I’ve found an error in the grader for the W1 quiz question “You have finished training a language model RNN and are using it to sample random sentences, as follows: [image of RNN] What are you doing at each time step t?”
The version of the quiz I used has this question as #5. It is a multiple select question, and I selected the correct answer, which states “(i) Use the probabilities output by the RNN to pick the highest probability word for that time-step as y-hat-t (ii) Then pass the ground-truth word from the training set to the next time-step.”
The grader, however, marked this answer as incorrect and in the explanation stated “Incorrect. The probabilities output by the RNN are not used to pick the highest probability word and the ground-truth word from the training set is not the input to the next time-step.”
Edit: Okay, now I’m going crazy. This also happened for the True-False version of the question. “True/False: In this sample sentence, step t uses the probabilities output by the RNN to pick the highest probability word for that time-step. Then it passes the ground-truth word from the training set to the next time-step.” I answered True.
For the sake of my understanding, can you please either acknowledge that I chose the correct answer or explain why the answer I chose is incorrect?
Both statements don’t hold truth as at each time t, it uses the probability output by the RNN to randomly sample a chosen word for that time step as y-hat. Then pass this selected word to the next time-step.
Right: the point is that what happens at training time is different than what happens at “inference” time, meaning when you use the trained model to generate samples. The multiple choice question is asking about inference time, not training time. You probably haven’t given the full context here, but my guess is that the T/F version of the question is asking about training time.
Oh my gosh. Thank you @Deepti_Prasad and @paulinpaloalto. I see. Apologies, I really thought I was going crazy. Now I get it. As you said, this is inference time sampling. Of course. Thank you both.