Pratik Y Chhatbar

Extraction of a reward expectation signal from cortical units following ballistic movements generated by a brain machine interface

Movement decoding algorithms used in todays brain-machine interface (BMI) technologies require movement-related neural activity in large quantities as training data to decode with sufficient accuracy the intended movements of the user. Because of physical disability the end users of BMI systems may be unable to readily provide such training data. Moreover, variability in the neural control of movements across patients with disability may result in individually unique training data. These issues limit the generalizability of movement decoding algorithms across BMI users. One potential method of circumventing this generalizability limitation and individualizing BMI technology is the use of reinforcement learning, a group of techniques that require minimal feedback in order to find solutions to an arbitrary problem. One promising means of providing feedback to a reinforcement learning-based BMI is via a neural reward signal found in multiple cortical and subcortical areas. Particularly attractive is the idea of parallel extraction of both the movement control signal and the reward signal from the same electrode array. We examined the neural signal underlying the expectation of reward depending on the probability of successfully reaching a target given the initial ballistic movement generated by a BMI. The real-time extraction of such signal could be used to determine if the user expects a movement generated by a BMI to succeed or fail. This information could then be used to update the control architecture of the BMI to generate an output more in line with the users intention.

Processing of a directionally dependent reward signal in motor and somatosensory units

Improving the control of neuroprosthetics to achieve biomimetic movements would significantly increase their utility and greatly improve the quality of life of their users. One potential addition to todays neuroprosthetics control systems would be an inclusion of the reward-based signal from motor or somatosensory cortex. The reward signal present in these cortices could indicate if a movement goal, such as reaching to and grasping a cup of coffee, was successful or not. Such a signal could be used as a component in reinforcement learning algorithms employed in brain-machine interfaces. This study seeks to determine the movement direction dependence of the reward signal. To accomplish this goal, we examined data recorded from the units of one bonnet macaque as it performed a center-out reaching task that has a fixed-probability reward assignment at the completion of a successful trial. By comparing the spiking neural activity with the actual receipt of reward, we examined if the change in firing activity can be attributed to the reward signal, and if this signal is also tied to the directionality of the movement. Including this information for reinforcement learning in brain-machine interfaces would bolster current efforts and lead to more realistic movement of neuroprosthetics.