Explainable AI

Denis de Montigny PhD

University College London

Co-founder Qlicks.io

Rem Sadykhov MSc

University College London

Co-founder Qlicks.io

This paper provides an introduction to Explainable AI. We aim to provide a more in depth discussion of related issues on this topic in separate videos and papers.

Explainable AI (XAI) is a subfield of artificial intelligence (AI) that focuses on developing systems that can provide clear and interpretable explanations for their decisions and actions. The goal of XAI is to create AI systems that are transparent, accountable, and trustworthy.

One of the main challenges in XAI is that many modern AI systems, such as deep learning models, are highly complex and their decision-making processes are difficult to understand. This can make it difficult for humans to trust the decisions made by these systems, especially in critical applications such as healthcare, finance, and criminal justice.

There are several approaches to XAI, including:

  • Model interpretability: This approach focuses on making the internal workings of AI models more transparent and understandable to humans. Techniques such as feature importance analysis, saliency maps, and layer-wise relevance propagation can be used to understand how the model is using input features to make decisions.

  • Post-hoc explanation: This approach involves generating explanations after the model has made a decision. This can be done using techniques such as rule extraction and case-based reasoning.

  • Transparent models: This approach involves designing AI models from the ground up to be more interpretable, such as decision trees and linear regression models.

  • Human-AI collaboration: This approach involves designing AI systems that can work together with humans to make decisions. The AI system can provide explanations for its decisions to help humans understand and trust the results.

XAI is a rapidly growing field and there is ongoing research in developing new methods and techniques for creating transparent and interpretable AI systems. Some important challenges that need to be addressed include:

  • Balancing transparency and performance: Finding the right balance between making AI models transparent and maintaining their performance can be difficult.

  • Handling complex and dynamic systems: Many real-world AI systems are highly complex and operate in dynamic environments. This can make it difficult to provide clear explanations for their decisions.

  • Addressing ethical and social issues: XAI also raises important ethical and social issues, such as ensuring that AI systems do not perpetuate existing biases or discriminate against certain groups.

The field of explainable AI (XAI) is still a relatively young and rapidly evolving field, and there is ongoing research and debate about the best approaches and techniques for building explainable AI systems. However, there are several papers that are often cited as canonical works in the field.

  • One of the most frequently cited papers on explainable AI is "Why Should I Trust You?": Explaining the Predictions of Any Classifier" by Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin, published in the Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining (KDD '16) in 2016. This paper introduces a technique called Local Interpretable Model-Agnostic Explanations (LIME), which can be used to generate human-interpretable explanations for the predictions of any machine learning model. LIME works by approximating the decision boundary of a model in the neighborhood of a particular prediction and using this approximation to build a simpler, interpretable model that can be used to explain the prediction.

  • Another important paper in the field of explainable AI is "Interpretable Machine Learning: A Brief History, State-of-the-Art and Challenges" by Christoph Molnar, published in the journal arXiv in 2019. This paper provides an overview of the history of interpretable machine learning, describes the current state of the field, and identifies key challenges and future directions for research.

  • Other notable papers in the field of XAI include "Towards A Rigorous Science of Interpretable Machine Learning" by Finale Doshi-Velez and Been Kim, published in the journal arXiv in 2017, and "Explainable Artificial Intelligence (XAI)" by David Gunning, published in the Defense Advanced Research Projects Agency (DARPA) Explainable Artificial Intelligence (XAI) program in 2017.

These papers provide an overview of recent research in the field of XAI and gives a good understanding of the current state of the field and the challenges it is facing. However, new papers are published frequently and it is always good to keep an eye on the latest publications in the field.

Overall, Explainable AI is a crucial aspect of AI development, it is important to ensure that AI systems are transparent, accountable, and trustworthy. This will help to build trust and confidence in these systems and make them more widely adopted in a variety of applications.

Join us today and help us build the future !