Optimal Brain Surgeon Algorithm Payne Y. Chang
What is OBS(Optical Brain Surgeon) ? To solve real-world problems with ANN : Usually Need Highly connected Larger Not alwayse good OBS algorithm An approach to prune the ANN Goal : 1. Set some synaptic weights to zero 2. Modify the remaining weights 3. Maintain good performance
How does OBS work ? 1. Train the MLP 2. Calculate the saliency ( Si ) of each synaptic weight ( wi ) Si : The increase in the cost function from the deletion of wi 3. If Si is smaller than a threshold, delete wi 4. Go to step 2 5. Stop when no more weights can be deleted
Program : OBS.exe 1. C++ 2. FLTK Input Output Hidden N Bias (1) Deleted w
Results and Summary 1. Efficient ( Time & Storage Saving ) Learning Rate = 0.01 Momentum = 0.8 Activation function : Hyperbolic tangen 1. Efficient ( Time & Storage Saving ) 2. Better Generalization