טכניון מכון טכנולוגי לישראל
הטכניון מכון טכנולוגי לישראל - בית הספר ללימודי מוסמכים  
M.Sc Thesis
M.Sc StudentBeiser Koren
SubjectEnhanced Neural Modulations during Operation of a Brain
Machine Interfaces: Optimal Feedback Control
Approach
DepartmentDepartment of Mechanical Engineering
Supervisor Professor Miriam Zacksenhouse
Full Thesis textFull thesis text - English Version


Abstract

The primary motor-cortex (M1) is believed to function as a state-feedback controller. Its input is an estimation of the body's state and its outputs are the motor-commands sent to the relevant muscles. Electrical signals produced by nerve-cells (neurons) in M1 are therefore believed to encode the incoming state estimations.

A Brain-Machine-Interface (BMI) is designed to decode such neuronal signals in real-time and to control an artificial actuator accordingly.

The data analyzed in this work was taken from BMI experiments in which monkeys were controlling a moving cursor displayed on a computer screen. 

During the operation of the BMI, neural modulations in the monkeys' motor cortex were found to be significantly increased compared to when the monkeys performed the same task by manipulating a hand-held pole. Furthermore, these increased modulations were less correlated with the actual cursor-movement compared to the manual task.

In this work I suggest that these enhanced modulations can be explained by an increase in the variance of the estimated cursor-movement.

The estimated movement is assumed to be the end-product of a prediction-correction process implemented outside of M1. Predictions are internal simulations which are based on internal modeling of the body and copies of the motor commands it will actually receive. These predictions are then corrected using sensory feedback.

Once the state estimation process is finished, its end-product is fed into M1 where it will be processed in order to generate a motor command. The exact transformation from sate estimations into motor commands is governed by an optimality tradeoff between saving control-energy and reaching quickly for a rewarding target.  

According to my analysis, the internal model used during BMI operation is characterized by a lighter mass in comparison to the one used during the performance of actual hand movements. As a result, the internal control optimality tradeoff during BMI operation is driven towards preferring faster reaching movements which in turn generate larger neural modulations.

Furthermore, during BMI operation, only the significantly delayed visual feedback is available while the faster proprioreceptive feedback is missing. The absence of updated feedback during BMI operation causes state estimation to become even more dependent on internal predictions that are based on the "light-body" model.

The decreased correlations between actual cursor movement and neural activity can be explained by inaccuracies in the new internal model and by the lack of the fast propriorecptive feedback.