I have a question about the lecture in week 3 about Error analysis under “Machine learning development process”. Around the 5:13 mark, he says you can use more data and add more features to combat the misclassification of emails. Why is this the case?
I thought if the model misclassified the email it meant the model could not grasp the complexity of this problem, and thus it is a high-bias problem. And earlier he said a high bias problem is typically not solved by adding more data. I do understand why you would use more features.
This lecture on Error Analysis is showing a different method to diagnose a learning algorithm performance problem. Having a view of which categories of data are mis-classified most often, then, getting more data or engineering more features is a way to help the algorithm to learn and improve its accuracy.
To begin with, I think you are right that adding more data wouldn’t help a high-bias problem, and that’s exactly why we should doubt whether the model misclassifying emails to be due to high-bias if someone said adding data helped.
If you watch the video again at 6:34, 7:00, and 8:00, Andrew had repeatedly mentioned variance too (especially 8:00). I recommend you to go through that video again and this one too. Sometimes, reviewing a video twice in another day can give us a different view For example, how would we diagnose a bias and a variance problem? This is a must-clearly-know.
For example, how would we diagnose a bias and a variance problem? This is a must-clearly-know.