Layerwise Conservation Principle

• which says that “a network’s output activity is fully redistributed through the layers of a DNN onto the input variables, i.e., neither positive nor negative evidence is lost.”