Depth-Adaptive Transformer

Maha Elbayad, Jiatao Gu, Edouard Grave +1

State of the art sequence-to-sequence models for large scale tasks perform a fixed number of computations for each input sequence regardless of whether it is easy or hard to proces…