Better Neural Network Expressivity: Subdividing the Simplex