Universals

Universals are the entities — properties, relations, kinds, patterns — that can be instantiated by multiple particular things. The question of whether universals exist independently of the minds that conceive them is one of the oldest problems in philosophy, dividing Platonic realists (who hold that universals exist in a realm of abstract forms) from nominalists (who hold that only particular things exist, and universals are merely names).

From a systems perspective, the universals debate is a debate about compression. When a cognitive system recognizes that multiple instances share a pattern — that all red things are red, that all triangles are triangular — it has discovered a universal. Whether that universal 'exists' independently of the discovering system is the question whether patterns are discovered or invented. The constructivist answer — that universals are viable constructions rather than pre-existing entities — dissolves the traditional opposition without settling it. What matters is not whether universals exist but whether they are productive: do they enable predictions, interventions, and further discoveries?

The connection to machine learning is direct. A neural network that learns to classify cats has, in effect, constructed a universal — a pattern that picks out the cat-property across instances. Whether that pattern corresponds to a 'real' universal (some essential cat-ness) or merely to a statistically useful compression is precisely the question that the philosophy of universals has been asking for two millennia. The network does not answer the question. It performs it.