Prof. Ng makes an error in the video lesson “Cleaning Up Incorrectly Labelled Data” at 3 minutes 28 seconds from the start when he refers to “…100 mislabelled dev set examples…” when what he really means is “…100 mis-classified dev set examples…”.
Mis-classified examples includes both mis-labelled examples and mis-classified examples. For example, if the classifier predicts a cat is not in the picture when it is presented with an image of a great cat but predicts it is not a cat then this is when the classifier mis-classifies the image. But when the classifier predicts that there is a cat in the picture when presented with a dog, then this is an example of a mis-labeled example.
Also, I have noticed another error. Prof. Ng presents four percentage values of 8%, 43%, 61% and 6% which adds up to over 100%!
I am paying £36 per month for this course and it should be error-free. Its not on.
Yes, the terminology “mislabelled” is a bit ambiguous and he says as much earlier in that same lecture (around :45 to :60). It’s clear enough what he means if you just listen to the flow of what he’s saying.
On the point about the percentages adding up to more than 100%, that’s simple: the categories are not mutually exclusive. It can be a picture of a great cat and it can be blurry at the same time, for example. We saw that in the previous lecture:
Its not ‘‘…ambiguous…”, it’s wrong.
Mis-labeled means the incorrect label has been assigned to a particular input dataset member.
Mis-classified means the model has incorrectly predicted the expected ground truth label for a particular input dataset member.
It would be clear what he meant if he hadn’t made a mistake. Simple as.
However, you are correct about the percentages.
I am seeing more and more errors by Prof. Ng, as I progress through his course. I”m paying £36 a month for this course and there should be zero errors. Its not on.
