I was reading this paper which applies a modified version of the transformers for traffic forecasting. I am somewhat familiar with the transformer architecture and how it functions, but, in the paper, the authors are using scheduled sampling. For those who are not familiar with the concept, it basically means that sometimes you're feeding the decoder the ground truth and other times its own prediction.
How can such a mechanism work? As far as I know, during training, the decoder should be fed the ground truth shifted to the right by one (where a symbolic "start" token is introduced). How could a scheduled sampling fit in here?
The only way I could see is to modify the training procedure to resemble the inference/testing where we are predicting one future observation at a time, but this will increase the training time significantly.