Scheduled Sampling

A sequence model predicting tomorrow's value from a history of past values is normally trained with "teacher forcing": at each step, it's fed the true previous value, even if its own prediction for that step was off. That makes training stable and fast, but it creates a mismatch — at real prediction time, the model has no true past to lean on beyond the point it started forecasting, so it must feed its own (possibly imperfect) predictions back in as input for the next step. A model trained only ever seeing perfect inputs can be fragile once its own small errors start compounding, a problem often called exposure bias.

Scheduled sampling closes that gap by randomly substituting the model's own previous prediction for the true value during training, with the probability of doing so increasing as training progresses. Early on, the model mostly sees ground truth (so it can still learn the basic task); later, it increasingly sees its own outputs, forcing it to get comfortable recovering from its own small mistakes — much closer to the real deployment condition. The "schedule" refers to this probability curve, which typically starts near zero and ramps up toward one over training.

For instance, a schedule might use true inputs 90% of the time in early epochs, dropping to 50% by the midpoint of training, and down to 10% by the end — so the model spends its final training stretch mostly practicing under realistic, self-generated conditions.