Representational Similarity via Interpretable Visual Concepts (RSVC)

Setup

conda create -n "RSVC" python=3.8.18
conda activate RSVC
bash setup.sh

Downloading Datasets

1. ImageNet: You can download imagenet with bash download_imagenet.sh.
2. NABirds: Please visit https://dl.allaboutbirds.org/nabirds to request access and download the dataset.

The datasets should be placed (or symlinked) in the data/ directory. The structure should look like this:

data/
│-- imagenet/
│   │-- train/
│   │-- val/
│   │-- ILSVRC2012_devkit_t12/
│   │-- ...
|-- nabirds/
    |-- images/
    |-- parts/
    |-- ...

Usage

Toy Concept Comparison

To run the toy concept comparison experiment, make sure the nabirds dataset is downloaded. You can download the custom trained models from M_ps and M_nc.

bash toy_concept_experiment.sh

RN18 vs. RN50

To run the resnet size comparison make sure the imagenet dataset is downloaded and placed in the correct location. These models should automatically download from the timm library.

bash compare_rn18_rn50.sh

TODO

• Add qualitative visualization code to resnet comparisons
• Add remaining comparison experiments

Citation

@inproceedings{kondapaneni2025representational,
  title={Representational Similarity via Interpretable Visual Concepts},
  author={Kondapaneni, Neehar and Mac Aodha, Oisin and Perona, Pietro},
  journal={The Thirteenth International Conference on Learning Representations},
  year={2025}
}