Re: Questions - meta-questions about knowledge representation

Hello everyone,

Hope this thread is still alive.

Good questions, indeed. Specially risen today in the hype era of LLMs, RNNs
and 'other' technologies that sometimes seems too complicated, even for
so-called easy examples.

The use of formal and symbolic logics, e.g. FOL, assures the coherence,
correctness, and evolveness capabilities of the axiomatic system, relying
in well defined and structured hierarchy of descriptive logic languages
well-stablisheds. As a branch of the semantic networks, these logics were
proven very useful to structure knowledge systems, orchestrate services,
and permit RNNs, MLs, DNNs, to learn on top of the produced knowledge,
information, and data.

Some examples are easy found on chatgpt, e.g. aws, azure cognitive
services, google knowledge-graph, etc.

--

But, maybe they are not enough to express the interactivity of knowledge,
principally when uncertain knowledge is concerned.

4) if you use google you may direct see its evidences, I suppose. They use
ontologies to interpret the user entry, then organize the output in
semantic categories, e.g. albums, sales, colors, etc. Also, KRR is also
very useful in NLP, due to the linguistic representation it offers.

My meta-level question are 2:

1) why there's so little speaking of KRR evolution traceability, patterns
of tracing graph evolutions;

2) theres a non-linguistic method to express knowledge?

best,

Em dom., 8 de set. de 2024 10:42, Paola Di Maio <paola.dimaio@gmail.com>
escreveu:

>
> Apologies, for the late reply to this thread
>
> These are valid questions, bur can be answered in a variety of ways
> At the same time, they are possibly a bit speculative - I mean, what
> motivates this inquiry/
> There is so much literature,  I suspect you would have to find your own
> answers
>
> Some very quick thoughts  (I have no spare neurons) apologies for the
> brevity
>
> The non symbolic KR well, KR is symbolic (in the sense that it is explicit)
> but it can be used to represent, say, stochastic or probabilistic
> reasoning? (Discuss)
>
> Can we use KR say to represent the predictions as to how a network of
> mycelium is going to develop
> (I think so, maybe using a neural network visualisation approach that can
> be tweaked to reflect different
> conditions/assumptions/scenarios)
>
> especially if we start looking at the neurosymbolic KR (as the explicit
> representation of neurosymbolic AI)
>
> But hay, let us know when the thinking evolves -
> somehow this is related to D Raggets post today perhaps
> DR wrote:
> Further work is needed on declarative means to express strategies and
> tactics for argumentation, along with associated work on machine learning.
> It is probably going to be easier and more scalable to work on neural
> networks that learn and reason like we do if the goal is to construct tools
> to help with the kinds of applications envisaged by Lenat and others.
>
>
>
> On Fri, Jul 5, 2024 at 9:33 AM <ontologos@protonmail.com> wrote:
>
>> Here are some meta-questions (background questions about the why and
>> foundations) and devil's advocate questions that are not often (?) asked,
>> but helpful for non-specialists, students, and I assume even experts. Note,
>> I know some answers the following, but I would like to see the answers of
>> others. Also, some answers i'm aware of are superficial, so hopefully we
>> can have more substantive ones here.
>>
>> Context:
>>
>>    - When I studied knowledge representation and reasoning (KRR) and
>>    philosophy of AI, KRR was presented as one topic in AI. And symbolic and
>>    non-symbolic approaches were mentioned.
>>
>>
>> 1) Are there non-symbolic approaches to KRR?
>> (Approaches that do not use symbolic logics?)
>>
>> If so, what are they, and what are the pro's and con's of both
>> (non)symbolic KRR?
>> (it may be helpful to define what you mean by symbolic and non-symbolic
>> if different from the use of FOL and other logics)
>>
>> 2) Why use formal/symbol logics such as first-order predicate calculus
>> (or others)?
>> Do you have quantitative evidence for the benefits of them?
>> Can you give examples of successful projects that publicize quantitaitve
>> evidence for the benefits?
>>
>> 3) What are alternatives for KRR techniques for achieving the same goals
>> KRR aims to?
>>
>> 4) Do you have quantitative evidence for the utility of KRR in general?
>>
>> 5) ...add your own meta-level questions...
>>
>> NOTE: I've seen answers to 2 & 4 are often found as qualitative
>> descriptions, buzz-phrases, or hand-waving (like in many disciplines) but
>> without clear and specific quantiattive evidence. So please focus on that.
>> Quantiative evidence may involve: showing increases in efficiency of this
>> or that computational process, (by the numbers), etc.
>>
>> Thanks.
>>
>> Roberto
>> --
>>
>>