As you've noticed from the previous section, there's a clear difference between the traditional way we apply machine learning and machine learning that involves TL (as shown in the following diagram). In traditional machine learning, you don't transfer any knowledge or representations to any other task, which is not the case in TL. Sometimes, people use TL in a wrong way, so we are going to mention a few conditions under which you can only use TL to maximize the gains.

The following are the conditions for applying TL:

• Unlike traditional machine learning, the source and target task or domains don't have to come from the same distribution, but they have to be similar
• You can also use TL in case of less training samples or if you don't have the necessary computational power