Abstract
Optimization methods are crucial to train deep neural networks. Adaptive optimization methods, especially Adam, are wildly used because they aren’t sensitive to the selection of learning rate and converge fast. Recent work point out Adam has a performance gap with SGD and even not converge because of the unstable and extreme learning rates. Many variants of Adam are proposed to solve the problem, such as AMSGrad, AdaBound and AdaBelief. In this paper, we propose a new variant of Adam, called AdaDiff. AdaDiff computing gradient descent step size by the exponential moving average(EMA) of gradient and differential of gradients aiming to make the training of networks more stable. We compare our method with other optimizers on various tasks. The results show AdaDiff outperforms Adam and minimizes the performance gap with SGD.
Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
