Latest posts
-
Taylor Expansion as a Local Bet
A Taylor expansion can describe a small transition from a base class toward an unseen one. Its usefulness depends on whether that local approximation predicts well, which the formulation alone cannot show.
-
What an Elephant Minus Horse Vector Says
The difference between elephant and horse embeddings can be drawn as a direction in semantic space. That direction suggests a route for extrapolation, but is not evidence that an unseen class will be recognized.
-
From Edges to Unseen Object Classes
An object representation built from edges, corners, and formal parts has a natural connection to compositional zero-shot learning. The connection is structural; it does not, by itself, establish recognition performance.