Abstract
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.
Users
Please
log in to take part in the discussion (add own reviews or comments).