécanismes neuronaux mis en jeu dans le codage de prédiction et dans l’influence descendante au sein des premiers étages sensoriels,

3) la plasticité des réseaux neuronaux induite par l’apprentissage et l’expérience,

4) les principes de la prise de décision qui relie perception et action.

Compte-rendu (en anglais)

Proceedings of the conference are summarized below in the form of a brief outline of the individual presentations.

Brice Bathelier (UNIC, CNRS, Gif-sur-Yvette, France) kicked off the meeting by presenting some data on uni- and multi-sensory representations in the primary auditory cortex of the mouse. Using large scale two-photon imaging and optogenetic techniques in mouse auditory cortex as well as computational modeling and behavior (operant conditioning), he showed that auditory cortex responds to intensity variations of sounds in a complex, nonlinear manner, illustrating the tremendous importance of temporal features (intensity and frequency variations in time) for recognition of sounds and the importance of auditory cortex for discriminating sound with temporal features. These responses build highly divergent representations of sounds that differ only in their intensity time course, as expected from the divergent percepts produced by these stimuli.

Michael Halassa (NYU, New York, USA) explored new directions in thalamic control of cognitive processes. Indeed, while the role of thalamus in transmitting sensory signals is well-studied, its broader engagement in cognition is unclear. In his talk, he showed evidence that the thalamus regulates functional connectivity within and between cortical regions, demonstrating how a cognitive process is implemented across distributed cortical microcircuits. In the framework of attention control of perception, he combined in vivo single unit recordings and optogenetic control of cortico-thalamic circuits during the execution of a cross-modal, divided attentional task and showed that thalamic circuits are important to filter and determine the categorical content of a cognitive process (e.g., sensory details in feature-based attention). Additionally, he showed that in the medio-dorsal thalamus to prefrontal cortex network the thalamus also provide a route by which task-relevant cortical representations are sustained and coordinated, showing for the first time a form of thalamic control of functional cortical connectivity.

Sophie Deneve (Ecole Normale Supérieure, LNC, Paris, France) explored the relevance on circular inference and excitatory/inhibitory balance in normal and altered perceptual representation. Using theoretical and modeling approaches, she first presented a hierarchical inference framework in which the percept is a function of the prior belief and top-down prediction error sends by the hierarchy. Then, in the context of a perceptual illusion, as observed in schizophrenia, she introduced the principle of circular inference, which refers to a corruption of sensory data by prior information and vice versa. This impairment in Bayesian inference mechanisms could lead to ‘see what we expect’ (through descending loops), to ‘expect what we see’ (through ascending loops), leading to an “over-representation” of one sensory feature or both. Interestingly, this model successfully captured not only the averaged behavior, but also the inter-participant variability in both control and schizophrenic patients.

Adam Kepecs (Cold Spring Harbor Laboratory, New-York, USA) discussed about confidence judgments and provided some evidences that confidence is framed in mathematics as an objective statistical quantity –the probability that a chosen hypothesis is correct– in rats, monkeys and human. In a Bayesian brain framework, predictions suggest that confidence should be characterized by three signatures of statistical decision confidence: 1) statistical confidence predicted the mean choice accuracy, 2) the statistical confidence for a given level of evidence discriminability increases for correct and decreases for incorrect choices, 3) for each level of evidence discriminability, accuracy for high-confidence choices is greater than for low-confidence choices. Interestingly, neuronal responses in the ventro-lateral orbito-frontal cortex also verify these three signatures and reversible inactivation of this region impairs confidence. Lastly, he showed that VTA dopaminergic neurons of the monkey also incorporate the signature of statistical confidence in addition to encoding a post-decision teaching signal.

