Faster Neural Network Training with Approximate Tensor Operations

Menachem Adelman, Kfir Yehuda Levy, Ido Hakimi, Mark Silberstein

Keywords: optimization for deep networks

TL;DR: We accelerate neural network training by using fast approximations for matrix multiplication and convolution

Abstract:

We propose a novel technique for faster deep neural network training which systematically applies sample-based approximation to the constituent tensor operations, i.e., matrix multiplications and convolutions. We introduce new sampling techniques, study their theoretical properties, and prove that they provide the same convergence guarantees when applied to SGD training.
We apply approximate tensor operations to single and multi-node training of MLP and CNN networks on MNIST, CIFAR-10 and ImageNet datasets. We demonstrate up to 66% reduction in the amount of computations and communication, and up to 1.37x faster training time while maintaining negligible or no impact on the final test accuracy.

Code Of Conduct: I certify that all co-authors of this work have read and commit to adhering to the NeurIPS Statement on Ethics, Fairness, Inclusivity, and Code of Conduct.

Code: https://github.com/acsl-technion/approx

Community Implementations: [ 3 code implementations](https://www.catalyzex.com/paper/arxiv:1805.08079/code)