Research approach and objectives
Existing computational models in motor control and motor learning are important for understanding how the central nervous system performs and learns movements. They are suitable to unveil basic principles the nervous system uses to deal with all the complex variables (muscles, joints, etc.) when performing purposeful motor actions. However, it remains elusive how these models are biologically implemented. We propose that this should be studied on the level of neural networks. The advantage of this approach is that it is accessible from the computational and also from a biological perspective. Consequently, neural networks can be the common link to bridge computational and biological analysis.
The starting point of our modelling efforts will be existing computational models of motor control and motor learning (Shadmehr et al., 2010; Wolpert et al., 2011). We will consider different neuronal network models that have been assigned to functions outside the motor domain as candidate models for motor control. A good starting point is the random recurrent network model comprised of leaky integrate-and-fire neurons (Brunel, 2000) that has established itself as a de facto standard in cortical modelling. A hallmark of this model is the biologically realistic asynchronous-irregular activity dynamics that results from a selforganized state, where excitatory and inhibitory forces are in almost perfect balance. Variants of this model have, for instance, recently been successful in pinning down the detailed neuronal mechanisms leading to contrast-invariant feature selectivity in the primary visual cortex of rodents (Sadeh and Rotter, 2014a, b, 2015). Model parameters of the motor networks we will focus on will be calibrated within anatomical and physiological constraints using existing behavioural, neurophysiological and anatomical data as well as predictions from the aforementioned computational models. Model predictions will be compared to experimental data on the behavioural, neuronal and anatomical levels. In particular, predictions based on these network models may suggest new experiments for further validation or falsification. Comparisons between predictions and experimental results will be used to refine models or dismiss certain model architectures.