Mathias Pessiglione (Motivation, Brain and Behavior lab, Institut du Cerveau et de la Moelle, Hôpital de la Pitié-Salpêtrière, Paris, France) discussed his work on reinforcement learning, motivation and apathy with a particular focus on the cost/benefit trade-off. Apathy can be defined as a reduction of goal-directed behavior and is frequently observed in psychiatric and neurological diseases. To understand the neural dysfunction underlying apathy, it is necessary to decompose this syndrome into elementary computational processes, which may rely on Bayesian principles. A key motivational process is the arbitrage between costs and benefits: apathy can result either from hyposensitivity to potential rewards (or benefit) or from hypersensitivity to potential efforts (or cost). Using a model-based analysis of pharmacological studies and fMRI data from Parkinson’s disease patients, he provided insights into the motivational effects of standard treatments. Notably, dopamine receptor agonists appeared to enhance reward attractiveness, whereas serotonin reuptake inhibitors seemed to alleviate effort cost.

Virginie Van Wassenhove (Cognitive Neuroimaging Unit, CEA DRF/Joliot, INSERM, Université Paris-Sud, Université Paris-Saclay, NeuroSpin center, Gif/Yvette, France) presented data on multisensory integration, time structures and abstraction in the human mind. She first discussed why computing the temporal co-modulation of multisensory inputs may be relevant for audiovisual speech processing and multisensory integration. Notably, she presented a recent MEG study relating the benefits of supra-modal processing in the discrimination of visual motion coherence. She then discussed how an individual’s perception of order and simultaneity across sensory modalities displays intrinsic biases that can be modulated by oscillatory activity, such that the phase of neural oscillations in an entrainment regime may indicate an individual’s perception of order.

Jean-Baptiste Masson (Institut Pasteur, Paris, France) discussed three aspects of his current work on decision and Bayesian computations. He first addressed Bayesian strategies to understand synaptic receptor dynamics in synapses at the single receptor level. Combining machine learning, graph theory, Bayesian inference and random walk theory, he showed how the physical environment of synapses can be probed and how a link between structures and dynamics can be understood using an example of the glycine receptors in inhibitory synapses. He then moved to drosophila larva decision-making in uncertain environments. Drosophila larva are ~10 000 neurons animals exhibiting a large repertoire of behavior and complex responses to changes in sensory environments. He showed that larvae are able to solve complex problems where a direction must be chosen after receiving noisy, side-dependent, nociceptive stimuli. Finally, he addressed amortizing Bayesian search strategies in complex environments. Searching for olfactory sources in turbulent environments is difficult since few cues can be detected and they provided only limited information pertaining to the position of the source. He illustrated how Vergassola’s Infotaxis can be amortized to allow searches with limited space perception and limited memory.

Georg Keller (FMI, Basel, Switzerland) presented his current work on predictive processing in primary visual cortex of mice. Vision is intrinsically an active process as most visual input is the direct consequence of self-generated movements. He speculated that the computational function of layer 2/3 of primary visual cortex of the mouse is that of a comparison between expected and actual visual input. Using two-photon imaging of visual cortex neuron responses of mice exploring a virtual environment, he showed that visual cortex combines motor-related predictions of visual flow with visual signals to generate mismatch responses that signal a deviation from expectation. Mismatch signals are likely generated through a combination of an excitatory feedback predictions and somatostatin-interneuron mediated feed-forward inhibition. In addition, he showed that both feedback predictions and mismatch responses are critically dependent on sensorimotor experience. All together these data illustrated that the function of the visual system may be the generation of an internal model of the visual environment that is continuously compared to and updated by visual input.

Daniel Huber (University of Geneva, Department of Basic Neuroscience, Geneva, Switzerland) presented his new approach to probe cortical circuit function with neuro-prosthetic learning. When reaching for an object, the motor command generating the action is guided by sensory feedback from the moving arm. To bypass the neural complexity associated with natural behavior he asked whether both the motor command and the sensory feedback could be artificially confined in the cortex. Mice were trained to activate a single neuron in the forelimb motor cortex (M1) to obtain rewards, while a proportional optogenetic rate code provided real-time feedback of its activity level by stimulating the forelimb somatosensory cortex. This fabricated cortical channel led to a progressive but rapid increase of activity in the operant conditioned M1 neuron. The artificial sensory feedback was necessary for detecting the reinforced activations and learning to produce them more frequently. Longitudinal imaging of neighboring M1 cells with two-photon microscopy revealed that the learning-related activity changes were specific to the conditioned neuron instead of engaging a neural population response. These findings demonstrate that cortical stimulation can be translated into informative real-time feedback of ongoing neural activity. It reveals the capacity of animals to use an artificial cortical communication channel in a behaviorally relevant way and showed the speed and specificity at which this can occur.

