Hands-On Machine Learning for Algorithmic Trading

Bidirectional RNNs

For some tasks, it can be realistic, necessary, and beneficial for the output to depend not only on past sequence elements as modeled so far, but also on future elements. Machine translation or speech and handwriting recognition are examples where subsequent sequence elements are useful and realistically available to disambiguate competing outputs.

For a one-dimensional sequence, bidirectional RNNs combine an RNN model that moves forward with another RNN that scans the sequence in the opposite direction. As a result, the output comes to depend on both the future and the past of the sequence. Applications to these domains have been very successful.

Bidirectional RNNs can also be used with two-dimensional image data. In this ...