10 March 2022 in IEEE Transactions on Neural Networks and Learning Systems
AdaBelief, one of the current best optimizers, demonstrates superior generalization ability over the popular Adam algorithm by viewing the exponential moving average of observed gradients. AdaBelief is theoretically appealing in which it has a data-dependent O(√T) regret bound when objective functions are convex, where T is a time horizon. It remains, however, an open problem whether the convergence rate can be further improved without sacrificing its generalization ability. To this end, we make the first attempt in this work and design a novel optimization algorithm called FastAdaBelief that aims to exploit its strong convexity in order to achieve an even faster convergence rate. In particular, by adjusting the step size that better considers strong convexity and prevents fluctuation, our proposed FastAdaBelief demonstrates excellent generalization ability and superior convergence. As an important theoretical contribution, we prove that FastAdaBelief attains a data-dependent O(log T) regret bound, which is substantially lower than AdaBelief in strongly convex cases. On the empirical side, we validate our theoretical analysis with extensive experiments in scenarios of strong convexity and nonconvexity using three popular baseline models. Experimental results are very encouraging: FastAdaBelief converges the quickest in comparison to all mainstream algorithms while maintaining an excellent generalization ability, in cases of both strong convexity or nonconvexity. FastAdaBelief is, thus, posited as a new benchmark model for the research community.
11 January 2022 in Cognitive Computation
23 May 2022 in IEEE Transactions on Emerging Topics in Computational Intelligence
The training process for deep learning and pattern recognition normally involves the use of convex and strongly convex optimization algorithms such as AdaBelief and SAdam to handle lots of “uninformative” samples that should be ignored, thus incurring extra calculations. To solve this open problem, we propose to design bandit sampling method to make these algorithms focus on “informative” samples during training process. Our contribution is twofold: first, we propose a convex optimization algorithm with bandit sampling, termed AdaBeliefBS, and prove that it converges faster than its original version; second, we prove that bandit sampling works well for strongly convex algorithms, and propose a generalized SAdam, called SAdamBS, that converges faster than SAdam. Finally, we conduct a series of experiments on various benchmark datasets to verify the fast convergence rate of our proposed algorithms.