Assess the effect of embedding uniformity on predictive utility

Determine whether applying uniformity- or entropy-based regularization to spread PlacePulse embeddings more evenly across the unit sphere improves retrieval and downstream prediction or instead removes meaningful variation.

Background

PlacePulse uses nearly all 256 embedding dimensions, but its variance remains concentrated in a smaller number of directions, so the embeddings occupy only a narrow region of the unit sphere. The paper proposes uniformity- or entropy-based regularizers as possible ways to alter this geometry while preserving local spatial structure.

The authors explicitly leave unresolved whether a more uniform embedding would improve practical retrieval and prediction. Greater uniformity could also eliminate meaningful density structure that currently supports interpretable clustering, making the direction of the effect an empirical question.