Practical Time Series Analysis

Bi-directional recurrent neural networks

The RNN discussed so far is unidirectional and traverses along the direction of the original time series. However, in many cases, capturing the sequential information and memory in the reversed series improves prediction. Such an RNN that uses both forward and backward traversal is called a bi-directional RNN that improves the ability of the network to capture memory over long ranges. The following figure shows a bi-directional RNN:

Figure 5.17: Bi-directional recurrent neural network with p timesteps