Neural network methods for error canceling in human-machine manipulation

Abstract

A neural network technique is employed to cancel hand motion error during microsurgery. A cascade-correlation neural network trained via extended Kalman filtering was tested on 15 recordings of hand movement collected from 4 surgeons. The neural network was trained to output the surgeon's desired motion, suppressing erroneous components. In experiments this technique reduced the root mean square error (rmse) of the erroneous motion by an average of 39.5%. This was 9.6% greater than the reduction achieved in earlier work, which followed the complementary approach of estimating the error rather than the desired component. Preliminary results are also presented from tests in which training and testing data were taken from different surgeons.