Why is reverse time RNN counter-intuitive?

[ThachNgocTran]

Hello @mp2893 ,

Your paper is interesting in the way it applies attentions to both visit-level and variable-level (within each visit). I'm reading it, and try to understand as much as possible. However, there is one thing I cannot grasp:

I believe: The "Reverse Time Attention Model" has something to do with the claim in the paper "attending the EHR data in a reverse time order so that recent clinical visits are likely to receive higher attention.". The RNN in the paper indeed goes reverse in time when processing the series of visits. This is, to me, counter-intuitive because it's like "first in first out": if RNN processes the last visit first (v_i), then goes backward in time to the first ever visit (v_1), at this point, it may have forgotten the last visit(s) (v_i, v_i-1...), which is important in predicting the medical codes (as per aforementioned claim).

Maybe I miss something in the paper? Or some assumption somewhere in the paper's references? Could you please give some explanation? Thank you. 😀