Java Implementation of Optimal Brain Surgeon ECE 539 Project Curtis Christie
What Is OBS Optimal Brain Surgeon Takes a relatively large trained neural network Calculates which weights can be removed with out introducing a lot of error Adjusts remaining weight to account for removal of a weight OBS can reduce number of weights by 70-90%
My Project Java Implementation of OBS Inputs Outputs Input Vector Trained Weights Partial Derivate’s of F(x,w) in terms of w Outputs New Weights
How It Works Calculates the inverse Hessian matrix using recursion Finds Saliency of each weight Increase in error that results when the weight is eliminated Finds lowest Saliency of all the weights, determines if weight should be removed Adjusts all weights for the removal of each a weight Repeats until no more weights can be removed
Progress Accomplished Remaining Found Hessian Inverse Found minimum Saliency Remaining Remove weight Adjust remaining weights Run test Networks