0
$\begingroup$

In Multitask Learning (MTL,) one factor that can have an effect on the performance of an MTL architecture is how correlated or uncorrelated tasks are. As far as I understand, the more correlated tasks are, the less performance improvement we will be expecting to see (correct me if I am wrong).

Then I wonder, when choosing different tasks for MTL, how should we know how correlated our tasks are? I'm asking since ideally we want to train a network to give us maximum performance improvement but without knowing whether tasks are highly correlated, we may not achieve this.

Is there any way to measure the "correlation" between tasks? Or should we just simply compare tasks' input and output and judge if they are similar or different?

Citation about task correlation (Caruana's MTL paper, 1997):
Caruana, R. Multitask Learning. Machine Learning 28, 41–75 (1997)
Improve this question
$\endgroup$
4
  • $\begingroup$ Can you explain what do you mean by task correlation? Tasks should share some features, they must be similar in some way, otherwise MTL will not make a sense $\endgroup$
    – Aksakal
    Commented Feb 13, 2020 at 21:06
  • $\begingroup$ @Aksakal in my case, tasks are NLP tasks. For example, a binary text classification task and a token-level task like Named Entity Recognition (just as an example here.) Back to your question, I'm exactly trying to figure out what "task correlation" means. You say they should share some features, what features are we talking about here? $\endgroup$
    – Pedram
    Commented Feb 13, 2020 at 22:14
  • $\begingroup$ You came up with task correlation concept, so you must define it. In a degenerated case, imagine that you divided the dataset into two subsets. then you run them as if they were different tasks in MTL setup. would this be better than running the subsets separately? you bet it should, if nothing else, then at least because the sample size increased. would you call these tasks correlated? I don't know, because you didnt explain what you mean by correlation $\endgroup$
    – Aksakal
    Commented Feb 13, 2020 at 22:35
  • 1
    $\begingroup$ @Aksakal I did not come up with the "task correlation" this is what I saw in the original MTL paper (I just added the citation for the paper to the question.) and trying to understand what it means or how it can be measured. $\endgroup$
    – Pedram
    Commented Feb 13, 2020 at 22:45

0

Reset to default

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.