I typically think of teacher forcing as optional when training an RNN. We may either:

  • use the output of time-step $t$ as the input to time-step $t+1$

  • use the $(t+1)$th input as the input to time-step $t+1$

When I actually sat down to write a bidirectional RNN from scratch today I realised it would be impossible to do without 100% teacher forcing, because each time step needs access to the "history" going back to the 0th time-step (forward direction) and going back (or forward - however you want to think of it) to the last time-step (backward direction).

Is that correct?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.