Optimal Brain Damage 1989
This paper is the ground breaking work of network pruning of Model Compression.
From this paper, I know the procedure of OBD :
(1). Choose the reasonable network architecture
(2). Train the network until a good solution is obtained
(3): Compute the second derivative hkk for each parameter
(4): Compute the saliencies for each parameters: sk = hkk uk2/2
(5): Iterate to step 2
A new technique called OBD for reducing the size of a learning network seletively deleting weights.
Networks with too many weights do not generalize well. However, networks with too few weights will not have enough power to represent the data accurately.
One technique to reach this tradeoff is to minimize the cost function composed of two terms: