By projecting data into a semantic space, it is possible to represent extra knowledge beyond the transformation from inputs to representations in a way that can be neatly disentangled. The author argues that learning about the relationships between variables in high-level semantic variables is an important next step in improving current neural nets.