# Explainability in Machine Learning: a Pedagogical Perspective

**Authors:** Andreas Bueff, Ioannis Papantonis, Auste Simkute, Vaishak Belle  
**arXiv:** [2202.10335](https://arxiv.org/abs/2202.10335)  
**Date:** 2022-02-21  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2202.10335v1/

## Key points

- Lack of pedagogical resources exploring the integration of explainability into machine learning
- Importance of teaching students and researchers about the advantages of explainability and how to implement techniques for interpreting a model's decision-making process
- Broad overview of explainable ML
- Exploration of transparent models vs opaque models frameworks
- Discussion on limitations and strengths of model-specific and model-agnostic post-hoc explainability approaches
- Insights on deep learning models and future research directions
- Exercises focused on six popular techniques for explainability using publicly available datasets
- Intermediate assignments evaluated based on workbooks utilizing post-hoc explainability methods discussed in the course
- Final assessment allows participants to choose their own project, incorporating at least two methods not covered in the course
- Use of Google Colab notebooks or similar platforms for implementation, no specific hardware required
- Participants work on their own projects or in groups throughout the course, going through all stages of posthoc explainability exploration
- Course consists of 10 lectures targeting different principles and techniques related to explainability
- Pre-recorded lectures followed by live sessions for discussion and debate, tutorial sessions for coding exercises and practical issues related to topics.
- Aims to promote inclusivity empowerment and responsibility in XAI by providing a comprehensive understanding of explainable ML techniques.
- Completion of the course gives data science professionals a birds-eye view of this rapidly developing area and confidence in deploying machine learning more widely.

## Summary

This course aims to address the lack of pedagogical resources exploring the integration of explainability into machine learning. It emphasizes the importance of teaching students and researchers in machine learning about the advantages of explainability and how to implement various techniques for interpreting a model's decision-making process. The course structure begins by providing a broad overview of explainable ML, followed by an exploration of two frameworks: transparent models vs opaque models. The limitations and strengths of model-specific and model-agnostic post-hoc explainability approaches are discussed, along with insights on deep learning models and future research directions. The course includes exercises focused on six popular techniques for explainability, allowing participants to gain hands-on experience with coding problems using publicly available datasets. Intermediate assignments are evaluated based on workbooks that utilize post-hoc explainability methods discussed in the course. For final assessment, participants can choose their own project, incorporating at least two methods not covered in the course. To facilitate implementation, students are encouraged to use Google Colab notebooks or similar platforms as no specific hardware is required. These notebooks provide a convenient coding environment for implementing various XAI techniques without worrying about individual hardware specifications or library dependencies. Throughout the course individuals work on their own projects or in groups going through all stages of posthoc explainability exploration. This includes exploring data and making initial findings, critiquing tools and identifying shortcomings, listing possible future features, and presenting an analysis. The course consists of 10 lectures targeting different principles and techniques related to explainability. Topics are provided as pre-recorded lectures following a flipped classroom model where live sessions allow for discussion and debate about the content from the pre-recorded lectures. Tutorial sessions provide opportunities for participants to engage in coding exercises and raise practical issues related to the topics. Overall this course aims to promote inclusivity empowerment and responsibility in XAI by providing a comprehensive understanding of explainable ML techniques. By completing the course data science professionals will gain a birds-eye view of this rapidly developing area and be confident in deploying machine learning more widely. The effectiveness of a recently delivered course following a similar structure is presented as evidence supporting the pedagogical approach.