Cervelet et apprentissage/ Cerebellum and learning
Séminaire organisé par Nicolas Brunel et Vincent Hakim, du 9 au 14 septembre 2024
Participants
Boris BARBOUR (Institut de Biologie de l’Ecole Normale Supérieure – Paris, France), Nicolas BRUNEL (Duke University – Durham, États-Unis), Alex CAYCO GAJIC (Institut de Biologie de l’Ecole Normale Supérieure – Paris, France), Chris DE ZEEUW (Erasmus MC – Rotterdam, Pays-Bas), Stéphane DIEUDONNE (Ecole Normale Supérieure / INSERM – Paris, France), Vincent HAKIM (Ecole Normale Supérieure & CNRS – Paris, France), David HERZFELD (Duke University School of Medicine – Durham, États-Unis), Court HULL (Duke University School of Medicine – Durham, États-Unis), Stephen LISBERGER (Duke University School of Medicine – Durham, États-Unis), Javier MEDINA (Baylor College of Medicine – Houston, États-Unis), Abigail PERSON (University of Colorado School of Medicine – Denver, États-Unis), Jennifer RAYMOND (Stanford University – Stanford, États-Unis), Wade REGEHR (Harvard Medical School – Boston, États-Unis), Aparna SUVRATHAN (McGill University – Montreal, Canada)
Résumé

©Boris Barbour
Le séminaire « Cervelet et apprentissage » a été consacré à la question fondamentale du fonctionnement et de l’apprentissage neuronal, sur l’exemple du cervelet, l’une des régions cérébrales les mieux comprises actuellement. Il a réuni quatorze experts éminents (6 européens et 8 nord-américains), dont deux jeunes scientifiques prometteurs, possédant une expertise en neurosciences expérimentales ou computationnelles, en sciences cognitives ou en physique. Les participants ont exposé leurs recherches en cours lors d’exposés d’une heure entrecoupés de nombreuses questions, ce qui a permis des discussions approfondies sur les résultats présentés. Les échanges et la confrontation des différents points de vue se sont poursuivis avec tous les participants lors des séances de fin d’après-midi (d’une durée de 1 à 2 heures chaque jour). L’atmosphère et le cadre enchanteur ont également encouragé les discussions en petits groupes pendant les repas, les pauses en début d’après-midi et les longues soirées. Les participants ont unanimement déclaré qu’ils avaient beaucoup apprécié le séminaire et qu’ils en avaient énormément profité.
Abstract
The Seminar « Cerebellum and learning » was devoted to the fundamental question of how the brain functions and learns, by focusing on the cerebellum, currently one of the best understood brain regions. It gathered fourteen eminent experts (6 from Europe, and 8 from North America), including two promising young scientists, with expertise in experimental or computational neuroscience, in cognitive science or in physics. The participants exposed their current research in one-hour talk interspersed by numerous questions, which allowed in-depth discussions of the presented results. The discussions and the confrontation of different viewpoints continued with all participants during the late afternoon sessions (1 to 2 hours long every day). The wonderful atmosphere and setting also encouraged discussions in smaller groups during meal times, early afternoon breaks and the long evenings. Participants unanimously reported that they very much enjoyed the Seminar, and felt that they had tremendously benefited from it.
Report
The cerebellum, a brain structure located at the back of the brain, is classically thought to be involved in fine motor control and has been found to be implicated in a variety of other brain functions, like language, attention or regulation of emotions, and brain dysfunctions like autism. Unlike most other brain regions, the cerebellum has a very regular “crystalline” anatomical structure which has long attracted the attention of experimentalists and theoreticians. Following early work by John Eccles and his collaborators, the classical theory of cerebellar operation and learning was proposed 50 years ago by David Marr and James Albus. Some of the basic tenets of this theory were verified in experiments by Masao Ito and his collaborators. The Marr-Albus-Ito Theory (MAIT) proposes that the Purkinje cells (PCs), the principal cells of the cerebellar cortex and its only outputs, learn to associate contextual information coming from the very numerous granule cells via their parallel fiber (PF) axons to generate proper motor outputs, in a “supervised” manner, guided by climbing fiber (CF) inputs coming from the inferior olive (IO), a nucleus of the brain stem. Several predictions of this theory have been verified experimentally since the pioneering work of Ito and collaborators. However, it has also been realized that MAIT leaves many issues unsolved and does not account for multiple features of the recorded data. These were the central questions discussed in the Seminar.
Lisberger in the starting talk of the Seminar summarized the decades of experiments from his own lab that document the transformation from a moving visual input to the appropriate eye movements in a task where the animal must follow a moving target with its eyes. This task has been shown to require the cerebellum to successfully learn to adapt to consistent changes in movement of the target. These experimental results are compatible with MAIT for fast learning acquisition in the cerebellar cortex, with a second, slower learning site with longer retention in the deep cerebellar nucleus (DCN). Lisberger emphasized that most of our current knowledge arises from studying the adaptation of eye movements in “simple’’ tasks like the vestibulo-ocular-reflex (VOR where, in normal conditions, the eyes rotate in opposite direction from the head to maintain a fixed gaze) or smooth pursuit of a moving visual target, and in very specific regions of the cerebellum like the flocculus. One central current question is whether the knowledge acquired from these experiments generalizes to more complex movements, other cerebellar regions or other cerebellar dependent non-motor tasks and, if so, how. This question was addressed in the following talks and constituted one of the central themes of the Seminar. Other important questions included the respective roles of the cerebellum and the DCN, which are the single targets of Purkinje cells, as well of the nature of the information representation in granule cells.
