Why transfer learning is used more often than muti task learning?

Hi Mentor and friend,

At the end of the video “multi task learning”, Professor Ng said many times that: transfer learning is used more often than muti task learning. I do not very follow why’s that. Can someone give me more idea?

My understanding is that, multi task learning has the “features” of transfer learning. for example, they both use “other class data” for low level feature learning, so that boost final classification target. Prof Ng also said, muti-task learning is not simple because too many classes at the same time. So, if there is one pic, I want to detect 4 objects, my first try should go for 4 separate NN learning instead of a single big NN muti-task?

Can someone give me more idea why transfer learning is used more than muti task? thank you.

It’s not anything deep or complicated: transfer learning is a more straightforward technique and there are lots more situations in which it is applicable. There are more cases in which it is useful, so it gets used more frequently. It’s been a while since I listened to the Course 3 lectures, but I’m pretty sure Prof Ng makes that general point when he is discussing all this.

thanks for the reply, Paul.
So, are you saying, when Prof says transfer learning is used more often than multi task learning, is just because more applications need transfer learning, NOT some tech disadvantages of multi task learning?

Yes, I believe that is the point Prof Ng is making. But all this is just “word games” really: what do you mean by a “tech disadvantage”? Does the fact that multitask learning does not apply in as many cases count as a “tech disadvantage”?

His goal is to educate us about as many different ideas and techniques as he can. Not every idea is applicable in every situation, but the more you know, the better chance you have of finding a solution that works to solve some particular important problem that you are working on. The more “tools you have in your toolbox”, the better off you are. Note that he always includes an explanation of how to figure out if a given technique is applicable or not. As with most everything here, there are no magic “silver bullet” or “one size fits all” solutions.

got it. thanks.
The reason I was saying “tech disadvantage”, because Prof Ng mentioned twice that “In practice, multi-task learning is used much less often than transfer learning”. So, I thought this multi-task is “bad” for some reasons.

But I got the idea now from you: just more real application needs transfer learning.