RE: Toward a web standard for XAI?

Paola Di Maio,

When considering explanations of artificial intelligence systems’ behaviors or outputs or when considering arguments that artificial intelligence systems’ behaviors or outputs are correct or the best possible, we can consider diagrammatic, recursive, component-based approaches to the design and representation of models and systems (e.g. Lobe). For such approaches, we can consider simple components, interconnections between components, and composite components which are comprised of interconnected subcomponents. For such approaches, we can also consider that components can have settings, that components can be configurable.

As we consider recursive representations, a question is which level of abstraction should one use when generating an explanation – when composite components can be double-clicked upon to reveal yet more interconnected components? Which composite components should one utilize in an explanation or argument and which should be zoomed in upon and to which level of detail? We can generalize with respect to generating explanations and arguments from recursive models of: (1) mathematical proofs, (2) computer programs, and (3) component-based systems. We can consider a number of topics for all three cases: explanation planning, context modeling, task modeling, user modeling, cognitive load modeling, attention modeling, relevance modeling and adaptive explanation.

Another topic important to XAI is that some components are trained on data, that the behavior of some components, simple or composite, is dependent upon training data, training procedures or experiences in environments. Brainstorming, we can consider that components or systems can produce data, e.g. event logs, when training or forming experiences in environments, such that the produced data can be of use to generating explanations and arguments for artificial intelligence systems’ behaviors or outputs. Pertinent topics include contextual summarization and narrative.

XAI topics are interesting; I’m enjoying the discussion. I hope that these theoretical topics can be of some use to developing new standards.

Best regards,

Schiller, Marvin, and Christoph Benzmüller. "Presenting proofs with adapted granularity." In Annual Conference on Artificial Intelligence, pp. 289-297. Springer, Berlin, Heidelberg, 2009.

Cheong, Yun-Gyung, and Robert Michael Young. "A Framework for Summarizing Game Experiences as Narratives." In AIIDE, pp. 106-108. 2006.

From: Paola Di Maio<>
Sent: Monday, November 12, 2018 9:59 PM
Cc:<>; semantic-web at W3C<>
Subject: Re: Toward a web standard for XAI?

Dear Adam
thanks and sorry for taking time to reply.
Indeed triggered some thinking
In the process of doing so,irealised whatever we come up with has to match the web stack, and then realised that we do not have a stack for the distributed web yet, as such
Is this what you are thinking, Adam Sobieski, please share more
sounds like in the right direction

Artificial intelligence and machine learning systems could produce explanation and/or argumentation [1].

Deep learning models can be assembled by interconnecting components [2][3]. Sets of interconnected components can become interconnectable composite components. XAI [4] approaches should work for deep learning models assembled by interconnecting components. We can envision explanations and arguments, or generators for such, forming as deep learning models are assembled from components.

What do you think about XAI and deep learning models assembled by interconnecting components?

Best regards,
Adam Sobieski


From: Paola Di Maio<>
Sent: Wednesday, October 31, 2018 9:31 AM
To:<>; semantic-web at W3C<>
Subject: Toward a web standard for XAI?

Just wondering

Received on Saturday, 17 November 2018 05:40:55 UTC