Global Geometry Is Not Enough for Vision Representations
Two recent studies examine the limitations of global geometry in vision representations and the role of spatial-frequency accessibility in vision-language models.
A study submitted to arXiv on February 3, 2026, and last revised on June 1, 2026, found that global geometry is often insensitive to how elements are composed[1]. The researchers tested the ability of geometric metrics to predict compositional binding across various vision encoders and discovered that standard geometry-based statistics exhibit near-zero correlation with compositional binding. In contrast, functional sensitivity reliably tracks this capability. Another study, submitted on June 2, 2026, examined how vision-language models alter the structure of visual information through spatial-frequency accessibility. The analysis used pretrained models with frozen parameters to reduce confounding effects from optimization and found consistent frequency-dependent changes in accessibility across CLIP and DINOv2 on ImageNet and MS-COCO datasets[2]. The study also observed that spectral accessibility follows a non-monotonic trajectory across depth, peaking at intermediate layers before decreasing toward the output representation. Furthermore, the researchers noted that CLIP's learned projection is spectrally neutral, whereas DINOv2's [CLS] pooling induces a structured loss across the spectrum.
research-paper
Background sources we checked (4)
- arxiv.org ↗ A common assumption in representation learning is that globally well-distributed embeddings support robust and generalizable representations. This focus has shaped both training objectives and evaluation protocols, implicitly treating global geometry as a proxy for representation…
- en.wikipedia.org ↗ In mathematics, a Lie group (pronounced Lee) is a group that is also a differentiable manifold, such that group multiplication and taking inverses are both differentiable. A manifold is a space that locally resembles Euclidean space, whereas groups define the abstract concept of…
- en.wikipedia.org ↗ Simultaneous localization and mapping (SLAM) is a process where a computer constructs or updates a map of an unknown environment while simultaneously keeping track of an entity's location within it. While this initially appears to be a chicken or the egg problem, there are severa…
- en.wikipedia.org ↗ In machine learning, a neural field (also known as implicit neural representation, neural implicit, or coordinate-based neural network), is a mathematical field that is fully or partially parametrized by a neural network. Initially developed to tackle visual computing tasks, such…