ML//neural network//embedding//directionality

Directions in embedding space encode semantic relationships: king - man + woman ≈ queen (gender direction), cat → cats (plural direction)

Directions in embedding space encode semantic relationships: king - man + woman ≈ queen (gender direction), cat → cats (plural direction)

Dot product = 0 means perpendicular = unrelated. High dot product = aligned = semantically connected.

These directions emerge during training — not hand-designed. The model discovers its own feature geometry.

The 12K+ dimensions of a typical embedding encode thousands of overlapping directional relationships simultaneously.

This is why cosine similarity works for search: similar concepts point in similar directions.