Re: KR, Logic in a few slides

Thank you Dave
the lecture slides I pointed to a classical texbook answer which summarises
my reply to your question
*how to use KR to identify Deepfakes (short answer: by using logic)*
Please show us how factfinding and reasoning  can be done without KR
(I may not be able to read/reply all the posts apologies but have given my
availability to participate in a panel so that we can have this discussion
pne day)


On Tue, Nov 1, 2022 at 10:23 PM Dave Raggett <dsr@w3.org> wrote:

> The lecture focuses on traditional logic and ignores the imprecision and
> context sensitivity of natural language. As a result it sounds rather dated.
>
> The other lectures in the course:
>
>     https://www.cs.princeton.edu/courses/archive/fall16/cos402/
>
> They cite Chomsky’s comments deriding researchers in machine learning who
> use purely statistical methods to produce behaviour that mimics something
> in the world, but who don't try to understand the meaning of that behaviour.
>
> This of course predates recent work on large language models which
> demonstrate a very strong grasp of language and world knowledge, but are
> rather weak in respect to reasoning. The brittleness of both hand-authored
> knowledge and deep learning, motivates work on machine learning of
> reasoning for everyday knowledge.
>
> The last twenty years have shown that machine learning is vastly superior
> to hand authoring when it comes to things like speech, text and image
> processing. However, we have still to successfully mimic human learning and
> reasoning.
>
> To get there, I believe that some hand authoring for small scale
> experiments can help illustrate what’s needed from more scalable
> approaches.  I don’t see anyone else on this list being interested, though,
> in helping with that. Where are the programmers and analysts when you need
> their help?
>
> I disagree that the slides you linked to are effective in arguing for
> applying KR to ML.  You could say that ML requires a choice of KR since ML
> software has to operate with information in some form or other, but that is
> like stating the obvious.
>
> Perhaps you are assuming that KR should use a high level theoretical
> formalism for knowledge? That’s wishful thinking as far as I am concerned.
> It certainly didn’t help when it came to former stretch goals for AI, e.g.
> beating chess masters at their own game.
>
> On 1 Nov 2022, at 13:25, Paola Di Maio <paola.dimaio@gmail.com> wrote:
>
>  Sweet and short set of slides that speaks CS language in making the
> argument of KR for ML, the authors  may be instructors at Princeton and do
> not cite any literature, but do a good job at summarizing the main point-
>
> Lecture 12: Knowledge Representation and
> Reasoning Part 1: Logic
>
> https://www.cs.princeton.edu/courses/archive/fall16/cos402/lectures/402-lec12.pdf
> <https://www.google.com/url?sa=i&rct=j&q=&esrc=s&source=web&cd=&cad=rja&uact=8&ved=0CAQQw7AJahcKEwi40bXYiI37AhUAAAAAHQAAAAAQAw&url=https%3A%2F%2Fwww.cs.princeton.edu%2Fcourses%2Farchive%2Ffall16%2Fcos402%2Flectures%2F402-lec12.pdf&psig=AOvVaw1dCxdvDmEmj3r5EiLn5UGV&ust=1667395111043516>
>
> Some may appreciate the presence of a penguin illustrating the argument
>
>
> Dave Raggett <dsr@w3.org>
>
>
>
>