In plain words: Instead of plain gradient descent, this method estimates how the loss curves in many directions at once, so it takes smarter steps and handles huge batches. It reached 75% ImageNet accuracy even with extremely large batches, matching tuned first-order training while converging faster.
Abstract · Large-Scale Distributed Second-Order Optimization Using Kronecker-Factored Approximate Curvature for Deep Convolutional Neural Networks
Large-scale distributed training of deep neural networks suffer from the generalization gap caused by the increase in the effective mini-batch size. Previous approaches try to solve this problem by varying the learning rate and batch size over epochs and layers, or some ad hoc modification of the batch normalization. We propose an alternative approach using a second-order optimization method that shows similar generalization capability to first-order methods, but converges faster and can handle larger mini-batches. To test our method on a benchmark where highly optimized first-order methods are available as references, we train ResNet-50 on ImageNet. We converged to 75% Top-1 validation accuracy in 35 epochs for mini-batch sizes under 16,384, and achieved 75% even with a mini-batch size of 131,072, which took only 978 iterations.
Kazuki Osawa, Yohei Tsuji, Yuichiro Ueno, Akira Naruse, Rio Yokota, Satoshi Matsuoka
arXiv:1811.12019 · cs.LG, cs.CV, stat.ML · submitted Nov 29, 2018 · updated Mar 30, 2019
abstract · pdf · html · 10 pages, 7 figures. Accepted at CVPR 2019, Long Beach, CA
