Subspace Image Synthesis for Style-Invariant Object Recognition
- 1The University of Tokyo
Abstract
Human-like behavior in visual models does not necessarily imply human-like internal computation. When a model recognizes objects robustly across changes in visual style, what internal computations support this behavior, and which are relevant to human perception? We addressed this question by pruning style-robust neural networks into candidate submodels corresponding to two computational strategies. The first retained channels whose responses were stable across style changes for fixed content, yielding a Content Stability Score (CSS)-based submodel. The second retained channels whose responses were stable across content changes for fixed style, yielding a Style Stability Score (SSS)-based submodel. Across ImageNet, Stylized ImageNet, and model-versus-human benchmarks, the two submodel families showed comparable object-recognition performance. We then used their exclusive feature subspaces to generate matched images with Gram-matrix-based neural style transfer and asked human observers which image was more similar in visual style to a reference image. Images generated from the CSS-based submodel were selected significantly more often than chance, even though CSS preferentially retained channels whose individual activation patterns were relatively insensitive to style. These results suggest that representations preserving content across style changes are more consistent with human judgments of visual style. More broadly, they demonstrate a framework for isolating candidate computational subspaces from large artificial models and testing their perceptual relevance through image synthesis.
Related
- Local recurrence accounts for extended processing during occluded-object recognition
- Mucosal-associated invariant T cells recognize a tumor-derived metabolite in the DNA synthesis pathway
- Facilitating Valid Statistical Inference in Biomedical Image Synthesis
- Linear Fitness Subspace in Protein Language Models Enables Sample-Efficient Directed Evolution
- Subspace communication in the hippocampal–retrosplenial axis
BibTeX
@misc{ono2026subspace,
title = {{Subspace Image Synthesis for Style-Invariant Object Recognition}},
author = {H. Ono and A. Yamashita and K. Amano and M. Sawayama},
howpublished = {bioRxiv},
year = {2026},
month = oct,
doi = {10.64898/2026.10.21.759598},
url = {https://www.biorxiv.org/content/10.64898/2026.10.21.759598v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Subspace Image Synthesis for Style-Invariant Object Recognition
AU - H. Ono
AU - A. Yamashita
AU - K. Amano
AU - M. Sawayama
PY - 2026
DA - 2026/10/21
PB - bioRxiv
DO - 10.64898/2026.10.21.759598
UR - https://www.biorxiv.org/content/10.64898/2026.10.21.759598v1
AB - Human-like behavior in visual models does not necessarily imply human-like internal computation. When a model recognizes objects robustly across changes in visual style, what internal computations support this behavior, and which are relevant to human perception? We addressed this question by pruning style-robust neural networks into candidate submodels corresponding to two computational strategies. The first retained channels whose responses were stable across style changes for fixed content, yielding a Content Stability Score (CSS)-based submodel. The second retained channels whose responses were stable across content changes for fixed style, yielding a Style Stability Score (SSS)-based submodel. Across ImageNet, Stylized ImageNet, and model-versus-human benchmarks, the two submodel families showed comparable object-recognition performance. We then used their exclusive feature subspaces to generate matched images with Gram-matrix-based neural style transfer and asked human observers which image was more similar in visual style to a reference image. Images generated from the CSS-based submodel were selected significantly more often than chance, even though CSS preferentially retained channels whose individual activation patterns were relatively insensitive to style. These results suggest that representations preserving content across style changes are more consistent with human judgments of visual style. More broadly, they demonstrate a framework for isolating candidate computational subspaces from large artificial models and testing their perceptual relevance through image synthesis.
ER -