CoChat

The DOI matches its Crossref record; 45 of the 45 references it cites were found in a registry or an index.

Universal dimensions of visual representation

Open access
Chen, Zirui, Bonner, Michael F. · 2025 · Science Advances

Abstract

Do visual neural networks learn brain-aligned representations because they share architectural constraints and task objectives with biological vision or because they share universal features of natural image processing? We characterized the universality of hundreds of thousands of representational dimensions from networks with different architectures, tasks, and training data. We found that diverse networks learn to represent natural images using a shared set of latent dimensions, despite having highly distinct designs. Next, by comparing these networks with human brain representations measured with functional magnetic resonance imaging, we found that the most brain-aligned representations in neural networks are those that are universal and independent of a network’s specific characteristics. Each network can be reduced to fewer than 10 of its most universal dimensions with little impact on its representational similarity to the brain. These results suggest that the underlying similarities between artificial and biological vision are primarily governed by a core set of universal representations that are convergently learned by diverse systems.

Overview

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1126/sciadv.adw7697

Cites

45 of 45 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig. 1. Overview of a method for computing universality and brain similarity of network dimensions. (A) Four sets of DNNs were analyzed, including three sets of trained models that varied in their random initializations, architectures, or task objectives and one set of untrained models with different initializations. (B) Universality and brain similarity were defined as the average prediction accuracy of a latent dimension from a target network when using the activations of other networks or the fMRI activations of the human brain as predictors. Dimensions that can be consistently predicted from the representations of other networks have high universality. Dimensions that can be consistently predicted from the representations of the human brain have high brain similarity.
Fig. 1. Overview of a method for computing universality and brain similarity of network dimensions. (A) Four sets of DNNs were analyzed, including three sets of trained models that varied in their random initializations, architectures, or task objectives and one set of untrained models with different initializations. (B) Universality and brain similarity were defined as the average prediction accuracy of a latent dimension from a target network when using the activations of other networks or the fMRI activations of the human brain as predictors. Dimensions that can be consistently predicted from the representations of other networks have high universality. Dimensions that can be consistently predicted from the representations of the human brain have high brain similarity.CC BY · PubMed Central
Fig. 2. Universality and brain similarity of network dimensions. Universality and brain similarity were computed for representational dimensions in four sets of DNNs. These included three sets of trained networks with varied initializations, architectures, and task objectives and one set of untrained networks. These metrics were computed for the PCs of network activations extracted from the sampled layers of each network. Universality scores reflect the degree to which a representational dimension is shared across all networks in a set, and brain similarity scores reflect the degree to which a representational dimension is shared with the human visual system. Measurements of the human visual cortex activity were obtained from the Natural Scenes fMRI Dataset using a general ROI that included all visually responsive voxels (13). Universality and brain similarity scores are plotted for all analyzed network dimensions. These plots show the density of dimensions on a logarithmic scale, with densities computed using kernel density estimation. The orange dots show the mean universality and brain similarity scores for equally sized quantiles of 100 dimensions along the x axis. These plots show similar trends for all three sets of trained models (the first three plots on the left). Specifically, they exhibit a high density of points near the origin, showing that most dimensions are idiosyncratic to each network and not shared with the human brain. However, they also contain a subset of dimensions with exceptionally high universality and brain similarity scores. These latter dimensions correspond to representations that are consistently learned by all networks within a set and also strongly shared with the visual representations of the human brain. In contrast, untrained networks (right) can also have shared dimensions, but these shared untrained dimensions have relatively weak brain similarity scores.
Fig. 2. Universality and brain similarity of network dimensions. Universality and brain similarity were computed for representational dimensions in four sets of DNNs. These included three sets of trained networks with varied initializations, architectures, and task objectives and one set of untrained networks. These metrics were computed for the PCs of network activations extracted from the sampled layers of each network. Universality scores reflect the degree to which a representational dimension is shared across all networks in a set, and brain similarity scores reflect the degree to which a representational dimension is shared with the human visual system. Measurements of the human visual cortex activity were obtained from the Natural Scenes fMRI Dataset using a general ROI that included all visually responsive voxels (13). Universality and brain similarity scores are plotted for all analyzed network dimensions. These plots show the density of dimensions on a logarithmic scale, with densities computed using kernel density estimation. The orange dots show the mean universality and brain similarity scores for equally sized quantiles of 100 dimensions along the x axis. These plots show similar trends for all three sets of trained models (the first three plots on the left). Specifically, they exhibit a high density of points near the origin, showing that most dimensions are idiosyncratic to each network and not shared with the human brain. However, they also contain a subset of dimensions with exceptionally high universality and brain similarity scores. These latter dimensions correspond to representations that are consistently learned by all networks within a set and also strongly shared with the visual representations of the human brain. In contrast, untrained networks (right) can also have shared dimensions, but these shared untrained dimensions have relatively weak brain similarity scores.CC BY · PubMed Central
