understandable-machine-intelligence-lab / NoiseGrad

NoiseGrad (and its extension NoiseGrad++) is a method to enhance explanations of artificial neural networks by adding noise to model weights

NoiseGrad and FusionGrad

NoiseGrad: enhancing explanations by introducing stochasticity to model weights

Pytorch implementation

---

Pytorch implementation for "NoiseGrad: enhancing explanations by introducing stochasticity to model weights". The paper introduces two novel methods NoiseGrad and FusionGrad which both improves attribution-based explanations by introducing stochasticity to the model parameters. See arXiv preprint: https://arxiv.org/abs/2106.10185.

Cite this paper

To cite this paper use following Bibtex annotation:

@misc{bykov2021noisegrad,
      title={NoiseGrad: enhancing explanations by introducing stochasticity to model weights},
      author={Kirill Bykov and Anna Hedström and Shinichi Nakajima and Marina M. -C. Höhne},
      year={2021},
      eprint={2106.10185},
      archivePrefix={arXiv},
      primaryClass={cs.LG}}

Requirements

To install requirements:

pip install -r requirements.txt

All experiments were conducted with Python 3.6.9.

Code structure

The source code can be found in the src/ folder and an example notebook in examples/ folder.