I am working on a data-set that has multiple labels associated with it (not necessarily independent of each other). During my development, I am confused if I should consider it as a multi-class multilabel data or a multi-class MTL kind of an approach. Is there any fundamental difference between the two?
$\begingroup$
$\endgroup$
1
-
1$\begingroup$ Hello. "MTL" stands for? Multi-task learning? $\endgroup$– nbroMar 10, 2022 at 23:07
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
1
Is there any fundamental difference between the two?
The difference is in the names:
- Multi task means that we are learning more than a single task, i.e. the labels we have will be used to compute different losses
- Multi label means that for a single task, more than one label is allowed as correct prediction. but in practice, the different labels will be used to compute a single loss.
You can also visualize the difference: a network trained on multi tasks will have more output layers, one per task,whereas a network trained on multi label classification require only one output layer. Note that in a multi task learning setting one of the tasks could very well be multi label classification.
I am confused if I should consider it as a multi-class multilabel data or a multi-class MTL kind of an approach
That depends on what you're trying to do, and we can't even give suggestions without knowing the data.
answered Mar 11, 2022 at 0:23
-
$\begingroup$ Thanks Edoardo for the explanation. To be specific, I have 3 sec speech audio inputs and labels which state if the particular audio input has sound repetition, prolongation and other types of speech disfluencies. And each disfluency is assigned a number between 0-3. A given sample can be categorized as sound repetition =2 and prolongation=1. So I think with an objective to classify the audio inputs into different classes of disfluency I can either choose a single task learning or multi task, I think. $\endgroup$ Mar 12, 2022 at 1:19