An Experimental Study of Semantic Continuity for Deep Learning Models