Why Michael's talk disappointed me.
gregsharp73
First off, please forgive the sensational headline. The talk was really good. I see how discovering mathematical analogs for ontological patterns could be really valuable.
But let me explain. Michael promised to tackle two points framed by Hilbert’s sixth problem:
1. ontology for a physical theory
2. ontology as a physical theory
In doing this he brought up an ontology of chemical molecules.
The first point was that this ontology should be adequate to model all the molecules that chemists have found within some area or research. It is also good if the ontology is precise so as not to include unintended molecules. Sounds good.
To the second point, we can suppose that there are molecules that the ontology suggests are either possible or impossible. Sounds great.
Right there, I was expecting a bold jump into idealism and instead I heard a retreat into pragmatism.
How so? Because when the ontology treads into the unknown, the only recourse he presented was empirical. We go back to the chemists and they tell us whether the molecule exists, doesn’t exist or couldn’t exist and then we place some extension or constraint on the ontology to account for this. Fine.
But given that LLMs are already being used to predict molecules (.ie AlphaFold), doesn’t this make ontologies like this feel a little meh (to unhiply quote my children)?
Should there not be a spark of idealism in the ontologist that wants to believe that their model has what it takes not just to describe reality but to anticipate it (I would say predict but AI already overloaded that term)?
What’s my point? If symbolic AI is to keep up with and inform statistical AI, don’t we need to ask more of our ontologies? Don’t they need to anticipate the molecules that do, don’t yet, and never could exist? Or will ontologizing always be a reactionary extension or constraint here or there, as needed, like Ptolemaic epicycles bound by pragmatic interests, or shouldn’t we continue to be searching for the correct, and dare I say ideal theories?
And where this leads is not to axiomatizing ontologies in first order logic, but rather axiomatizing first order logic in ontologies. The former approach leaves us right where we are in pragmatism. Yeah we can express stuff, but that only gets us so far. How about we crack open some new ideas about what an ontology can do, like anticipate science? Pragmatism pays the bills. I respect that. But a little idealism can change the world. And aren’t we hungry for that?
Michael Denny
Greg,
Would the work since the early 2000’s building and exploiting the Gene Ontology provide an example of the potential utility of ontology for predictive purposes? One use of that ontology is as a tool to predict the likely functions of proteins based on their structure and other properties. An example of this use in conjunction with more recent machine learning based computational protein function prediction is addressed in the citation below.
Mike
--
All contributions to this forum are covered by an open-source license.
For information about the wiki, the license, and how to subscribe or
unsubscribe to the forum, see http://ontologforum.org/info
---
You received this message because you are subscribed to the Google Groups "ontolog-forum" group.
To unsubscribe from this group and stop receiving emails from it, send an email to ontolog-foru...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/ontolog-forum/a5349840-4537-43fb-ad16-1280756ccc66n%40googlegroups.com.
Alex Shkotin
Greg,
I think it's a kind of misunderstanding, as we are talking here about formal ontologies - a special kind of computer based artifact.
So it's better to read MG's sentences like this
1. formal ontology for a physical theory
2. formal ontology as a physical theory
Today, formal ontology (see for ex. OBO Foundry) is a sophisticated dictionary of terms, in our case from one or another physical theory.
What is important and unique, formal ontology keeps some knowledge formalized, i.e. in the form of formulas.
First of all we have physical theory and we create formal ontology for it.
Secondly, we begin to formalize knowledge (axioms, definitions) in formal ontology and this means that now we look at formal ontology as a formal physical theory.
It is hard to formalize physical theory. Even the Static part of Mechanics.
Formal ontology is just formalized theory.
Consider one time we open a Chemistry textbook and it's written on FOL, actually HOL.
There are some nuances.
For example OWL2 ontology keeps theory in TBox and can keep data in ABox, but usually we keep data in DB.
But to check data against a theory we import a part needed in ABox.
If data are accepted by the theory we call it a model of this theory, following A. Tarski.
If data corresponds to the part of reality it is also a model of this part if theory is good enough for this particular part.
I hope we get "a spark of idealism" on Track 2
https://ontologforum.com/index.php/OntologySummit2025#Track_2:_Theoretical_Knowledge_and_Reality.
