| dc.description.abstract | The basal ganglia are a dynamic neural network of telencephalic subcortical nuclei,
involved in adaptive control of behaviour.
There has been much experimental evidence on the anatomy and physiology of the
basal ganglia published over the last 25 years showing that the basal ganglia are
involved in the learning of many adaptive behaviours, including motor planning,
working memory and cognitive functions. Current qualitative basal ganglia models
of the box and arrow type, whilst explaining much of the anatomical data, do not
give enough insight into the mechanisms involved in basal ganglia function either in
health or in disease states.
The striatum is the main input nucleus of the basal ganglia, integrating widespread
cortical and thalamic inputs to perform behaviour selection. Convergent data from
control theory learning models and experimental data have shown that the phasic
dopamine signal in the striatum could be performing a similar function to a scalar
teaching signal in reinforcement learning models, both signals indicating the
occurrence of reward. Similarly, both models and electrophysiological data have
shown how the timing of this reward signal can be changed during learning so as to
occur at the point in time of the earliest predictor of forthcoming reward. These
models do not, however, show how this teaching signal is used in the striatum to
learn to select the action most likely to lead to reward.
Computational models have been produced to investigate the circuitry involved in
striatal action selection. These models have tended to be of winner-takes-all
networks, using a mechanism of recurrent lateral inhibition between the medium
spiny cells of the striatum to select the winner and thus releasing the behavioural
action judged to be correct in the current environmental context. However, the
necessary biological circuitry to implement a winner-takes-all network is absent in
the striatum. This leads to a requirement for new models of striatal function
incorporating current biological data to provide a more realistic mechanism for
behavioural selection.
This thesis develops a biophysically inspired, minimal current model of a striatal
medium spiny neuron which utilises transitions between two membrane potential
states, both below firing threshold, to filter excitatory input. The behaviour of the
model is first validated against experimental electrophysiological data and then used
to demonstrate how the striatum could perform two of the tasks required for
behaviour selection; accurately timed release of behaviours and learning a sequence
of action selections to obtain reward. In the first series of simulations timed release
of behaviours is demonstrated to be linearly related to the timing of firing of
feed forward inhibitory interneurons. In a second set of simulations learned sequences
of action selection, using a simulated dopamine reward signal, are shown in a reward
location task performed by a small network of model medium spiny striatal neurons.
Taken together these studies show that this simple model of a striatal medium spiny
neuron is capable of simulating the basic functionality required for behaviour
selection in a manner which has greater biological plausibility than previously
published models. | en |
