- From: Paola Di Maio <paola.dimaio@gmail.com>
- Date: Tue, 13 Oct 2020 13:39:59 +0800
- To: W3C AIKR CG <public-aikr@w3.org>, public-cogai <public-cogai@w3.org>
Greetings all, apologies for lack of activity on the lists, I ll soon post a summary of AI KR related contributions and a bottom line conclusions of the exploratory work done in the last couple of years we can then decide whether ther is enough material for a report and the future of the AI KR CG I summarize this email the feedback for the NIST draft on explainability 9due 15 Oct) as it is related to KR https://www.nextgov.com/emerging-tech/2020/08/nist-releases-core-principles-judge-explainable-ai/167833/ with a draft voice narration (14 minutes) which I think I ll record again as this is a bit ranty "https://www.loom.com/share/0f2cd37e8f854cd9b788def429b7e283" FEEDBACK NOTES FOR NIST ON EXPLAINABILITY Draft NISTIR 8312 from PAOLA DI MAIO. 13 October 2020 PREAMBLES a) before explanaibility can be addressed in the context of AI, AI should be better understood/defined. The reality is that we may not yet have AI after all b) In addition to the distinction between narrow and general AI, the distinction closed vs open system AI is also necessary. This particularly applies to the point Knowledge limits in the draft. GENERAL COMMENTS ON THE PRINCIPLES IN THE DRAFT 1. EXPLANATION type mismatch among the principles for example explanation, is a noun, while meaningful is an adjective, would be advisable to have some consistency in the naming conventions? 2. MEANINGFUL explanation is described as a prinicple that mandates an explanation for AI, and meaningful is described as a principle that the explanation is meaningful, but it does not describe criteria/pameters for meaningfulness. This does not seem up to standard. Looks to me that meaningful is a qualifier for explanation (1) 3. EXPLANATION ACCURACY - same as above, this does not seem a principle more like a qualifier for principle 1. Looks to me that 2 and 3 are qualifiers for 1. hoever they should be better defined 4. Knowledge Limits - this is new (ie. unheard of) Is there a reference for such a notion? Where does it come from? who may have come up with such an idea? Intelligence can be said to overcome knowledge limits, ie, given limited knowledge an intelligent process relies on logical inferences deduction, abduction to achieve a conclusion. Reasoning with limited knowledge is a defining characteristic of intelligent systems. Furthermore in open systems, knowledge is not limited, by contrast, it is continually updated with new knowledge. To consider limited knowledge for intelligent systems/AI is a contradiction in terms. A knowledge limit applies to closed database systems not to AI. OTHER = In addition to meaningful and accurate, explanations should also be timely, accessible, updatable etc - symbolic KR is central to subsymbolic explainability and should be mentioned in this document -there should be a standard for systems explainabilty ------------------------------------- On Thu, Aug 20, 2020 at 7:44 PM Paola Di Maio <paola.dimaio@gmail.com> wrote: > > I could not understand the word limits in the context of this list, but seems to be used in this report > https://www.nextgov.com/emerging-tech/2020/08/nist-releases-core-principles-judge-explainable-ai/167833/ > > if anyone want to provide input to the report lets gather them here first? > > PDM
Received on Tuesday, 13 October 2020 05:40:59 UTC