← Florian P. Mahner
Nature Machine Intelligence · 2025

Dimensions underlying the representational alignment of deep neural networks with humans

Florian P. Mahner*, Lukas Muttenthaler*, Umut Güçlü, Martin N. Hebart
* equal contribution

Comparing humans and deep nets via the latent dimensions that drive their behaviour. The same task — same data — yet humans organise objects by what they are, while networks organise them by what they look like.

Framework overview: parallel pipelines for humans and DNNs through a triplet odd-one-out task to recover latent dimensions, then compared dimension-by-dimension.
The framework. Humans and a deep network make the same odd-one-out judgements over thousands of natural images. From these behavioural responses we recover a low-dimensional representation for each system — and then we can compare them, dimension by dimension.
The Question

Two systems can score the same on a benchmark and still represent the world differently. We wanted to know how.

Recovered dimensions for humans and DNN side by side, with example images that score most strongly on each dimension.
Each row is one recovered dimension. The images that load most strongly on it tell you what that dimension means to that system.
The Approach

A common embedding from a common task. No model surgery, no probes — only the behaviour.

Pairwise alignment between human and DNN dimensions, with semantic ones in cool tones and visual ones in warm tones.
Mapping each network dimension to its closest human counterpart. Some line up tightly. Others have no match at all.
What We Found

Networks lean visual; humans lean semantic. The match is good in the middle and breaks down at the edges — and the breakdown is principled, not noise.

Distribution of dimensions across visual versus semantic axes for humans and the DNN, showing the network's bias toward visual properties.
Where each system's dimensions sit on the visual–semantic axis. Humans spread across both. The network clusters on the visual side.
BibTeX
@article{mahner2025dimensions,
  title   = {Dimensions underlying the representational alignment of deep neural networks with humans},
  author  = {Mahner, Florian P. and Muttenthaler, Lukas and G\"u\c{c}l\"u, Umut and Hebart, Martin N.},
  journal = {Nature Machine Intelligence},
  volume  = {7},
  pages   = {848--859},
  year    = {2025},
  doi     = {10.1038/s42256-025-01041-7}
}