Recurrent neural networks (RNNs) are useful to learn from paired input/target sequences, that have correlations between neighboring vectors. Regular RNNs have the disadvantage that they can use input information only from one side of the currently estimated output in the sequence. This talk introduces a bidirectional recurrent neural network structure (BRNN), that has shown to be most efficient for learning from time series with taking any available amount of input information into account. This new structure removes disadvantages of approaches that used fixed windows (MLP, TDNN) or only information from one side (RNN). This talk was first held in Feb. 1997. A paper describing BRNNs in detail appeared as:

Mike Schuster, Kuldip K. Paliwal, "BIDIRECTIONAL RECURRENT NEURAL NETWORKS", IEEE Transactions on Signal Processing, November 1997, Vol. 45, Page 2673-2681

Download the talk: (gzipped postscript, 12 pages, 40kB)