What Does Explainable AI Really Mean? A New Conceptualization of Perspectives

Transcript

1. What does Explainable AI Really Mean? Doran, Schulz, and Besold,

   2017 Golestan Sally Radwan PWL NYC, 27 June 2018

2. The Paper What does Explainable AI really mean? A new

   conceptualization of perspectives. By: Derek Doran, Sarah Schulz and Tarek Besold Pre-print: arXiv:1710.00794v1, 2 October 2017

3. Main Premise Explainable AI is a hot topic in research

   and parts of industry But there is no uniform understanding of what it really entails Also a polarizing topic: some strong supporters and some staunch opponents

4. Why I love it Topic is near and dear to

   my heart :-) Fairly short and accessible Gives a good overview of current research and uses Attempts to clarify and classify types of explainability Sets the groundwork for further development on the topic

5. Intro “ If you were held accountable for the decision

   of a machine in contexts that have financial, safety, security, or personal ramifications to an individual, would you blindly trust its decision? How can we hold accountable Artificial Intelligence (AI) systems that make decisions on possibly unethical grounds, e.g. when they predict a person’s weight and health by their social media images or the world region they are from as part of a downstream determination about their future, like when they will quit their job, commit a crime, or could be radicalized into terrorism.”

6. What is explainability? Oxford Dictionary definition of ‘explainable’: A statement

   or account that makes something clear; a reason or justification given for an action or belief. Based on this definition, none of the existing AI systems are actually explainable!

7. 3 types of AI systems

8. Example: Doctors Opaque: Provide diagnosis and treatment with no elaboration

   Comprehensible: Give high-level indicators on symptoms and test results Interpretable: Step-by-step walkthrough of reasoning and decisions Your target audience determines the level of explainability you need!

9. Additional traits of an explainable system

10. “Truly Explainable AI” Current models enable interpretation but leave it

    to the user to apply their own knowledge, bias, and understanding. This can be dangerous as based on the above factors, different people might come up with different explanations for the same decision. They lack a line of reasoning that explains the decision-making process of a model using human-understandable features of the input data.

11. The Solution Augmenting comprehensible models with a reasoning engine, combining

    symbols emitted with a domain-specific knowledge base encoding relationships between concepts presented by the symbols.

12. What is your idea of Explainable AI?