Reviews: Phased LSTM: Accelerating Recurrent Network Training for Long or Event-based Sequences

LSTMs (and GRUs) are increasingly used as basic building blocks in neural network architectures, both in inherently sequential problems but also in other applications as many other problems can usefully be decomposed into sequential problems using mechanisms such as attention. Despite being devised some time ago, LSTMs have proved to be difficult to beat as a general purpose tools for modeling sequential structures (e.g. This paper presents an interesting idea for improving the performance of LSTMs, particularly on tasks which contain cyclical structure. It is novel and explains the model and motivations well. There are aspect of the analysis and experimental results which could be improved on, but it is a novel approach that will be of interest to the field. I have several suggestions for improvements below, but these do not significantly detract from the work, which is of a high standard.