26th Annual Computational Neuroscience Meeting (CNS*2017): Part 3

Abstract

Neuro computational models represent a powerful tool for bridging
\nthe gap between functions of the neural circuits and observable
\nbehaviors [1]. Once the model has been built, its output is compared
\nwith the observations either to validate the model itself or to propose
\nnew hypotheses. This approach has led to building a multi-scale
\nmodel of the sensorimotor system from muscles, proprioceptors to
\nskeletal joints, spinal regulating centers and central control circuits
\n[2–6].
\nIn this framework, we propose a neural network architecture to
\nsimulate the selection of actions performed by the motor cortex in
\nresponse to a sensory input during a reward-based movement learning.
\nThe network has as many input nodes as the number of different
\nstimuli, each node being a combination of the sensory inputs, and as
\nmany output nodes as the number of different actions that can be
\nperformed, each node being a combination of the motor commands.
\nThe network is fully connected, so that each stimulus concurs to the
\nselection of each action and each action is selected concurrently by
\nall the stimuli. The weights are updated by taking into account both
\nthe expected reward and the actual reward, as suggested in [7]. By
\nadopting this architecture, the percept is represented by a combination
\nof sensory inputs, while the action is represented by a combination
\nof motor commands. Thus, it reproduces faithfully the condition
\nof experiments of motor learning when a set of sensory inputs, such
\nas semantically neutral visual stimuli, are presented to the subject
\nwhose response is merely a motor action, such as pushing a button.
\nUnder such conditions, it then becomes possible to fit the data provided
\nby the experiments with the model to both estimate the validity
\nof the model and to infer the role of the parameter on behavioral
\ntraits.
\nThe simulations were compared to the behaviors of human subjects
\nwhile learning which out of two buttons to press in response to a collection
\nof visual stimuli containing edges and geometric shapes in a
\nreward based setting. The results showed that the behavior of the
\ncomplete system is the one expected under the hypothesis that the
\nreward acts by modulating the action selection triggered by the input
\nstimuli during motor learning. Moreover, differently from most literature
\nmodels, the learning rate varies with the complexity of the task,
\ni.e. the number of input stimuli. It can be argued that the decrease in
\nlearning rate seen in humans learning large set of stimuli could be due
\nto an attenuation of memory traces in real synapses over time. In our
\nfuture investigations, we will work to improve the model by adding
\nsuch an effect in our network.