Fig. 3. Universality and brain similarity across network layers. These plots show the universality and brain similarity scores for individual network layers spanning the full depth of each network. Four sets of DNNs were examined, including three sets of trained networks with varied initializations, architectures, and tasks and one set of untrained networks. The analyses are the same as in Fig. 2, but here, the results are plotted as the average values for individual layers, which are labeled according to their relative depth. Further details of these network layers are included in the Supplementary Materials (46). Average universality and brain similarity scores were computed for equally sized quantiles of 100 dimensions along the x axis for each layer. Panels on the sides of each plot show the density of dimensions on a logarithmic scale computed using kernel density estimation. As in Fig. 2, these plots exhibit a high density of points near the origin, which means that across all sampled layers, most dimensions are idiosyncratic and not shared with the human brain. However, the three sets of trained networks (the first three plots on the left) also contain a subset of dimensions at the right end of each plot that has exceptionally high universality and brain similarity scores. These layerwise plots show that a consistent trend is observed across all sampled layers and that universal dimensions are not restricted to early network layers.
Fig. 3. Universality and brain similarity across network layers. These plots show the universality and brain similarity scores for individual network layers spanning the full depth of each network. Four sets of DNNs were examined, including three sets of trained networks with varied initializations, architectures, and tasks and one set of untrained networks. The analyses are the same as in Fig. 2, but here, the results are plotted as the average values for individual layers, which are labeled according to their relative depth. Further details of these network layers are included in the Supplementary Materials (46). Average universality and brain similarity scores were computed for equally sized quantiles of 100 dimensions along the x axis for each layer. Panels on the sides of each plot show the density of dimensions on a logarithmic scale computed using kernel density estimation. As in Fig. 2, these plots exhibit a high density of points near the origin, which means that across all sampled layers, most dimensions are idiosyncratic and not shared with the human brain. However, the three sets of trained networks (the first three plots on the left) also contain a subset of dimensions at the right end of each plot that has exceptionally high universality and brain similarity scores. These layerwise plots show that a consistent trend is observed across all sampled layers and that universal dimensions are not restricted to early network layers.CC BY · PubMed Central
Fig. 4. 2D visualization of high-level universal representations. Image activations for the 100 most universal dimensions from a high-level network layer were embedded in two dimensions using UMAP. Specifically, image activations were obtained for the top 100 dimensions with the highest universality scores in the penultimate layer from the set of models trained on different tasks. For visualization purposes, this figure only includes images shown to a single subject. This plot shows that universal dimensions do not reflect low-level image features but instead capture high-level properties that group images into semantically related clusters, some of which are highlighted here, including animals, food, sports, and people. In contrast, fig. S8 shows a visualization of the 100 least universal dimensions from the same network layer, and it shows no clear semantic organization.
Fig. 4. 2D visualization of high-level universal representations. Image activations for the 100 most universal dimensions from a high-level network layer were embedded in two dimensions using UMAP. Specifically, image activations were obtained for the top 100 dimensions with the highest universality scores in the penultimate layer from the set of models trained on different tasks. For visualization purposes, this figure only includes images shown to a single subject. This plot shows that universal dimensions do not reflect low-level image features but instead capture high-level properties that group images into semantically related clusters, some of which are highlighted here, including animals, food, sports, and people. In contrast, fig. S8 shows a visualization of the 100 least universal dimensions from the same network layer, and it shows no clear semantic organization.CC BY · PubMed Central
Fig. 5. Universal dimensions underlie the results of conventional RSAs. RSA was used to compare the representations of neural networks and the visual cortex. These analyses were performed using the same general ROI in the visual cortex and the same sets of neural networks as in Figs. 2 and 3. RDMs were created by calculating Pearson correlation distances for pairwise comparisons of image representations within each network and each fMRI subject. Representational similarities were obtained by calculating the Spearman correlation between the RDMs for a network and an fMRI subject. These representational similarities were averaged across subjects. For each network, the best-performing layer was selected using a set of training data, and the final representational similarity was computed on the held-out test data. These analyses show the results of RSA for networks whose representations were either intact or reduced to subspaces of their top 10 or 5 universal dimensions. In these plots, each dot is a network, and lines connect different versions of a network containing all, 10, or 5 dimensions. The violin plots show distributions of representational similarities across networks. Even after markedly reducing the networks to just 10 or 5 universal dimensions, the representational similarities exhibit little or no decrease—rather, for all three sets of trained networks, the representational similarities slightly improve. These results demonstrate that conventional measures of representational similarity between neural networks and the visual cortex are largely driven by the subspaces of universal dimensions contained within each network. Similar trends were observed within each individual subject and in other ROIs, as shown in figs. S10 and S11.
Fig. 5. Universal dimensions underlie the results of conventional RSAs. RSA was used to compare the representations of neural networks and the visual cortex. These analyses were performed using the same general ROI in the visual cortex and the same sets of neural networks as in Figs. 2 and 3. RDMs were created by calculating Pearson correlation distances for pairwise comparisons of image representations within each network and each fMRI subject. Representational similarities were obtained by calculating the Spearman correlation between the RDMs for a network and an fMRI subject. These representational similarities were averaged across subjects. For each network, the best-performing layer was selected using a set of training data, and the final representational similarity was computed on the held-out test data. These analyses show the results of RSA for networks whose representations were either intact or reduced to subspaces of their top 10 or 5 universal dimensions. In these plots, each dot is a network, and lines connect different versions of a network containing all, 10, or 5 dimensions. The violin plots show distributions of representational similarities across networks. Even after markedly reducing the networks to just 10 or 5 universal dimensions, the representational similarities exhibit little or no decrease—rather, for all three sets of trained networks, the representational similarities slightly improve. These results demonstrate that conventional measures of representational similarity between neural networks and the visual cortex are largely driven by the subspaces of universal dimensions contained within each network. Similar trends were observed within each individual subject and in other ROIs, as shown in figs. S10 and S11.CC BY · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Show all 36 references

