English

Stochastic Gradient Methods with Layer-wise Adaptive Moments for Training of Deep Networks

Machine Learning 2020-02-10 v3 Machine Learning

Abstract

We propose NovoGrad, an adaptive stochastic gradient descent method with layer-wise gradient normalization and decoupled weight decay. In our experiments on neural networks for image classification, speech recognition, machine translation, and language modeling, it performs on par or better than well tuned SGD with momentum and Adam or AdamW. Additionally, NovoGrad (1) is robust to the choice of learning rate and weight initialization, (2) works well in a large batch setting, and (3) has two times smaller memory footprint than Adam.

Keywords

Cite

@article{arxiv.1905.11286,
  title  = {Stochastic Gradient Methods with Layer-wise Adaptive Moments for Training of Deep Networks},
  author = {Boris Ginsburg and Patrice Castonguay and Oleksii Hrinchuk and Oleksii Kuchaiev and Vitaly Lavrukhin and Ryan Leary and Jason Li and Huyen Nguyen and Yang Zhang and Jonathan M. Cohen},
  journal= {arXiv preprint arXiv:1905.11286},
  year   = {2020}
}

Comments

Preprint, under review