Long-term temporal averaging for stochastic optimization of deep neural networks

Deep learning models are capable of successfully tackling several difficult tasks. However, training deep neural models is not always a straightforward task due to several well-known issues, such as the problems of vanishing and exploding gradients. Furthermore, the stochastic nature of most of the used optimization techniques inevitably leads to instabilities during the training process, even when state-of-the-art stochastic optimization techniques are used. In this work, we propose an advanced temporal averaging technique that is capable of stabilizing the convergence of stochastic optimization for neural network training. Six different datasets and evaluation setups are used to extensively evaluate the proposed method and demonstrate the performance benefits. The more stable convergence of the algorithm also reduces the risk of stopping the training process when a bad descent step was taken and the learning rate was not appropriately set.