Compelling ReLU Networks to Exhibit Exponentially Many Linear Regions at Initialization and During Training | Read Paper on Bytez