The activity-weight duality in feed forward neural networks: The geometric determinants of generalization