Bounding the Complexity of Formally Verifying Neural Networks: A Geometric Approach