On Geometric Structure of Activation Spaces in Neural Networks