Skip to content

README.md

Latest commit

 

History

History
130 lines (85 loc) · 2.62 KB

README.md

File metadata and controls

130 lines (85 loc) · 2.62 KB

Path Attributions for KAN

This is the repository for the [TMLR'26] paper entitled "Interpreting Kolmogorov-Arnold Networks in Neuroimaging: A Path-Based Attribution Framework", which is an attribution method for Kolmogorov–Arnold Networks (KANs) and other similar additive networks. The framework computes global feature importance by propagating edge scores across network paths. The method produces interpretable global importance maps at input level encompassing all the information learnt by the model. The paper can be found at https://openreview.net/pdf?id=cPtKpNdYc2. Please cite the same if referencing this work.

Installation

Clone the repository and install dependencies:

git clone https://github.com/val-iisc/KAN-PathAttribution.git
cd KAN-PathAttribution
pip install -r src/requirements.txt

Or using conda:

conda env create -f src/environment.yml
conda activate pathkan

MNIST Demo

An example MNIST experiment is provided to demonstrate the full pipeline.

Run:

python misc/mnist_demo/mnist_demo.py

This script:

  • trains a KAN model
  • computes edge scores
  • performs path-based attribution
  • generates a feature importance heatmap

Attribution Pipeline

Compute edge scores:

from src.methods.edge_scoring import attribute_edges

edge_scores = attribute_edges(model, method="l2", tanh=True)

Aggregate importance across network paths:

from src.methods.path_attribution import path_attribution

global_importances = path_attribution(edge_scores)

Output format:

{"Input_0": importance, "Input_1": importance, ...}

Visualization

2D spatial heatmaps:

from misc.visualization.heatmap_2d import create_heatmap_from_dict

heatmap = create_heatmap_from_dict(global_importances, height=28, width=28)

3D neuroimaging visualization:

from misc.visualization.brain_3d_interactive import dict_to_3d_array

Interactive example:

misc/assets/OASIS_Volume_Interactive.html

Datasets

Experiments were conducted on:

  • OASIS-1
  • ADNI
  • Mendeley Neuroimaging Dataset
  • MNIST (demo)

Datasets are not included due to licensing restrictions. See:

misc/data/dataset_instructions.md

License

MIT License.

Citation

@article{murthy2026interpreting,
  title={Interpreting Kolmogorov-Arnold Networks in Neuroimaging: A Path-Based Attribution Framework},
  author={Suhrud Murthy and Venkatesh Babu Radhakrishnan and Neelam Sinha},
  journal={Transactions on Machine Learning Research},
  issn={2835-8856},
  year={2026},
  url={https://openreview.net/forum?id=cPtKpNdYc2},
  note={}
}