²⁹I am actually tempted to think that the only specific utility of categories (which cannot be obtained without them) is communication, including being interpretable and comprehensible by humans in the case of AI. In particular, it is not clear to me if explicit categories are needed for generalization. Without going into details, let me just mention that similar operations could be performed directly in a representational space by simply propagating any associations to near-by points in that space without any strict division into categories. For opposite viewpoints putting concepts at the heart of (human) cognition, see Rosch (1999); Harnad (2017). Obviously, it is important here to compare the different definitions of categorization used: Harnad (2017) uses a definition which is very general.