We present an efficient coresets-based neural network compression algorithm that sparsifies the parameters of a trained fully-connected neural network in a manner that provably approximates the network's output. Our approach is based on an importance sampling scheme that judiciously defines a sampling distribution over the neural network parameters, and as a result, retains parameters of high importance while discarding redundant ones. We leverage a novel, empirical notion of sensitivity and extend traditional coreset constructions to the application of compressing parameters. Our theoretical analysis establishes guarantees on the size and accuracy of the resulting compressed network and gives rise to generalization bounds that may provide new insights into the generalization properties of neural networks. We demonstrate the practical effectiveness of our algorithm on a variety of neural network configurations and real-world data sets.
|Title of host publication||International Conference on Learning Representations (ICLR) 2019|
|State||Published - 2019|
|Event||7th International Conference on Learning Representations, ICLR 2019 - New Orleans, United States|
Duration: 6 May 2019 → 9 May 2019
|Conference||7th International Conference on Learning Representations, ICLR 2019|
|Period||6/05/19 → 9/05/19|
Bibliographical noteFunding Information:
This research was supported in part by the National Science Foundation award IIS-1723943. We thank Brandon Araki and Kiran Vodrahalli for valuable discussions and helpful suggestions. We would also like to thank Kasper Green Larsen, Alexander Mathiasen, and Allan Gronlund for pointing out an error in an earlier formulation of Lemma 6.
© 7th International Conference on Learning Representations, ICLR 2019. All Rights Reserved.
ASJC Scopus subject areas
- Computer Science Applications
- Linguistics and Language
- Language and Linguistics