I wonder if you have examples of situations where it is ok to only use train/dev sets and not train/dev/test sets? Andrew says in the video that it might be ok in situations where you do not need to have an unbiased estimate of the performance of your model. In what situation would you not want an unbiased estimate of the performance?
i will try to answer you , i think it’s depends on the main knowledge domain you are interested in . Sometimes , generalization it’s not something that easy and we’re not sure it’s there that’s why we need another test sets with train/dev to have more accuracy that we are on the correct path.
therefore, for example on the environmental domain the scale is important factor that’s make generalization more difficult so only train/dev it’s not enough
i hope that’s helps you a little with your question
Thanks for your input
I get the importance of generalisation, and so the importance of a test set, but I struggle imagining concrete situations where it would be acceptable not to try to make sure we have an unbiased estimate of the performance.
I understand that sometimes generalisation is very hard or practically impossible to reach.
I guess another reason might be the number of examples you have access to. If you do not have a lot of them, building a test set on top of the train/dev set will be complicated.