C5W1 - bidirectional RNN alternatives?

Course Q&A Deep Learning Specialization Sequence Models
1

Hello,
so I was wondering whether this version of “many to many” introduced in the course can also serve as bidirectional RNN

image
image708×288 47.4 KB

because from this schematic representation, we should have Tx+Ty+1 activations (considering only 1 layer), which is similar to what we have in bidirectional RNNs when Tx=Ty.
And also even the first y_hat will have full information on all the x sequences.

But maybe some drawbacks such as vanishing gradients?

2

Bidirectional nature gives the ability to look at the inputs from both ends. So, as long you have computational resources, this could be a good approach unless having unidirectional approach is sufficient on the problem.