Dependent Bidirectional RNN with Extended-long Short-term Memory
In this work, we first conduct mathematical analysis on the memory, which is defined as a function that maps an element in a sequence to the current output, of three RNN cells; namely, the simple recurrent neural network (SRN), the long short-term memory (LSTM) and the gated recurrent unit (GRU). Based on the analysis, we propose a new design, called the extended-long short-term memory (ELSTM), to extend the memory length of a cell. Next, we present a multi-task RNN model that is robust to previous erroneous predictions, called the dependent bidirectional recurrent neural network (DBRNN), for the sequence-in-sequenceout (SISO) problem. Finally, the performance of the DBRNN model with the ELSTM cell is demonstrated by experimental results.
PDF Abstract