This Ph.D. research project bridges neuroscience and bioengineering, aiming to explore the brain neural mechanisms through a neuro-computational approach. Indeed, computational neural models are fundamental to explain how ensemble brain functions might emerge from elementary neuronal components. Since the entire nervous system is too complex and wide to be studied as a whole in a PhD-thesis-span, the research is circumscribed to the cerebellum, because of its crucial learning skills. The project is focused on the generation and tuning of cerebellar computational models, which had been gradually refined to increase their biological realism. Eventually, these computational models have been embedded within simulated and real robotic platforms and challenged in various closed loop protocols. The models were developed as realistic Spiking Neural Networks (SNN) and have been tested in different sensorimotor tasks: Eye Blink Classical Conditioning (EBCC), Vestibulo-Ocular Reflex (VOR), movements perturbed by force fields, and motor correction. Eventually, fitting real experimental data coming from human or animal subjects. A mechanistic quantitative interpretation of the dynamic evolution of cerebellar plasticity during skill acquisition is still lacking, and computational models can be very effective to provide new insights. Thus, an extensive analysis of the relations between model parameters and behavioral outputs have been exploited. As a first step, a realistic model was developed to test neurophysiological theories, such as the confirmation of the role of multiple plasticity sites in cerebellar learning. The simulations demonstrated that the addition of nuclear plasticity sites improved model performance and showed the different timescales of learning. We validated the model robustness in learning associative responses with different inter stimuli intervals and we have shed light on acquisition, extinction and consolidation mechanisms, associable to the different active plasticity sites. As a second step, we used the cerebellar model to generate expectations from potential ad-hoc experiments and to generate hypotheses about the mechanisms underlying behavioral components and modifications. By manipulating the model parameters, it was possible to tune the model to fit the behavior of either healthy subjects or subjects perturbed by Transcranial Magnetic Stimulation (TMS). We have shown that a macroscopic measurement during a behavioral task can be successfully explained by using an appropriate model constructed at the microscopic level. In particular, we have been able to tune the cerebellar model against human EBCC data before and after perturbation with TMS, in two different protocols. The results supported the emerging view that cerebellar plasticity is a dynamic and distributed process, in which cortical plasticity is more rapidly activated and drives a set of changes that reverberate onto the more slowly adapting nuclei. Changes or deficiencies occurring in one site are compensated by the others suggesting possible interventional sites for therapy and repair. The model predictions are warranted future experimental investigations, e.g. performing \textit{in vivo} multi-electrode recordings of the plastic evolution of neural discharges in cortical and deep cerebellar nuclei neurons. As a third step, the neural network model was challenged in different learning paradigms, integrated with robotic platforms. These provided bodies to the simulated "little" brain, closing the loop between action and perception. As a result, we have efficiently linked low-level cerebellar circuits with high-level functions integrating a SNN into a neurorobot operating in real-time. The main added value of learning in our neurorobot is that it succeeded in reproducing how biological systems acquire, extinguish and express knowledge of a noisy and changing world, in multiple cerebellum-driven learning tasks performed by a real robot moving in perturbed environments. The real world is always noisier than the worst case simulation can accomplish, and learning is actually a long-lasting experience-dependent change in behavior, which can realistically be observed only from an embodied system. This approach has the challenging potential to represent the initial building block of a more complex brain-inspired controller, into which other realistic SNNs emulating different brain structures involved in sensorimotor integration could be embedded. During the project development, multiple features have been implemented in the model, taking inspiration from physiology, in order to increase its realism and potentiality. The network size was increased, including geometrical information and plausible convergence/divergence ratios and plastic mechanisms. We have exploited a new SNN simulator, the NEST simulator, a popular and more advanced tool, improving the realism of the model, including new neural populations and synapses. Leveraging High-Performance Computing systems, we succeed in generating a full reconstruction of the mouse flocculus (almost 1 million neurons and 360 million synapses) and simulating it in two learning tasks. The technological platform that has been developed in this project can lead to accelerated computational brain research since we have demonstrated the use of an effective tool-flow in the research of complex brain models. This platform showcases a paradigm for large-scale simulations that can be adapted to any other neuron modeling problems, greatly accelerating the scientific process and the development of brain research.

