deep neural networks have demonstrated superior performance in artificial
intelligence applications, but the opaqueness of their inner working mechanism
is one major drawback in their application. The prevailing unit-based
interpretation is a statistical observation of stimulus-respons