Dear fellow mentors/classmates,
would like to seek for help to check my understanding.
I get the point the the RNN is at fault, so we need to address RNN rather than beam search.
Yet, the questions stated that we are evaluating the “DEV” set, NOT the training set, I wonder why we should look for tuning network architecture. In DLS course 3, we were told that we should split the data into training/dev/test set, and evaluate the network architecture in the training set, NOT the dev set. If the training set performance is poor, we should tune the network architecture. After we have make a proper network architecture, and go to evaluate in dev set and if the performance is still poor, we could look for solution such as getting more data in the training set or use regularization. Thus, I wonder why the explain in this quiz would say use a different network architecture in the DEV set could help in this case.