Vision models show mixed understanding of abstract concepts across contexts
Do vision models know that three is three?
Computer Vision and Pattern Recognition
Summary
This paper investigates whether vision models recognize abstract concepts, like the number three or clean versus dirty, similarly across different visual forms such as handwritten digits and hand gestures. The researchers found that while these models generally find directions in their feature space representing concepts locally within one type of image, the same concept does not always have a shared representation across very different visual styles. Some models represent a concept differently depending on the context, a phenomenon termed "fragmented linearity." Models that have stronger local concept representations tend to develop more unified representations that transfer across visual styles.
What this means in practice
- •For computer vision engineers: Improve model design to enhance consistent recognition of concepts across visually different inputs like digits and gestures.
- •For machine learning platform teams: Develop evaluation tools for detecting whether models represent concepts uniformly across varied image domains to ensure robustness.
Authors
Ankit Sonthalia, Arnas Uselis, Seong Joon Oh
Abstract
Do vision models encode the transition $2{\to}3$ the same way whether it appears in handwritten digits or in hand gestures, and do they "understand" that clean${\to}$dirty is the same transformation regardless of whether it is applied to a kitchen or a carpet? The linear representation hypothesis (LRH) answers this for language models: concept changes correspond to directions in (language) embedding space. In vision, the same concept can manifest in visually unrelated "regimes" that rarely co-occur: images of digits and hands do not share much in common in pixel space. To understand whether vision models still encode the underlying abstract concepts, we study 46 vision encoders across architectures, training objectives, and scales, on 10 concept families including numerosity, age, object identity, and MIT-States attributes. Within each regime, we find statistically significant concept directions near-universally across our tested models. However, two distinct regimes do not always share the same concept direction. Several models learn multiple linear structures for the same concept transition, split across regimes. We call this phenomenon "fragmented linearity": strong local linearity coexisting with near-absent global linearity. Fragmentation is, however, not fixed: models with stronger local linearity progressively escape it (local-global linearity correlated with $τ\in [0.38, 0.89]$ across all concepts), and where global linearity does emerge, concept directions can transfer as linear probes across regimes.