English

Ultimate tensorization: compressing convolutional and FC layers alike

Machine Learning 2016-11-11 v1

Abstract

Convolutional neural networks excel in image recognition tasks, but this comes at the cost of high computational and memory complexity. To tackle this problem, [1] developed a tensor factorization framework to compress fully-connected layers. In this paper, we focus on compressing convolutional layers. We show that while the direct application of the tensor framework [1] to the 4-dimensional kernel of convolution does compress the layer, we can do better. We reshape the convolutional kernel into a tensor of higher order and factorize it. We combine the proposed approach with the previous work to compress both convolutional and fully-connected layers of a network and achieve 80x network compression rate with 1.1% accuracy drop on the CIFAR-10 dataset.

Keywords

Cite

@article{arxiv.1611.03214,
  title  = {Ultimate tensorization: compressing convolutional and FC layers alike},
  author = {Timur Garipov and Dmitry Podoprikhin and Alexander Novikov and Dmitry Vetrov},
  journal= {arXiv preprint arXiv:1611.03214},
  year   = {2016}
}

Comments

NIPS 2016 workshop: Learning with Tensors: Why Now and How?

R2 v1 2026-06-22T16:47:55.563Z