explAIner

This repository contains the source code for explAIner -- the framework for explainable AI and interactive machine learning.

Architecture

The framework consists of four plugins, which represent the stages of explanation, namely

• Understanding
• Diagnosis
• Refinement
• Reporting

Repository Structure

Folders

The repository contains 4 folders:

• backend/ This folder contains the python backend for the high-level (model in-/output) explanations.

• tensorboard-explainer-plugin/
  This folder contains the actual explAIner code. It has the following sub-folders:

  • explainer_plugins/
    • _1_understanding/
      Plugin for understanding. Data-independent explanations.
    • _2_diagnosis/
      Plugin for diagnosis. Debugging of NN graph.
    • _3_refinement/
      Plugin for refinement. Recommendations on improvements.
    • _4_reporting/
      Plugin for reporting. Summarizes the findings from previous steps.
    • common/
      Parts that are used in more than one plugin.
  • explainer_tensorboard/
    The modified TensorBoard executable, with explAIner plugins injected.

Building and Starting

To create example logs for explAIner, run the following command and wait for it to finish:

docker-compose up --build --remove-orphans explainer_summary

To build and start the explAIner TensorBoard executable (together with custom backend servers):

docker-compose up --build --remove-orphans -d explainer_tensorboard

Although the containers should be up and running after a few seconds, it might take a while until the code is fully compiled and the system gets available under http://127.0.0.1:6006.

Citing this Repository

To reference this repository, please cite the original explAIner publication (pre-print available on arXiv.org):

T. Spinner, U. Schlegel, H. Schafer, and M. El-Assady, “explAIner: A Visual Analytics Framework for Interactive and Explainable Machine Learning,” IEEE Trans. on Vis. and Computer Graphics, vol. 26, no. 1, Art. no. 1, 2020, doi: 10.1109/tvcg.2019.2934629.

BibTeX

@ARTICLE{SpinnerEtAl2020,
  author = {Thilo Spinner and Udo Schlegel and Hanna Schäfer and Mennatallah El-Assady},
  title = {{explAIner}: A Visual Analytics Framework for Interactive and Explainable Machine Learning},
  journal = {{IEEE} Transactions on Visualization and Computer Graphics},
  year = {2020},
  volume = {26},
  number = {1},
  pages = {1064-1074},
  doi = {10.1109/TVCG.2019.2934629},
}