Hugues Bersini (Université Libre de Bruxelles, Bruxelles, Belgium) addressed the relevance of artificial intelligence for neurosciences and the importance of neural networks for deep learning. After a brief historical overview of artificial intelligence, he presented the different forms of artificial intelligence, their architectures and their outcomes in the context of big data processing. Then he discussed what more recent research in neural networks, such as deep learning (deep in space and time), self-adapting nets, and chaotic Hopfield networks, might bring to neuroscience.

Mariana Alonso (Institut Pasteur, Paris, France) talked about the functional significance of adult-born neurons in the olfactory system for odor-reward associations. In mammals, most odors elicit responses acquired following learning, as the assignment of a positive value to a given odorant. To be effective, these odor-reward associations require multiple forms of plasticity. In rodents, one of the mechanisms underlying plasticity in the main olfactory bulb consists in recruiting new neurons every day through life. Using optogenetic and psychophysic, she presented new data showing that exposure to reward-associated odors specifically increases activity of adult-born neurons, but not neurons generated around birth. Remarkably, activating the adult-born neurons by optogenetic specifically during rewarded odor presentation boosted discrimination learning and odor value update during reversal association, suggesting that adult-born neurons may be involved in facilitating odor-reward association during adaptive learning.

Venkatesh Murthy (Harvard University, Boston, USA) discussed about olfactory object analysis and categorization in mice. In the context of invariable object detection in a variable background, he showed how a linear decoder with sparse weights could match mouse performance using just a small subset (~15) of the glomerular inputs to the olfactory system. However, when such a decoder, or mice are trained only with single odors, it generalizes poorly to mixture stimuli due to nonlinear mixture responses. He then asked what are the neuronal substrates for odor categorization in the olfactory system? He showed that the olfactory tubercles (OT), a ventral striatum area, display categorical reward-related responses that discriminate more efficiently reward from unrewarded odors compared to responses in posterior piriform cortex. Learning a new reward category is tracked by category decoding in the OT, suggesting that OT can encode valence categorization.

Takaki Komiyama (UCSD, San Diego, USA) explored the neural substrate of learning in the motor cortex. Using wide-field calcium imaging of dorsal cortex during motor learning, he characterized a sequential activation of different cortical modules, with the pre-motor cortex (M2) activation progressively leading the activation sequence during learning. He then presented new data showing that, in the course of a motor learning process, while animals progress from the naive to expert condition, the involvement of the primary motor cortex (M1) is gradually reduced.

Gabriel Lepousez (Institut Pasteur, Paris, France) discussed the functional anatomy and relevance of top-down pathways in the olfactory system, focusing on cortical projections to the olfactory bulb (OB). Using genetic labeling of cortico-bulbar inputs and optogenetics, he first described how cortical top-down inputs target OB adult-born granule cells and how this cortico-bulbar connectivity is enhanced after odor-reward association. With in vivo electrophysiology and optogenetics, he then showed that activation of cortical top-down inputs recruits feedforward inhibition onto OB principal cells in a frequency-dependent manner and is modulated by GABAB receptors. Lastly, he characterized the presence of a new inhibitory top-down circuit supported by long-range GABAergic projections from olfactory cortex interneurons to the OB. These inhibitory long-range projections mainly target local OB interneurons and may participate in the control of odor perception.

Florent Haiss (RWTH Aachen University, Aachen, Germany) discussed about neocortical response dynamics in somatosensory cortex during perception. In sensory cortex, responses typically adapt to repeated sensory stimulation, improving sensitivity to stimulus changes, but the relevance of this adaptation is not clear. He showed that discrimination performance is enhanced using non-adapting optogenetic cortical stimulation compared to whisker stimulation, which causes frequency-dependent adaptation. This improvement persisted when temporal precision of optically evoked spikes was reduced, but not when applying adaptation rules mimicking sensory-evoked responses to optical stimuli. This suggested that sensory adaptation is relevant for the perception of stimulus patterns, decreasing fidelity under steady-state conditions, in favor of change detection.

