Machine Learning

Recent Advances in Recurrent Neural Networks

Recurrent neural networks (RNNs) are capable of learning features and long term dependencies from sequential and time-series data. The RNNs have a stack of non-linear units where at least one connection between units forms a directed cycle. A well-trained RNN can model any dynamical system; however, training RNNs is mostly plagued by issues in learning long-term dependencies. In this paper, we present a survey on RNNs and several new advances for newcomers and professionals in the field. The fundamentals and recent advances are explained and the research challenges are introduced.

Recent Advances in Recurrent Neural Networks
by Hojjat Salehinejad, Julianne Baarbe, Sharan Sankar, Joseph Barfett, Errol Colak, Shahrokh Valaee
https://arxiv.org/pdf/1801.01078v2.pdf