Fagg, AndrewLopez - Santillana, Manuel2020-08-062020-08-062020-07-30https://hdl.handle.net/11244/325367The transformation from visual stimulus to muscle recruitment is a non- linear one: it must take into account the configuration of the body and the lines of action of the muscles. The primary cortex plays a role in the mus- cle recruitment process. Any limb movement can be represented in multiple coordinate frames; a musculoskeletal representation, where forces are enacted through muscles, which will be referred to as an intrinsic representation, or by tracking the resulting visual movement, taking the perspective of a onlooker, providing an extrinsic representation of the movement. A coordinate system can also be drawn in respect to the joint orientation. How limb movements are represented in the primary motor cortex has been an intense focus of research. Georgopoulos et al. (1982) established that the activity of individual neu- rons in the primary cortex can be fit to a cosine curve, where the peak rep- resents an extrinsic direction for which they are maximally activated when performing limb movements. Due to the association of the neuron’s activity with an extrinsic direction, Georgopoulos et al. (1982) referred to the peak activation as the preferred direction (PD), favoring an extrinsic representation of neural activity. Later on, Kakei et al. (1999) focused on wrist movement and the corresponding neural activity in a center-out wrist task. Kakei et al. (1999) observed that the preferred direction of neurons in certain wrist postures: the supinated, the pronated, and the midrange. Through this observation Kakei et al. (1999) found that a subgroup of the neurons seem to shift the location of the preferred direction depending on the wrist posture. In this experiment, the neurons in the primary motor cortex (MI) region of the brain had their preferred direction shift less than the actual rotation of the wrist. Oby et al. (2012) expanded upon this, finding that the corresponding wrist force task also had PD shifts, with both authors concluding that the resulting PD shifts, on average, were less than the wrist rotation, with the average MI PD shift being less than the average wrist force PD. To model what sort of transformations are possible in the primary motor cortex, Craig (2013) devised a neural network model as an abstract represen- tation of this transformation process. Craig’s model is an recurrent neural network (RNN) that computes the transformation from visual input to forces, having no assumptions of the transformations that should occur. The model of Craig seeks to minimize the MSE of the forces produced and regularization, the activity of the RNN is based upon the minimization of the loss. In this thesis, I provide an extension of the original neural network model of force generation in a center-out wrist task that was developed by Craig (2013). The Craig’s model inputs are the extrinsic representation of the target, wrist configuration, and task cues. The model is tasked with generating the corresponding extrinsic wrist force trajectories given those inputs. My model takes advantage of the Python library of Keras (François, 2015), allowing for an ease of access to deep learning tools. The model is trained from the force trajectory of Todorov and Jordan (1998) allowing for a more modern approach by employing dropout and having a more accurate force depiction. My model also makes the distinction between target direction and the onset of forces. The disassociation between these two inputs allows for a better understanding of the mechanics and transformations that are associated with each cue. Incorporation of the supinated position into my model permits the model’s activity to be evaluated in regard to this posture. In my model, a sub population of neurons consistently exhibit cosine tuning through various neuron choices and regularization parameters. The sub pop- ulation was present for all postures, with the supinated posture consistently containing less stable neurons than the pronated and midrange postures Other neurons of the model exhibited activity shifts that mainly were time depen- dent, with little change in activity with respect to target location. The MI neurons did not exhibit a PD shift across changes in wrist posture. However, the muscles exhibited a significant difference in PD shift.Attribution 4.0 InternationalRecurrent Neural NetworksWrist ForcesMachine LearningPrimary Motor Cortex