From Comparison to Composition: Towards Understanding Machine Cognition of Unseen Categories

Minghao Fu, Sheng Zhang, Guangyi Chen, Zijian Li, Fan Feng, Yifan Shen, Shaoan Xie, Heng Huang, Kun Zhang

Keywords: Concept Learning; Cognitive Science; Compositional Generalization

TL;DR: We develop a theory of unseen category cognition via comparison and re-composition with identifiability guarantees.

Abstract: Humans acquire visual concepts through a natural compare-then-compose process, enabling effortless generalization to novel categories. Whether machines can achieve similar genuine semantic generalization—recognizing unseen categories without any training exposure—remains a fundamental open question. We formalize the cognitive compare-then-compose mechanism into Comparison–Composition Cognition (C³), an identifiability framework grounded in two complementary conditions: comparison requires sufficient cross-category contrast for latent concept identification; composition requires disjoint concept supports for reliable unseen category recognition. Under mild, nonparametric assumptions, we prove these conditions yield both necessary and sufficient guarantees. On eight fine-grained benchmarks under a genuine generalization protocol, our instantiation achieves +3.8% average accuracy over state-of-the-art, with ablations confirming each component’s contribution. C³ provides the first principled characterization of when and why learned representations generalize to unseen categories.

Submission Number: 22