Questo progetto di ricerca unisce le neuroscienze e la bioingegneria, con l'obiettivo di esplorare i meccanismi cerebrali con un approccio neuro-computazionale. Infatti, i modelli neurali computazionali sono fondamentali per spiegare come le funzioni cerebrali possano emergere dalle componenti neurali elementari. Poiché l'intero sistema nervoso è troppo complesso e ampio per essere studiato interamente, nell'arco di una tesi di dottorato, la ricerca è stata circoscritta al cervelletto. Questo sistema è stato scelto a causa delle sue particolari capacità di apprendimento motorio. Il progetto è incentrato sulla generazione e messa a punto di modelli computazionali cerebellari, che sono stati gradualmente perfezionati per aumentare il loro realismo biologico. Infine, questi modelli computazionali sono stati integrati in piattaforme robotiche, sia simulate sia reali, e testati in vari protocolli in anello chiuso (closed-loop). I modelli sono stati sviluppati tramite reti neurali artificiali dette Spiking Neural Networks (SNN) e sono stati testati in diverse attività sensorimotorie: il condizionamento dell'ammiccamento (Eye Blink Classical Conditioning), il riflesso vestibulo oculare (VOR), movimenti perturbati da campi di forza e correzione del piano motorio. Inoltre, sono stati modellati dati sperimentali reali provenienti da soggetti umani e animali. Non è ancora stata raggiunta un'interpretazione quantitativa e meccanicistica dell'evoluzione dinamica della plasticità cerebellare durante l'apprendimento motorio. In questa direzione, i modelli computazionali possono essere molto efficaci per fornire nuovi suggerimenti. Pertanto, è stata sfruttata un'analisi approfondita delle relazioni tra parametri del modello e gli output comportamentali. Durante lo sviluppo del progetto, sono state implementate molteplici caratteristiche nel modello, ispirandosi alla fisiologia, al fine di aumentarne il realismo e le potenzialità. La dimensione della rete è stata aumentata, includendo informazioni geometriche e rapporti di convergenza/divergenza plausibili e meccanismi plastici distribuiti. Sfruttando i sistemi di calcolo ad alte prestazioni, abbiamo generato una ricostruzione completa del flocculo del topo (quasi 1 milione di neuroni e 360 milioni di sinapsi) e di simularlo in due attività di apprendimento motorio. La piattaforma tecnologica che è stata sviluppata in questo progetto può portare ad un'accelerazione nella ricerca nel settore delle neuroscienze. Infatti, abbiamo dimostrato che l'uso di modelli computazionali è uno strumento indispensabile per lo studio di processi cerebrali complessi.

Computational cerebellar models and their embodiment in behavioral loops to understand neural bases of motor learning

