Simulation of nervous centres in closed-loop of perception-action
Metadatos
Mostrar el registro completo del ítemEditorial
Universidad de Granada
Departamento
Universidad de Granada. Departamento de Arquitectura y Tecnología de ComputadoresMateria
Sistema nervioso Cerebelo Neurociencia computacional Biología computacional Inteligencia artificial Simulación Redes neuronales (Informática)
Materia UDC
681.3 330406
Fecha
2017Fecha lectura
2017-06-06Referencia bibliográfica
Naveros, F. Simulation of nervous centres in closed-loop of perception-action. Granada: Universidad de Granada, 2017. [http://hdl.handle.net/10481/47117]
Patrocinador
Tesis Univ. Granada. Programa Oficial de Doctorado en: Tecnología de la Información y la Comunicación; This work has been partially supported by the FPU national grant program, the European projects REALNET (FP7-270434) and HBP (FP7-604102), the Marie Curie grant (658479- Spike Control) and the French government research program “Investissements d'avenir through the Robotex Equipment of Excellence (ANR-10-EQPX-44)”. We gratefully acknowledge the support of NVIDIA Corporation with the donation of two Titan GPUs for current EDLUT development.Resumen
The human brain, thanks to the evolution, has developed efficient biological structures able to
perform a wide range of complex tasks. This is also the case of one of its centres: the
cerebellum. This structure plays a fundamental role in different motor control tasks such as
coordination of movements or calibration of sensorimotor relationship. In this thesis
computational models of the cerebellum have been developed in order to achieve better
understanding of the biological mechanisms that confer the cerebellum its motor control and
learning capabilities. Simulated and real robots have been used in this work as emulated
bodies to control. Thus we aim to validate several hypotheses about the cerebellum operation
when performing different motor control tasks such as object manipulation, eye blink classical
conditioning (EBCC) or vestibulo-ocular reflex (VOR) experiments.
The scope of this thesis is the development of biologically inspired control systems based in
cerebellar models able to perform different motor control tasks using biomorphic robots in
real time. This work can be subdivided in three main blocks: (i) development of all the tools
needed for this study, (ii) development of a cerebellar model based in data obtained with
biological experiments, and (iii) validation of the cerebellar models embedding these ones in
control schemes able to perform different motor control tasks with biomorphic robots in real
time.
An upgraded version of the EDLUT simulator has been used as the main simulation tool for this
study. EDLUT is an efficient spiking neural network simulator developed by our research group
at the University of Granada. It was conceived as a small tool capable of efficiently simulate
medium-scale spiking neural networks. In this thesis we have systematically improved the
efficiency and functionality of EDLUT. We have parallelized its simulation in multicore CPUGPU
co-processing architectures, developed new and efficient event-driven and time-driven
simulation methods and implemented new neuron models and learning rules. Additionally, we
have included new modules and features in EDLUT related with the robotic control in real
time. EDLUT-simulation spiking neural models can now connect with many (simulated or real)
robotic devices using TCP/IP connections. Communication interfaces able to translate the
cerebellar signals (spikes) in robotic signal (analogical signals) and vice versa have also been
implemented within EDLUT. Finally, EDLUT incorporates a real time supervisor able to ensure
that a simulation is performed in real time. Thus, EDLUT is now more than a simple spiking
neural network simulator. It is a simulation tool able to create biologically inspired control
schemes based in spiking neural networks to perform different motor control tasks using
biomorphic robots in real time.
Starting from a cerebellar model previously developed by our research group, two new
plasticity mechanisms at deep cerebellar nuclei (DCN) level have been proposed and
implemented, conferring to the cerebellar model with learning consolidation and gain
adaptation capabilities. We have also propose a new neural model for Purkinje cells able to
replicate its tri-modal spike modes (tonic, silence and bursting) and a new Inferior Olive (IO)
layer interconnected with electrical coupling able to better codify the error signal. Finally, a
new synaptic connection from the IO to the DCN cells have been proposed and included. All
these new elements, based in theoretical hypotheses and experimental results in the
literature, have increased the biological plausibility of our cerebellar model. Finally we have analysed how each one of the above-mentioned elements affects the
behaviour of the whole cerebellar model when performing motor control experiments as a
test-bench: a manipulation object task with a robotic arm, an EBCC experiment with a
simulated environment or a VOR experiment with a robotic head.