Alfredo Fontanini (Stony Brook University, New York, USA) then invited us to the gustatory cortex to examine anticipatory signals in the context of associative learning. As animals navigate their environment, they associate various stimuli with the presence of food. These associations transform neutral stimuli into predictive cues. He showed how his group has studied predictive coding of taste change activity in the gustatory cortex and it relevance for taste processing and behavior. Using behavioral electrophysiology, manipulations of neural activity, and computational methods, he showed that following cue-taste associations, gustatory cortex neurons started to respond to non-gustatory stimuli with these cue-responses encoding general expectations of gustatory stimulation as well as expectations of specific gustatory outcomes. When the ability of different modalities to engage into cue-taste associations was tested, he observed that odors and tactile stimuli were learned more rapidly than sound and light stimuli. He also found that cue responses play a role in guiding consummatory behaviors and in speeding-up taste coding, changing attractor dynamics of gustatory cortex networks. This suggests that gustatory cortex integrates sensory information with cognitive signals to better serve feeding behaviors and ingestive decisions.

Patrick Haggard (Institute of Cognitive Neuroscience, University College, London, UK) discussed about action, volition and agency in humans. Voluntary actions are often defined as actions that are internally-generated, rather than directly triggered by an external stimulus. Classically, we consciously deliberate and form intentions to drive our actions. To measure volition, he first presented the difficulty to study this experimentally, showing some experiments to illustrate the experience of being about to act as a direct readout of ongoing neural preparation. He then discussed about the sense of agency defined as this mental capacity to link actions to outcomes on the outside world. As a marker of the sense of agency, he focused on distortions of time perception in a task involving the ‘intentional binding’ effect. In this task, participants are asked to report the perceived time either of a voluntary action or of a subsequent sensory event (such as a tone). It has been shown that voluntary actions, but not involuntary movements, are perceived as shifted in time towards their subsequent outcomes and that the outcomes themselves are perceived as shifted towards the voluntary actions that caused them. This shift is analyzed in comparison to control conditions in which the action and outcome occur independently. As a result, the sense of agency can be quantified as a compression of the interval perception between action and outcome. He then explored the contribution of the sense of agency for improving instrumental learning and pointed out the importance of dopamine on this effect.

Concluding remarks

This symposium “The Bayesian Sensory Brain and Beyond” clearly demonstrated the decisive progress in Neuroscience which is now in a new era. The synergistic contribution of a variety of experimental approaches, sensory systems, models and levels of analysis (from molecules to behavior), along with the facilitated dialogue between theoreticians and experimentalists, have been reinforced nowadays by major methodological advances in psychophysics, neuronal physiology and genetics tools for monitoring and manipulating specific neuronal circuits during behavior. These unique opportunities have strongly increased our capacity to understand brain mechanism of perception and draw causal inference between circuits, computations and cognitive processes. The pace of discovery and publications is impressive and raise exiting expectations for the future. Importantly, the diversity of the audience expertise, the format of the symposium, with hour-long presentations, and the large time allocated for discussion, the hospitality and efficiency of the fondation staff, along with the inspiring beauty of the site, were all key factors to make this symposium both a stimulating, interactive and productive conference as well as a pleasant and memorable moment. For all these reasons, we would like to thank the “Fondation Les Treilles” for making feasible this amazing social/scientific meeting.

Citer ce billet
Fondation des Treilles (2017, 22 avril). Le cerveau sensoriel: visions bayésiennes et au-delà / The Bayesian sensory brain and beyond. Les carnets de la Fondation des Treilles. Consulté le 20 mai 2024, à l’adresse https://doi.org/10.58079/qv2i

Rechercher dans OpenEdition Search

Vous allez être redirigé vers OpenEdition Search