Abstract

In this paper, we will show an unprecedented method to accelerate trainingand improve performance, which called random gradient (RG). This method can beeasier to the training of any model without extra calculation cost, we useImage classification, Semantic segmentation, and GANs to confirm this methodcan improve speed which is training model in computer vision. The central ideais using the loss multiplied by a random number to random reduce theback-propagation gradient. We can use this method to produce a better result inPascal VOC, Cifar, Cityscapes datasets.

 

Quick Read (beta)