I was reading DT-LET: Deep transfer learning by exploring where to transfer, and it contains the following:
It should be noted direct use of labeled source domain data on a new scene of target domain would result in poor performance due to the semantic gap between the two domains, even they are representing the same objects.
Can someone please explain what the semantic gap is?
In terms of transfer learning, semantic gap means different meanings and purposes behind the same syntax between two or more domains. For example, suppose that we have a deep learning application to detect and label a sequence of actions/words $a_1, a_2, \ldots, a_n$ in a video/text as a "greeting" in a society A. However, this knowledge in Society A cannot be transferred to another society B that the same sequence of actions in that society means "criticizing"! Although the example is very abstract, it shows the semantic gap between the two domains. You can see the different meanings behind the same syntax or sequence of actions in two domains: Societies A and B. This phenomenon is called the "semantic gap".
The wiki has a concise quote by Andreas Hein, where the gap is defined by "the difference in meaning between constructs formed within different representation systems". This connotes the core problem of translating meaning between an informal language (typically natural language) and a formal language (programing language or other formal symbolic system).
Informally, the problem could be defined as "the gap in meaning between two different contexts", and we might observe this in something as simple as gestures having different meanings in different cultures. It would be less of a problem translating meaning between two formal systems.
At its core, "Semantic gap" seems to relate to the difficulty in formalizing certain fuzzy concepts in symbolic systems.