Research works printed but not linked: 4

  1. Y. Le, X. Yang, Tiny imagenet visual recognition challenge. CS 231N 7, 3 (2015).
  2. A. Krizhevsky I. Sutskever G. E. Hinton “ImageNet classification with deep convolutional neural networks ” in Advances in Neural Information Processing Systems F. Pereira C. Burges L. Bottou K. Weinberger Eds. (Curran Associates Inc. 2012) vol. 25 pp. 1097–1105.
  3. J. S. Prince G. Fajardo G. A. Alvarez T. Konkle “Manipulating dropout reveals an optimal balance of efficiency and robustness in biological and machine visual systems ” in The Twelfth International Conference on Learning Representations (Open Review 2024); https://openreview.net/pdf?id=ADDCErFzev.
  4. B. S. Robinson N. Drenkow C. Conwell M. Bonner “A sparse null code emerges in deep neural networks ” in Proceedings of UniReps: The First Workshop on Unifying Representations in Neural Models M. Fumero E. Rodolá C. Domine F. Locatello K. Dziugaite C. Mathilde Eds. (PMLR 2024) vol. 243 of Proceedings of Machine Learning Research pp. 302–314.

Cited by

See all 25

Metadata

Authors
Chen, Z., & Bonner, M. F.
Affiliations
Johns Hopkins University (US)
Corresponding author
Chen, Zirui
Venue
Science Advances
Published
2025-07-04
Volume
11
Issue
27
Pages
eadw7697-eadw7697
ISSN
2375-2548 (ISSN-L)
OpenAlex
W4411948297
Cited by
16
References
36
Citations by year
2026: 12 · 2025: 3
FWCI
8.51
Type
journal article
Language
en
Open access
Licence
cc-by
Locations
  • Science Advances (publishedVersion, cc-by)
  • PubMed Central (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • arXiv (Cornell University) (submittedVersion, cc-by-nc-nd)
  • OSF Preprints (OSF Preprints) (submittedVersion)
  • arXiv (Cornell University)
Topics
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Visual perception and processing mechanisms
  • Physical Sciences › Computer Science › Computer Vision and Pattern Recognition › Advanced Neural Network Applications
  • Physical Sciences › Computer Science › Artificial Intelligence › Domain Adaptation and Few-Shot Learning
Keywords
representational similarity, latent dimensions, universal representation, dimensionality reduction, functional MRI
MeSH terms
Brain, Brain Mapping, Humans, Image Processing, Computer-Assisted, Magnetic Resonance Imaging, Vision, Ocular, Visual Perception, Neural Networks, Computer

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Universal dimensions of visual representation. https://app.cochat.ai/papers/10.1126/sciadv.adw7697
The paper
Chen, Z., & Bonner, M. F. (2025). Universal dimensions of visual representation. Science Advances. https://doi.org/10.1126/sciadv.adw7697

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.