ImageNet did not invent its category structure from scratch. It borrowed the noun hierarchy of WordNet, the hand-built lexical database created by George A. Miller at Princeton, which organizes English words into concept sets (“synsets”) and links them through is-a relations. ImageNet’s design was to attach a large set of labeled images to each meaningful WordNet noun concept, so the image categories and their hierarchy inherited WordNet’s structure directly.
This is why a single resource from the symbolic, knowledge-engineering tradition of AI - WordNet, rooted in 1980s and 1990s lexical research - ended up as the skeleton of the dataset that helped launch the deep-learning era. The semantic scaffold came from decades of careful human encoding of word meanings; the images and scale came later. The connection is a neat reminder that AI’s “symbolic” and “learned” threads are not as separate as they are often portrayed: one quietly enabled the other.