eXplainable Artificial Intelligence

A novel classification system

The success and widespread application of Artificial Intelligence (AI) in several fields, spanning from medicine to autonomous transports. Despite this, most of the predictive models beloging to the AI field lack intepretability and transparency because of the non-linearity and complexity of their underlying structures. This can hinder the acceptance of these models by the wider public as people do not trust what they cannot understand. Thus, eXplainable Artificial Intelligence (XAI) has become a vital sub-field of AI. Its ultimate goals is building a unified approach to learn predictive models that possess high accuracy and a high degree of explainability. A plethora of XAI methods, proposing different strategies to reach this goal, have been developed so far. This has led to the need to organise this abudance of knowledge into a comprehensive system capable to structure at high-level the current state-of-the-art in XAI.

As part of my PhD project in XAI, I carried out a thorough literature review that evolved into a journal paper, currently under peer-review (check out its pre-print version). In this study, I propose a new hierarchy that organises XAI in four main categories mirroring the research activities currently performed by scholars: developing new XAI methods, based on a set of notions and attributes linked to the teoretical construct of explainability, evaluating their efficacy by proposing formal, objective metrics or human-centred evaluation approaches involving end-users and model designers and, eventually, organising all these methods in reviews. The following sunburst chart shows the number of studies for each category. Some studies are classified under more than one category and they are counted more than once, thus the numeric values of the inner layers are affected by multiple counting.

The proposed hirearchy is depicted in the explandable tree below. Click on the blue circles to expand each branch into its sub-branches. The leaves represent scientific articles and, by clicking on them, you are redirected to the original study. The full list of the scientific articles can be examined here.

Other resources

GITHUB repositories:

• Lucid, a collection of tools for neural network interpretability
• Awesome Interpretable Machine Learning
• DeepExplain: attribution methods for Deep Learning
• Interpretability and explainable machine-learning models
• Responsible Machine Learning
• SCIKIT-OPTIMIZE. Sequential model-based optimization in Python
• Diverse Counterfactual Explanations (DiCE) for ML
• What-If Tool

Online articles on XAI:

• An introduction to explainable AI, and why we need it
• The What-If Tool: Code-Free Probing of Machine Learning Models
• A New Approach to Understanding How Machines Think
• Inceptionism: Going Deeper into Neural Networks
• Hands-on Machine Learning Model Interpretation
• The AI Black Box Explanation Problem
• Learning explanatory rules from noisy data
• Google’s AI explains how image classifiers made their decisions
• How can we make sure that algorithms are fair?
• AI and the Question of Explainability
• The How of Explainable AI: Pre-modelling Explainability
• The How of Explainable AI: Explainable Modelling
• The great AI debate: Interpretability
• If it’s interpretable it’s pretty much useless.
• How Explainable Artificial Intelligence (XAI) Can Help Us Trust AI
• Interpretable Machine Learning with XGBoost
• SHAP: How to interpret machine learning models with Python
• LIME vs. SHAP: Which is Better for Explaining Machine Learning Models?

XAI researchers, online version of scientific papers and dedicated websites:

• Feature Visualization
• Been Kim, senior research scientist at Google Brain
• Understanding Neural Networks Through Deep Visualization
• The Building Blocks of Interpretability
• heatmapping.org
• LSTMVis - Visual Analysis for Recurrent Neural Networks
• Understanding Neural Networks Through Deep Visualization
• Learning Interpretable Models
• Symbolic Graph Reasoning Meets Convolutions, NeurIPS 2018
• Leverhulme Centre for the Future of Intelligence
• EasyMiner. Easy association rule mining, classification and anomaly detection
• Google Teams. People + AI Research (PAIR)
• A visual introduction to machine learning

Books on XAI:

• Transparent Computers: Designing Understandable Intelligent Systems
• Interpretable Machine Learning
• Explainable and Interpretable Models in Computer Vision and Machine Learning