Reviews: Learning to Prune Deep Neural Networks via Layer-wise Optimal Brain Surgeon

Summary: This paper adapts Optimal Brain Surgeon (OBS) method to a local version, and modified the objective function to be the target activation per each layer. Similar to OBS, it uses an approximation to compute Hessian inverse by running through the dataset once. Compare to prior methods, it finishes compression with much less retraining iterations. A theoretical bound on the total error based on local reconstruction error is provided. Pros: - The paper explores a local version of OBS and shows effectiveness of proposed method in terms of less time cost for retraining the pruned network.