- Which computational models can be expressed as neural network models? There exist a number of computational models in motor control (e.g. forward predictive models 8 (Wolpert et al., 1995), optimal control (Todorov and Jordan, 2002), model-free reinforcement (Krakauer and Mazzoni, 2011)) and motor learning (e.g. structure learning (Braun et al., 2009a), learning of error sensitivity (Herzfeld et al., 2014)). We will investigate the possibilities and constraints of translating the assumptions made in these computational models into neural network models.
- Which constraints on the network models can be derived from existing knowledge on the anatomy and physiology of the motor system? When designing a biological neuronal network model one faces an enormous number of degrees of freedom: various architectures can be used to describe biological neural networks (e.g. feed-forward networks, all-to-all connected networks, random networks, small-world networks); each of these can be realized in various ways and even a specific architecture has many parameters which need to be specified. Therefore, it is crucial to identify possible anatomical and physiological constraints on the network architecture. To this end, existing experimental results on the anatomy and physiology of the motor system will be analysed. Such analysis can yield different possible conclusions: (i) existing literature provides sufficient constraints on network parameters; (ii) some brain areas related to motor control are more accessible for parsimonious modelling than others; (iii) existing knowledge is insufficient, but specific recommendations for further experiments can be made to facilitate future network modelling. Regardless of the outcome, this analysis shall therefore provide valuable insights and useful information for future research on the motor system.
- Which model predictions can be experimentally assessed and how? With the xperiment being the final judge on any model, we need to identify which predictions of the network models can be experimentally assessed using which methods. It will also be of particular interest here to pinpoint deviations between the predictions of any given computational model and the corresponding neuronal network model.
- Which neural network learning rules can mimic biological motor learning? Different plastic neuronal network architectures can be investigated with regards to motor learning. One potential approach will be to use networks with sparse dynamics as an efficient way to achieve low-dimensional representations of high-dimensional systems, which could for example mimic structural motor learning. Low-level synaptic plasticity rules (Bi and Poo, 1998; Gutig et al., 2003; Feldman, 2012) or high-level learning rules will be studied with respect to their potential to train these networks and constrain their dynamics to lowdimensional
manifolds or to learn activity dynamics trajectories (Sussillo and Abbott, 2009; Laje and Buonomano, 2013). In this context it will also be of particular interest to understand how the different time scales of motor learning (Smith et al., 2006; Kobak and Mehring, 2012) can be related to microscopic parameters of the neuronal network and network learning rules.
- Which motor tasks should be studied? In previous experiments, most studies validated their computational models based on behavioural data in reaching and grasping. For our approach, it is important to define the motor task(s) where biological variables can be assessed. The level of complexity of the defined tasks is related to the level of complexity that can be achieved by neural network models.