Does efficient coding extend to deep hierarchies to explain universal dimensions?

Establish whether the efficient-coding hypothesis for natural images extends beyond first-layer orientation- and frequency-tuned features to deep hierarchical layers, thereby explaining the universal dimensions of natural image representation observed across trained vision networks as a consequence of optimal image encoding.

Background

Efficient-coding theory has been used to explain why early visual representations in both biological and artificial systems resemble V1-like filters tuned to orientation, frequency, and color. The present work shows that highly shared, brain-aligned dimensions also emerge in deeper layers across diverse trained networks.

If efficient coding generalizes to deep hierarchies, it could offer a principled account for the emergence of these universal dimensions as optimal encodings of natural image statistics across multiple levels of representation.