Multi-task vs multi-label classification

Hello! I can understand the difference between multi-task and multi-class, but I get confused about what’s the difference between multi-task and multi-label classification. Is it the same?
Thanks in advance!

Multi-task does several tasks, multi-label has one task but the outputs can be various labels for eg. Dog, cat, parrot :grin:

Thank you for your answer! It’s still not clear to me because in Andrew’s example he is using multiple labels too (pedestrian, car, stop sign, traffic lights), but in that case it is multi-task. I still can’t see the difference :disappointed:

Hello @superhop,

In my opinion, the multi-task example in the lecture can also be seen as a multi-label classification. However, the two concepts are not identical. For example, mult-task is not limited to doing classification, but also regression, right?

Sometimes people can argue their difference by defining a “multi-label classification” as one “task” which can be performed by a multi-task model alongside other “tasks” that can be anything like a regression task, a embedding learning task, or a single-label classification task. In their argument, the multi-task learning is a family that hosts multi-label classification.

I am not a guy very interested in differentiating names like those two :wink:, so just in case you are doing a serious literature review, analysis, or seriously preparing for explaining them to anyone, I recommend you to google for more articles or academic papers about them.


Hello Raymond,
Thanks for your answer, I think now I got the idea a little better. I’m not so interested in doing a deep analysis about the concepts, I just wanted to make sure that I was not missing any important concept from the lesson.
Thanks for your patience

You are welcome, Hector :slight_smile: @superhop