also offered as (Old) Lecture 16 | Connectionist Temporal Classification
https://www.youtube.com/watch?v=A8IhGQCurPc&list=PLp-0K3kfddPzNdZPX4p0lVi6AcDXBofuf&index=16
image.png
pretend that the output shows up for more than 1 time, and divgergence is everywhere
however, this assumption may not hold in question answering: answer will not be reached before the question is completed
two assumptions here: output is order-synchronous with input, and output number is smaller than input number
just look at the row we are interested in
(unfinished)