Suvrathan showed that timing requirements of plasticity at the PC-PF synapse are regionalised and related these variations to the delays of sensory feedback arriving via the IO. The plasticity could thus be adapted to serve specific roles of Purkinke cells in controlling diverse types of movement. She also reported her current work in characterizing the heterogeneity of time scales of synaptic signaling in Purkinje cells, focusing particularly upon variations of mGluR-mediated responses and how those variations were correlated with the regionalised marker zebrin II..
Regehr reported the progress of his lab on the classification of the different cell types of the cerebellum and the characterization of their connectivity. He emphasized his discovery of two types of molecular layer interneurons that do not correspond to the classical distinction between stellate and basket cells and obey distinct connectivity rules, between themselves, to Purkinje cells, and from the IO. This begs the question of the functional roles of these two types of interneurons in controlling Purkinje cell activity and learning.
Medina described the complex anticipatory firing of IO cells after learning during the classical eye-blink conditioning paradigm. In this experimental setting, a protective eyelid closure reflex to an air puff is associated in a Pavlovian fashion to a neutral anticipatory stimulus such as a flash of light or a sound. Complex spikes in Purkinje cells, reflecting IO inputs between the anticipatory stimulus and the air puff are observed to decrease or increase in neighboring PCs. Further experiments with repeated neutral stimuli show also the predictive timing abilities of IO spikes. These complex IO firing patterns are not predicted by MAIT and may reflect reinforcement learning or the formation of internal models in the cerebellum. A novel “oddball” learning behaviour was reported, in which the animal learns to react to the absence of an expected repetitive stimulus.
Equally puzzling observations about the firing properties of IO cells that are not consistent with MAIT during multiple experimental conditions, including learning of complex movements and those guided by rewards not errors, were reported by both Hull and Dieudonné. The latter highlighted the role of doublets of complex spikes in the adaptation of spontaneous movements. Explanations of these observations remain to be fully developed and they raise many mechanistic questions that were discussed at length among the participants.
Barbour stressed that MAIT does not address the “credit-assignment problem”, namely the determination of the specific synapses to be modified to learn a complex movement with poorly informative sensory feedback. He described his proposal that this problem could be solved using stochastic perturbation of PC by IO cells during movement learning, as well as the current efforts of his team to distinguish between the predictions of this theory and MAIT in eyeblink conditioning experiments.
Hakim discussed the convergence properties of this new proposed learning algorithm, which falls into the general class of node-perturbation algorithms. The outcome of the analysis is that the learning time grows with the number of output variables to be adjusted. This suggests that learning can be faster if complex movements are built from movement primitives.
Person took up the analysis of movement control using of a reaching task. Reaching movements can be affected by an oscillatory approach to target in cerebellar patients (dysmetria). She showed using optogenetic perturbations how the interposed nucleus activity causally supports braking behavior close to the movement endpoint, and how this is related to the activity of groups of PCs that increase or decrease their firing during movement. Notably, the data in this case could be explained by a model first developed in the context of eye smooth pursuit by Herzfeld and Lisberger.
Herzfeld described his current efforts to classify and monitor the activities of all different types of cerebellar cell activities during smooth pursuit to obtain a detailed quantitative description of how the proper output signal is constructed and how information from eye position in the MFs is transformed into velocity coordinates in PCs.
The observation of heterogeneous timing requirements for cerebellar plasticity led Raymond to discuss metaplasticity in the cerebellum, namely how synaptic plasticity rules themselves can be modified by experience. Her experiments in both mutant mice and wild-type mice using pharmacological perturbations are consistent with “threshold metaplasticity” a shift in the thresholds characterizing synaptic plasticity rules. Similarly, experiments with dark-reared mice led her to conclude that precise timing of plasticity rules could be the result of “timing metaplasticity” and proposed a simplified biochemical model that could account for such metaplasticity.
Brunel described how a model combining two variables—presynaptically released nitric oxide (NO) and postsynaptic calcium level—could reproduce and predict the results of different plasticity protocols. He further showed that reducing the amplitude of the calcium transients would account for the plasticity rule involving doublet olivary complex spikes reported by Barbour, and that heterogeneity in the parameters of this plasticity model could account for some of the diversity observed in experiments by Suvrathan and Raymond. He also presented a model with two plasticity sites, one in PF-to-PC synapses, another one in the cerebellar nuclei, that can reproduce behavioral and electrophysiological data in the VOR adaptation paradigm obtained in de Zeeuw’s lab
Cayco Gajic took up the theoretical challenge of describing how the specific architecture of the cerebellum could be useful for a variety of different tasks. After describing her own analysis of experimental data showing that MF activity is sparse and high-dimensional, as predicted by MAIT, she proposed that the real interest of the granule cell representation, beyond linear separability advocated by MAIT and more recent work, is to allow for universal function approximation. She illustrated how this could produce both stable and adaptable dynamics when coupling the cerebellum to motor cortex, modelled as a recurrent neural network.
In the final talk of the Seminar, De Zeeuw reminded the participants of the complex anatomy of the IO. He then reported the current work of his lab showing, using a discrimination task, that propagation of activity through loops involving the cerebellar cortex, DCN, and IO allow associations between different modalities and presumably coordination between different muscles.
A long discussion session between all participants concluded the Seminar. In the light of the different contributions, it appears that much work is still needed to decipher how the cerebellum repurposes its canonical circuit for different tasks. Progress in this direction will allow us to understand why and how the cerebellum is essential to fine-tune many behaviors, from simple and complex movements up to social interactions.
OpenEdition vous propose de citer ce billet de la manière suivante :
ldiebold (9 octobre 2024). Cervelet et apprentissage/ Cerebellum and learning. Les carnets de la Fondation des Treilles. Consulté le 11 février 2025 à l’adresse https://doi.org/10.58079/12fzx