Stimulus symmetries can confound representational similarity analyses

This folder contains the codes and the associated data to replicate the main findings of the paper:

Stimulus symmetries can confound representational similarity analyses, Authors: Farhad Pashakhanloo and Jacob A. Zavatone-Veth https://arxiv.org/abs/2605.21324

There are two main jupyter notebooks that can be executed as standlone Python scripts:

• RunToyModels.ipynb: This script replicates all the results for the toy model. The figures are saved under output/ folder.

• RunRealData.ipynb: This script replicates all the results related to the section with latent symmetry in image data . The saved neural network models are currently under models/ folder. However, new models can be trained and saved (use appropriate flags in the code to enable this).

These codes were tested with:

• python 3.10.18
• numpy: 1.26.4
• torch: 2.8.0
• matplotlib: 3.10.5
• scikit-learn: 1.7.1
• scipy: 1.15.2
• Pillow (PIL): 11.3.0