The method is also appropriate for non-stationary objectives and problems with very noisy and/or sparse gradients. Diederik P. Kingma, et al., "Adam: A Method for Stochastic Optimization" https://arxiv.org/abs/1412.6980 学習を最適化させる…