Dropout algorithms for recurrent neural networks
Nathan Watt, Mathys C. du Plessis · 2018
In the last decade, hardware advancements have allowed for neural networks to become much larger in size. Dropout is a popular deep learning technique which has shown to improve the performance of large neural networks. Recurrent neural networks are powerful networks specialised at solving problems which use time series data. Three different approaches to incorporating Dropout with recurrent neural networks have been suggested. However, these approaches have not been evaluated under identical experimental conditions. This article investigates the performance of these Dropout approaches using a 2D physics simulation benchmark. After applying statistical tests it was found that using Dropout did improve network performance on the benchmark. However, contrary to the literature, the Dropout approach which was expected to perform poorly, performed well, and the approach which was expected to perform well, performed poorly.