When we implement the NN with dropout, we end up getting a higher accuracy on the test set compared to the train. We did not talk about the interpretation of this result in terms of bias and variance. Is this caused by the a degree of randomness between the contents of the test and train set? Or is there an interpretation in terms of bias and variance? Thanks! Love the course so far
The interpretation is that dropout (when tuned appropriately) has reduced the overfitting, which means there is somewhat less variance in the model. You hope for the “Goldilocks” amount of reduction in variance: just enough and not too much! 
1 Like