Here, we report that the depth and the width of a neural network are dual from two perspectives. First, we employ the partially separable representation to determine the width and depth. Second, we use the De Morgan law to guide the conversion between a deep network and a wide network. Furthermore, we suggest the generalized De Morgan law to promote duality to network equivalency.