# What is the relation between self-taught learning and transfer learning?

I am new to transfer learning and I start by reading A Survey on Transfer Learning, and it stated the following:

according to different situations of labeled and unlabeled data in the source domain, we can further categorize the inductive transfer learning setting into two cases:

case $$(a)$$ (It is irrelevant to my question).

case $$(b):$$ No labeled data in the source domain are available. In this case, the inductive transfer learning setting is similar to the self-taught learning setting, which is first proposed by Raina et al. [22]. In the self-taught learning setting, the label spaces between the source and target domains may be different, which implies the side information of the source domain cannot be used directly. Thus, it’s similar to the inductive transfer learning setting where the labeled data in the source domain are unavailable.

From that, I understand that self-taught learning is inductive transfer learning.

But I opened the paper of self-taught learning that was mentioned (i.e paper by Raina et al. [22].), and It stated the following in the introduction:

Because self-taught learning places significantly fewer restrictions on the type of unlabeled data, in many practical applications (such as image, audio or text classification) it is much easier to apply than typical semi-supervised learning or transfer learning methods.

And here it looks like transfer learning is different from self-taught learning.

So what is the right relation between them?

• I'm not familiar with the term self-taught learning, but I suspect it is related to self-supervised learning, which is turn is related to transfer learning, as SSL can be used for transfer learning. See this post for more info about self-supervised learning.
– nbro
May 11 at 0:09
• @nbro It is good to know about self-supervised learning, Thank you. But what about the distribution of the data that used for it? and the actual labels of them? In Self_Taught learning, there is no need for the unlabeled data (that used for training) to follow the same class labels or generative distribution as the labelled data that will be used for the supervised task. May 11 at 8:47