Characterizing universal object representations across vision models
Abstract
Deep neural networks trained with different architectures, objectives, and datasets have been reported to converge on similar visual representations. However, what remains unknown is which visual properties models actually converge on and which factors may underlie this convergence. To address this, we decompose the object similarity structure of 162 diverse vision models into a small set of non-negative dimensions. To determine universal versus model-specific dimensions, we then estimate how often each dimension reappears across models. In contrast to model-specific dimensions, universal dimensions are more interpretable and more strongly driven by conceptual image properties, indicating the relevance of interpretability and semantic content as implicit factors driving universality across models. Differences in architecture, objective function, training data, model size, and model performance do not explain the emergence of universal dimensions. However, models with more universal dimensions also better predict macaque IT activity and human similarity judgments, suggesting that universality reflects representations relevant to biological vision. These findings have important implications for understanding the emergent representations underlying deep neural network models and their alignment with biological vision.
Related
- Acceleration of vision changes object representations from infancy to adulthood
- Prior visual context accelerates the emergence of object representations
- Sustained multidimensionality of object representations as a scaffold for diverse behaviours
- Disambiguation of two-tone images aligns object representations in the EEG
- Two dominant axes structure high-dimensional object representations in the human ventral temporal cortex
BibTeX
@misc{mahner2026characterizing,
title = {{Characterizing universal object representations across vision models}},
author = {Florian P. Mahner and Johannes Roth and Ka Chun Lam and Michael F. Bonner and Francisco Pereira and Martin N. Hebart},
howpublished = {arXiv},
year = {2026},
month = nov,
doi = {10.48550/arxiv.2605.13675},
eprint = {2605.13675},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2605.13675},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Characterizing universal object representations across vision models
AU - Florian P. Mahner
AU - Johannes Roth
AU - Ka Chun Lam
AU - Michael F. Bonner
AU - Francisco Pereira
AU - Martin N. Hebart
PY - 2026
DA - 2026/11/05
PB - arXiv
DO - 10.48550/arxiv.2605.13675
UR - https://arxiv.org/abs/2605.13675
AB - Deep neural networks trained with different architectures, objectives, and datasets have been reported to converge on similar visual representations. However, what remains unknown is which visual properties models actually converge on and which factors may underlie this convergence. To address this, we decompose the object similarity structure of 162 diverse vision models into a small set of non-negative dimensions. To determine universal versus model-specific dimensions, we then estimate how often each dimension reappears across models. In contrast to model-specific dimensions, universal dimensions are more interpretable and more strongly driven by conceptual image properties, indicating the relevance of interpretability and semantic content as implicit factors driving universality across models. Differences in architecture, objective function, training data, model size, and model performance do not explain the emergence of universal dimensions. However, models with more universal dimensions also better predict macaque IT activity and human similarity judgments, suggesting that universality reflects representations relevant to biological vision. These findings have important implications for understanding the emergent representations underlying deep neural network models and their alignment with biological vision.
ER -