I’m trying to wrap my head around how this question is formulated.
Is it even correct? Read a few online articles about inverted dropout and I cannot get to the correct answer … I could simply iterate over all of them but I guess thats not the point. Could any of the mentors please chime in here? Thanks!!
Ahhh… I see the issue… Is states “at test time”… I totally missed that part …
1 Like
Exactly! That is the key point that the question is getting at. You never do any kind of regularization (dropout, L2, or whatever) at test time: you only do it at training time. That’s the point: regularization affects the results of training by giving you a different model than you would otherwise get. Then when you evaluate the performance of the resulting model on the test data, you just use the trained model as it is.
Another followup point is that you don’t even use regularization when you evaluate the training accuracy: you just use the trained model to predict.
1 Like