Sensie: Probing the sensitivity of neural networks

Introduction

Sensie’s goal is to quickly interrogate a trained machine learning model, determining the sensitivity to a parameter or perturbation of the data.

Sensie probes the sensitivity of a network to inputs with a particular property, p. This property can be a feature of the data; an otherwise known property of a test or training set that is not provided explicitly in training; or a function that can artificially vary this property for a supplied test set. The effect of a particular property is measured according to the variance it introduces to the correct output of the network, such as the score for the correct class . Quantitatively, we would like to know the function *mean_score = f(p) for some property p; Sensie can calculate a linear approximation to this unknown function.

For more information and examples, see the GitHub repository at https://github.com/coljac/sensie.

Installation

Check out the repository and install with:

pip install .

(or add the sensie directory to your PYTHONPATH.)

Dependencies are listed in requirements.txt included in the repository. Sensie requires python 3.6 or above.

Optionally, install pytest with pip install pytest, then run the tests with pytest test from the repository root.

API documentation:

• sensie package
  • Module contents

Issues, Questions and Contributions

Any problems or questions? Email colin@coljac.net, or open an issue on GitHub at https://github.com/coljac/sensie.

Contributions are welcome and encouraged. Fork the GitHub repository to your own machine, make some changes, and push your work back up to the fork and open a pull request so that I can review and incorporate the changes.

Index

• Index
• Module Index
• Search Page