Skip to content
ML Academy

Introduced on day 37

Teacher forcing

Training a sequence decoder by feeding it the true previous output at each step instead of its own prediction.

In detail

It makes each step a one-step prediction from correct context, which trains fast and stably. At inference the decoder receives its own outputs, so the training and inference inputs differ (exposure bias). Evaluate multi-step use in free-running mode.