RE: Toward a web standard for XAI?

Dave,

Thank you for the interesting points on combining computational statistics with symbolic representations and reasoning.

With respect to species of animals and how they can sense, perceive and exhibit behavior so proximate to birth, one might find interesting the topics of evolution, developmental neurogenesis [1], instinct [2] and ethology [3].

Brainstorming with respect to XAI, there might also be value in considering computer simulation, event stream processing and complex event processing. That is, we can consider generating events as connectionist systems compute so that other software systems can process and interpret the events into descriptions of what the connectionist systems are doing instantaneously or over the course of time.

With respect to considering new formats and Web standards for XAI explanations and arguments, one might find interesting natural language generation as pertaining to explanation [4] and argumentation [5][6].


Best regards,
Adam

[1] https://en.wikipedia.org/wiki/Neurogenesis#Developmental_neurogenesis
[2] https://en.wikipedia.org/wiki/Instinct
[3] https://en.wikipedia.org/wiki/Ethology

[4] http://www.phoster.com/mixed-initiative-dialogue-systems-explanation-mechanistic-reasoning-mental-simulation-and-imagination/generating-explanations/
[5] http://www.phoster.com/linguistics/argumentation/
[6] Reed, Chris Anthony. "Generating arguments in natural language." PhD diss., University of London, 1998.

________________________________
From: Dave Raggett <dsr@w3.org>
Sent: Thursday, November 1, 2018 5:14:43 AM
To: Adam Sobieski
Cc: paoladimaio10@googlemail.com; public-aikr@w3.org; semantic-web at W3C
Subject: Re: Toward a web standard for XAI?

It is certainly interesting, but I expect that there are larger opportunities if we look at opportunities to combine computational statistics with symbolic representations and reasoning.  An example is vision, where deep learning works well if you have large training sets and the data you want to apply the trained network to reflects the same statistics as the training set. Unfortunately, that is often not the case, e.g. due to changes in lighting, weather.

Many species of animals are able to see very soon after birth and clearly aren’t reliant on huge training sets. Moreover, they are able to perceive objects that they have never seen before.  This calls for a more sophisticated architecture than stochastic back propagation, one that can embody induction based upon commonalities and abduction for inferring the presence of a previously learned object from cues. Moreover, for survival, animals need to recognise behaviour and to distinguish predators from others. This means learning to spot patterns of behaviour and to associate then with a class of things that have been learned by induction.


On 1 Nov 2018, at 07:35, Adam Sobieski <adamsobieski@hotmail.com<mailto:adamsobieski@hotmail.com>> wrote:

Artificial Intelligence Knowledge Representation Community Group,
Semantic Web Interest Group,
Paola Di Maio,

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
http://www.phoster.com/contents/

[1] https://www.w3.org/community/argumentation/
[2] https://www.lobe.ai/
[3] https://www.youtube.com/watch?v=IN69suHxS8w
[4] https://www.darpa.mil/program/explainable-artificial-intelligence


Dave Raggett <dsr@w3.org<mailto:dsr@w3.org>> http://www.w3.org/People/Raggett
W3C Data Activity Lead & W3C champion for the Web of things