Conjecture on Hierarchical Learning Explaining Influence Patterns in Vision Models

Test and validate the conjecture that, in hierarchically structured domains such as ImageNet classification with Inception‑V1, models first learn broad categories and only later refine class-specific distinctions, thereby explaining observed influence patterns where high-influence training images can include visually similar but differently labeled examples.

Background

In qualitative comparisons of influence patterns for Inception‑V1, the authors observe that the most influential training images for a given query sometimes include visually similar images from different classes (e.g., oranges influencing a lemon query).

They conjecture that hierarchical learning dynamics—coarse category formation followed by finer-grained differentiation—could explain these cross-class positive influences, suggesting a developmental interpretation aligned with known cognitive and learning phenomena.