Technocraft Corporation

Hierarchies That Generalize

An AGI research blog tracing how hierarchical object representations relate to zero-shot learning and compositional generalization.

Compositional pipeline

From Edges to Class Concepts

The proposed pipeline reads an input through four stages: edges φ(x), corners ψ(φ), formal parts z(ψ), and class concepts β(z). Each stage factorizes the representation into primitives that can recombine.

Edges φ(x)

Primitive contours extracted from the raw input.

Corners ψ(φ)

Junctions formed by combinations of edges.

Formal Parts z(ψ)

Corner assemblies that describe object components.

Class β(z)

Class concepts composed from formal parts.

Directional extrapolation

Vector Differences in Semantic Space

The proposed shift Δβ = v(Elephant) − v(Horse) defines a direction in semantic embedding space. Projecting that direction is a candidate route into unseen class coordinates.

This is a proposed direction, not evidence on its own that an unseen class will be recognized.