- motivation: we want to be info from the future, which isn't possible with vanilla [[Recurrent Neural Networks|RNNs]]
- for each input word, you create a forward component (purple box) & backward component(green component)
- connecting them both to the output at that timestep enables a decision informed by all words before & after that point
![[CleanShot 2024-07-09 at
[email protected]|400]]