Reward and Aversion Representation in the Primary Somatosensory, Primary Motor, and Dorsal Premotor Cortices of Non-Human Primates Completing a Motor Task

Abstract

Signals of reward and punishment have been recorded in a number of areas in the brain. This thesis explores how these variables are represented in the hand and arm regions of the dorsal premotor (PMd), primary motor (M1), and primary somatosensory (S1) cortices. Two non-human primates (NHPs) were trained to complete a gripping task in a virtual robotic environment, where they manually applied and maintained a given level of force for a specified period of time. Prior to each trial, visual cues were displayed to inform the NHP if the trial would result in a juice reward if completed successfully, a time-out period punishment if completed unsuccessfully, or no reward or punishment, where the task would move immediately to the next trial. Uncued periods were recorded as well. Modulation due to punishment was observed in these three regions, with significant differences in firing rate between punishing and non-punishing trials observed in 4-17% of units post-cue and 4-27% of units post-result. While this represents a substantial subpopulation of units, this was less than the 30-64% of units post-cue and 6-32% of units post-result with significant differences between rewarding and non-rewarding trials. Linear models and non-linear divisive normalization models were used to investigate the representation of the interaction between reward and punishment. Divisive normalization models incorporating both reward and punishment performed better than linear models or divisive normalization models incorporating only reward or punishment. For cued trials, the best performing model showed motivational salience encoding in 47-52% of units with significant model fits in PMd, 40-75% in M1, and 41-80% in S1, with the remainder encoding valence. Dimensionality reduction analysis demonstrated a representation of both reward and punishment in these regions. Along with the success or failure of the trial, these factors were predictable with a gradient boosting classifier at greater than an adjusted chance accuracy, indicating that this information can be extracted and utilized. Investigating the intricacies of these signals in PMd, M1, and S1 will allow future brain-machine interfaces (BMIs) to capture the breadth of this activity in a limited number of cortical regions that also contain sensorimotor information, rather than requiring implants in multiple regions. Taking full advantage of the range of information in these regions will be useful in creating algorithms for more robust, nuanced, and naturalistic BMI control.