Decoupling Semantic Similarity from Spatial Alignment for Neural Networks