FAT Forensic Events

An interactive demonstration of machine learning explainers for The Alan Turing Institute's AI UK 2021

Learn more about FAT Forensics: Source Code Documentation

Resources: Recordings Slides Jupyter Notebooks Slack

Did You Get That?

Reviewing Intelligibility of State-of-the-art Machine Learning Explanations

An interactive demonstration of machine learning explainers for The Alan Turing Institute's AI UK 2021.
Where: Zoom, The Alan Turing Institute, London, UK.
When: 11.00–11.15am on Wednesday, March 24th, 2021.

Table of Contents

About the Demonstration

To better understand data-driven predictive models and their decisions, the artificial intelligence and machine learning communities are developing dedicated transparency and explainability methods. While such techniques tend to provide appealing insights, their provenance and correctness are rarely ever assessed. In this session I demonstrate how distinct – and sometimes contradictory – explanations can be generated with one explainability approach for a single black-box prediction. I show how simple changes to the parameterisation of explainability methods can yield such disparate results1. This observation suggests that each application may require a bespoke explainer built and tuned for the problem at hand. Moreover, understanding these dependencies helps with designing robust and trustworthy explainers, whose insights can be relied upon.

FAT Forensics (Software)

To support the goals of this demonstration, we employ FAT Forensics – an open source Python package that can inspect selected fairness, accountability and transparency aspects of data (and their features), models and predictions. The toolbox spans all of the FAT domains because many of them share underlying algorithmic components that can be reused in multiple different implementations, often across the FAT borders. This interoperability allows, for example, a counterfactual data point generator to be used as a post-hoc explainer of black-box predictions on one hand, and as an individual fairness (disparate treatment) inspection tool on the other. The modular architecture23 enables FAT Forensics to deliver robust and tested low-level FAT building blocks as well as a collection of FAT tools built on top of them. Users can choose from these ready-made tools or, alternatively, combine the available building blocks to create their own bespoke algorithms without the need of modifying the code base.

Resources

The presentation is provided as an interactive set of slides utilising iPyWidgets and built with a Jupyter Notebook based on RISE and reveal.js. Additionally, an independent Jupyter Notebook with all of the interactive widgets is available to facilitate hands-on experimentation with the explainers introduced during the demonstration. These notebooks (hence the presentation) can be executed locally on one’s own machine or launched directly in the web browser through Google Colab or MyBidner. The recording of the demonstration session is available on YouTube.

Resources
Slides
Notebook
Recording

Instructors

Kacper Sokol

Kacper is a research associate with the TAILOR project at the University of Bristol. His main research focus is transparency – interpretability and explainability – of machine learning systems. In particular, he has done work on enhancing transparency of logical predictive models (and their ensembles) with counterfactual explanations. Kacper is the designer and lead developer of the FAT Forensics package.

Contact
K.Sokol@bristol.ac.uk

References

  1. Kacper Sokol, Alexander Hepburn, Raul Santos-Rodriguez, and Peter Flach. 2019. bLIMEy: Surrogate Prediction Explanations Beyond LIME. 2019 Workshop on Human-Centric Machine Learning (HCML 2019) at the 33rd Conference on NeuralInformation Processing Systems (NeurIPS 2019), Vancouver, Canada (2019). https://arxiv.org/abs/1910.13016 ↩

  2. Kacper Sokol, Alexander Hepburn, Rafael Poyiadzi, Matthew Clifford, Raul Santos-Rodriguez, and Peter Flach. 2020. FAT Forensics: A Python Toolbox for Implementing and Deploying Fairness, Accountability and Transparency Algorithms in Predictive Systems. Journal of Open Source Software, 5(49), p.1904. https://joss.theoj.org/papers/10.21105/joss.01904 ↩

  3. Kacper Sokol, Raul Santos-Rodriguez, and Peter Flach. 2019. FAT Forensics: A Python Toolbox for Algorithmic Fairness, Accountability and Transparency. arXiv preprint arXiv:1909.05167. https://arxiv.org/abs/1909.05167 ↩