A Question for Week 1 Quiz

Course Q&A Deep Learning Specialization Improving Deep Neural Networks: Hyperparameter tun
1

æˆȘć±2024-12-27 例捈4.17.22
æˆȘć±2024-12-27 例捈4.17.222068×522 46.3 KB

I would like to point out a potential issue with the phrasing of this question. It states that to reduce high variance, the regularization hyperparameter lambda MUST be increased. However, this wording could be misleading, as increasing lambda is not the only method to address high variance (overfitting).
Or is it my misunderstanding of this question ?

2

The question here clearly states to establish a relation if increase lambda can reduce high variance, to which the answer should be true as lambda is a regularisation term which is function of the weights. So increasing your lambda going to increase your regularization term, there by addressing the variance issue.

question wasn’t asking to find ways to reduce high variance. Question is very direct in terms of lambda hyperparameter.

1 Like
3

Thanks for your reply. However I would take that the word “MUST” is too strong that indicates there’s no other way around for high variance. It may states like “lambda could be increased” to reduce misleading.

4

the must is again in relation to hyperparameter lambda to high variance issue to be address.

The must is not mentioned as high variance can only be address, then you must increase the lambda.

The must is related to lambda hyperparameter in order to reduce high variance issue.

when a question is asked, it clearly divided into two parts, the first part to reduce high variance and the second part mentions regularisation hyperparameter lambda which should be addressed by increasing and it is a must condition if one has to reduce high variance if or I should say the question states clearly to lambda hyperparameter only here.

5

But as Deepti says, the point is that increasing lambda will reduce variance. So your proposed “could” phrasing can also be viewed as incorrect. The other edge of your “must” sword. :laughing: If you make the case that “must” is too strong, then “could” is too weak. So what is the “Goldilocks” wording in this case? How about this:

If you are using L2 regularization to address overfitting, then you should increase lambda in order to reduce variance. T or F?

The other thing to note here is that no harm is really done in that you can take the quizzes as many times as you want, although only 3 attempts in any 8 hour period.