Rethinking Crowd-Sourced Evaluation of Neuron Explanations

This is the official repo of our paper Rethinking Crowd-Sourced Evaluation of Neuron Explanations. Full code and data will be released later, please stay tuned.

• In this work, we conduct the first crowd-sourced neuron explanation evaluation utilizing a proper evaluation metric, namely correlation coefficient that measures whether the explanation matches neuron activations beyond just the very highest activations
• We show how to effectively utilize Importance Sampling to select most important inputs to show raters, leading to ∼30× reduction in labeling cost over uniform sampling.
• We develop a Bayes-based method to aggregate predictions of different raters to deal with noisy labels, further reducing the cost of reaching a certain accuracy by ∼5×.

Cite this work

T. Oikarinen, G. Yan, A. Kulkarni and T.-W. Weng, Rethinking Crowd-Sourced Evaluation of Neuron Explanations, arXiv preprint, 2025.

@misc{oikarinen2025rethinking,
  title={Rethinking Crowd-Sourced Evaluation of Neuron Explanations}, 
      author={Tuomas Oikarinen and Ge Yan and Akshay Kulkarni and Tsui-Wei Weng},
      year={2025},
      eprint={2506.07985},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2506.07985}, 
}