Does Data Scaling Lead to Visual Compositional Generalization?

TL;DR

This paper investigates whether data scaling improves compositional generalization in vision models, finding that data diversity, not scale, is key to learning compositional structures that enable efficient generalization.

Contribution

It demonstrates that compositional generalization depends on data diversity and concept coverage, and that a linearly factored representational structure underpins efficient compositional learning.

Findings

Data diversity drives compositional generalization.

Increased combinatorial coverage induces a factored representational structure.

Pretrained models show partial evidence of this structure.

Abstract

Compositional understanding is crucial for human intelligence, yet it remains unclear whether contemporary vision models exhibit it. The dominant machine learning paradigm is built on the premise that scaling data and model sizes will improve out-of-distribution performance, including compositional generalization. We test this premise through controlled experiments that systematically vary data scale, concept diversity, and combination coverage. We find that compositional generalization is driven by data diversity, not mere data scale. Increased combinatorial coverage forces models to discover a linearly factored representational structure, where concepts decompose into additive components. We prove this structure is key to efficiency, enabling perfect generalization from few observed combinations. Evaluating pretrained models (DINO, CLIP), we find above-random yet imperfect performance, suggesting partial presence of this structure. Our work motivates stronger emphasis on constructing diverse datasets for compositional generalization, and considering the importance of representational structure that enables efficient compositional learning. Code available at https://github.com/oshapio/visual-compositional-generalization.