ANTONIETTI, ALBERTO

Abstract

This Ph.D. research project bridges neuroscience and bioengineering, aiming to explore the brain neural mechanisms through a neuro-computational approach. Indeed, computational neural models are fundamental to explain how ensemble brain functions might emerge from elementary neuronal components. Since the entire nervous system is too complex and wide to be studied as a whole in a PhD-thesis-span, the research is circumscribed to the cerebellum, because of its crucial learning skills. The project is focused on the generation and tuning of cerebellar computational models, which had been gradually refined to increase their biological realism. Eventually, these computational models have been embedded within simulated and real robotic platforms and challenged in various closed loop protocols. The models were developed as realistic Spiking Neural Networks (SNN) and have been tested in different sensorimotor tasks: Eye Blink Classical Conditioning (EBCC), Vestibulo-Ocular Reflex (VOR), movements perturbed by force fields, and motor correction. Eventually, fitting real experimental data coming from human or animal subjects. A mechanistic quantitative interpretation of the dynamic evolution of cerebellar plasticity during skill acquisition is still lacking, and computational models can be very effective to provide new insights. Thus, an extensive analysis of the relations between model parameters and behavioral outputs have been exploited. As a first step, a realistic model was developed to test neurophysiological theories, such as the confirmation of the role of multiple plasticity sites in cerebellar learning. The simulations demonstrated that the addition of nuclear plasticity sites improved model performance and showed the different timescales of learning. We validated the model robustness in learning associative responses with different inter stimuli intervals and we have shed light on acquisition, extinction and consolidation mechanisms, associable to the different active plasticity sites. As a second step, we used the cerebellar model to generate expectations from potential ad-hoc experiments and to generate hypotheses about the mechanisms underlying behavioral components and modifications. By manipulating the model parameters, it was possible to tune the model to fit the behavior of either healthy subjects or subjects perturbed by Transcranial Magnetic Stimulation (TMS). We have shown that a macroscopic measurement during a behavioral task can be successfully explained by using an appropriate model constructed at the microscopic level. In particular, we have been able to tune the cerebellar model against human EBCC data before and after perturbation with TMS, in two different protocols. The results supported the emerging view that cerebellar plasticity is a dynamic and distributed process, in which cortical plasticity is more rapidly activated and drives a set of changes that reverberate onto the more slowly adapting nuclei. Changes or deficiencies occurring in one site are compensated by the others suggesting possible interventional sites for therapy and repair. The model predictions are warranted future experimental investigations, e.g. performing \textit{in vivo} multi-electrode recordings of the plastic evolution of neural discharges in cortical and deep cerebellar nuclei neurons. As a third step, the neural network model was challenged in different learning paradigms, integrated with robotic platforms. These provided bodies to the simulated "little" brain, closing the loop between action and perception. As a result, we have efficiently linked low-level cerebellar circuits with high-level functions integrating a SNN into a neurorobot operating in real-time. The main added value of learning in our neurorobot is that it succeeded in reproducing how biological systems acquire, extinguish and express knowledge of a noisy and changing world, in multiple cerebellum-driven learning tasks performed by a real robot moving in perturbed environments. The real world is always noisier than the worst case simulation can accomplish, and learning is actually a long-lasting experience-dependent change in behavior, which can realistically be observed only from an embodied system. This approach has the challenging potential to represent the initial building block of a more complex brain-inspired controller, into which other realistic SNNs emulating different brain structures involved in sensorimotor integration could be embedded. During the project development, multiple features have been implemented in the model, taking inspiration from physiology, in order to increase its realism and potentiality. The network size was increased, including geometrical information and plausible convergence/divergence ratios and plastic mechanisms. We have exploited a new SNN simulator, the NEST simulator, a popular and more advanced tool, improving the realism of the model, including new neural populations and synapses. Leveraging High-Performance Computing systems, we succeed in generating a full reconstruction of the mouse flocculus (almost 1 million neurons and 360 million synapses) and simulating it in two learning tasks. The technological platform that has been developed in this project can lead to accelerated computational brain research since we have demonstrated the use of an effective tool-flow in the research of complex brain models. This platform showcases a paradigm for large-scale simulations that can be adapted to any other neuron modeling problems, greatly accelerating the scientific process and the development of brain research.
ALIVERTI, ANDREA
REDAELLI, ALBERTO CESARE LUIGI
CASELLATO, CLAUDIA
10-mag-2018
Questo progetto di ricerca unisce le neuroscienze e la bioingegneria, con l'obiettivo di esplorare i meccanismi cerebrali con un approccio neuro-computazionale. Infatti, i modelli neurali computazionali sono fondamentali per spiegare come le funzioni cerebrali possano emergere dalle componenti neurali elementari. Poiché l'intero sistema nervoso è troppo complesso e ampio per essere studiato interamente, nell'arco di una tesi di dottorato, la ricerca è stata circoscritta al cervelletto. Questo sistema è stato scelto a causa delle sue particolari capacità di apprendimento motorio. Il progetto è incentrato sulla generazione e messa a punto di modelli computazionali cerebellari, che sono stati gradualmente perfezionati per aumentare il loro realismo biologico. Infine, questi modelli computazionali sono stati integrati in piattaforme robotiche, sia simulate sia reali, e testati in vari protocolli in anello chiuso (closed-loop). I modelli sono stati sviluppati tramite reti neurali artificiali dette Spiking Neural Networks (SNN) e sono stati testati in diverse attività sensorimotorie: il condizionamento dell'ammiccamento (Eye Blink Classical Conditioning), il riflesso vestibulo oculare (VOR), movimenti perturbati da campi di forza e correzione del piano motorio. Inoltre, sono stati modellati dati sperimentali reali provenienti da soggetti umani e animali. Non è ancora stata raggiunta un'interpretazione quantitativa e meccanicistica dell'evoluzione dinamica della plasticità cerebellare durante l'apprendimento motorio. In questa direzione, i modelli computazionali possono essere molto efficaci per fornire nuovi suggerimenti. Pertanto, è stata sfruttata un'analisi approfondita delle relazioni tra parametri del modello e gli output comportamentali. Durante lo sviluppo del progetto, sono state implementate molteplici caratteristiche nel modello, ispirandosi alla fisiologia, al fine di aumentarne il realismo e le potenzialità. La dimensione della rete è stata aumentata, includendo informazioni geometriche e rapporti di convergenza/divergenza plausibili e meccanismi plastici distribuiti. Sfruttando i sistemi di calcolo ad alte prestazioni, abbiamo generato una ricostruzione completa del flocculo del topo (quasi 1 milione di neuroni e 360 milioni di sinapsi) e di simularlo in due attività di apprendimento motorio. La piattaforma tecnologica che è stata sviluppata in questo progetto può portare ad un'accelerazione nella ricerca nel settore delle neuroscienze. Infatti, abbiamo dimostrato che l'uso di modelli computazionali è uno strumento indispensabile per lo studio di processi cerebrali complessi.
Tesi di dottorato
File allegati
File Dimensione Formato  
20180420_ANTONIETTI_ALBERTO_PhD_Thesis.pdf

solo utenti autorizzati dal 21/04/2021

Descrizione: Thesis Manuscript
Dimensione 29.68 MB
Formato Adobe PDF
29.68 MB Adobe PDF   Visualizza/Apri

I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.

Utilizza questo identificativo per citare o creare un link a questo documento: https://hdl.handle.net/10589/139681