Weighted Regularization for Efficient Neural Network Compression

Yanxin Fu, Longxing Yang, Jilin Mei, Tiantian Jiang, Yan Xing, Wenxiao Zhao, Yu Hu

Keywords: Weighted Regularization, Neural Network Compression

TL;DR: A weighted regularization method for network compression is proposed and theoretical analysis is given.

Abstract: Regularization is widely applied to model complexity reduction and neural network compression. Existing $L_1$ and nuclear norm regularizations can achieve favorable results, but these methods treat all parameters equally and ignore the importance of the parameters. Taking the trained parameters as prior information to construct weights, a weighted regularization method is proposed in this paper. Theoretically, we establish the bounds on the estimation errors for values of the global minimum for a fully connected single hidden layer neural network. Further we prove the estimates generated from the weighted $L_1$ regularization and the weighted nuclear norm regularization can recover the sparsity and the low rank structure of a global minimum of the neural network with a high probability, respectively. The effectiveness of the algorithm is validated by conducting a numerical simulation and experiments with popular neural networks on public datasets from real-world applications.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Deep Learning and representational learning