The article discusses the concept of interpretation in artificial intelligence (AI) and its significance in designing systems that can generate understanding via explanations post interpretations. The authors argue that the term "explainability" is often used interchangeably with "understandability," but they propose a framework to differentiate between the two concepts. They suggest that explainability refers to the ability of a system to provide reasons or details to make its functioning clear, while understandability refers to the ease with which an audience can comprehend the explanation.
The authors identify three main categories of XAI approaches: (1) model-based explanations, which rely on the internal workings of the AI model; (2) example-based explanations, which provide concrete examples to illustrate how the model arrived at its predictions; and (3) hybrid approaches, which combine both model-based and example-based explanations. They also propose a unified view of XAI, where the AI system can serve multiple kinds of audiences with different expectations regarding explanations.
The authors emphasize that interpretation is a crucial step in converting poor explanations into good explanations that lead to ease of understanding. They highlight the importance of considering the audience’s perspective and their level of understanding when designing XAI systems. The article concludes by suggesting that XAI has the potential to improve trust in AI systems and enhance their transparency and accountability.
Analogy
Imagine you are trying to explain a complex recipe to a friend who is not familiar with cooking. You could provide detailed instructions, such as "first, preheat the oven to 350 degrees" or "next, mix together flour, sugar, and eggs." These explanations would be similar to model-based explanations in XAI, as they rely on the internal workings of the recipe. Alternatively, you could provide concrete examples, such as "imagine a cake with chocolate frosting" or "picture a warm cookie straight from the oven." These examples would be more similar to example-based explanations in XAI, as they illustrate how the recipe can be applied in practice. By combining both types of explanations, you could provide a more comprehensive understanding of the recipe and make it easier for your friend to follow.