Alex
--
Gary Berg-Cross
>How so? Because when the ontology treads into the unknown, the only recourse he presented was empirical. We go >back to the chemists and they tell us whether the molecule exists, doesn’t exist or couldn’t exist and then we place >some extension or constraint on the ontology to account for this
--
Alex Shkotin
To view this discussion visit https://groups.google.com/d/msgid/ontolog-forum/CAMhe4f2ALMj71Sw5vC5cis3RfXKXF1etRv9LKz6S2Rvb%2B8TECA%40mail.gmail.com.
Avril Styrman
- Methodology and progress of philosophy in general
- How can philosophy foster progress in science?
- How can AI (artificial intelligence) foster progress or accelerate research in science and philosophy?
- How can scientific methods foster progress in philosophy?
- Evaluation criteria for philosophical theories
- From pluralism to syntheses
- Philosophical theories as axiomatic systems
- Induction of ontological commitments
- Causal-mechanical explanations in philosophy
- Experimental philosophy
- Avril Styrman (chair), University of Helsinki
- Rögnvaldur Ingthorsson (vice-chair), University of Helsinki
- Paavo Pylkkänen, University of Helsinki
- Marja-Liisa Kakkuri-Knuuttila, Aalto University
- Anssi Korhonen, University of Helsinki
- Mirja-Leena Zgurskaya (conference secretary), University of Helsinki
- Eetu Sipilä (conference secretary), University of Helsinki
Gilles Kassel
===================
CALL FOR TUTORIALS
===================
15th International Conference on Formal Ontology in Information Systems
08-12 September 2025 (Catania, Italy)
Website: https://www.dmi.unict.it/fois2025/
================================
Submission deadline: 28 February 2025
Questions to: tutorials...@easychair.org
================================
The Tutorial Chairs for FOIS 2025 invite proposals for tutorials to be
held on-site in conjunction with FOIS 2025, during the week of the
main conference which will take place 08-12 September 2025 in Catania,
Italy.
FOIS 2025 seeks original and highly relevant proposals for tutorials
to complement the conference program. The tutorial program will
address the diverse interests of its audience: ontologies, formal
ontology, and knowledge management and their application in
information science or other areas. Moreover, tutorials can be an
excellent opportunity to cover basic or advanced topics related to
ontology research, ranging from cognitive science to knowledge
representation, natural language processing, artificial intelligence,
logic, philosophy, and linguistics as well as well-established or
emerging research areas, techniques, methodologies and perspectives
related to the ontology field.
In 2025, we are looking for individuals and teams of presenters who
would like to offer a tutorial specifically designed as a basic or an
advanced description of a topic, or as a showcase of relevant
techniques. We privilege diversity in the organizing team as well as
the presence of young researchers and PhD students. Moreover, up to 1
instructor per tutorial can have their registration fees waived for
the tutorials as well as for JOWO 2025, the Joint Ontology Workshops
(https://www.dmi.unict.it/fois2025/?page_id=422) which are co-located
with FOIS 2025.
======
Format
======
Tutorials are intended to be highly interactive, and educational, and
should actively engage the audience to participate in the discussion.
We accept short and long tutorials, which are expected to last 1h30
and 3h respectively.
===================
Submission instructions
===================
Proposals should be no more than 2 pages in length, and must contain
the following information:
- title;
- names and affiliations of the presenters;
- a brief CV of the presenters, including their experience in tutorial
organization; Please indicate a primary contact person.
- description of the topic (max 1 page);
- brief description of the expected audience. Please give an estimate
of the expected number of participants and their background and
interests; if applicable, please relate this to participation in
similar prior events;
- 200-word abstract highlighting why the tutorial is attractive to
FOIS 2025 participants – Why is it novel? What are key learning
outcomes?
Proposals should be submitted via EasyChair by 28 February 2025 as a
single PDF file. A notification regarding your submission will be
received on 14 March 2025.
===============
Evaluation criteria
===============
The decision on acceptance or rejection of tutorial proposals will be
made on the basis of the overall quality of the proposal and its
appeal to a reasonable fraction of the FOIS community.
In particular, tutorials should satisfy each of the following criteria:
- The research topic falls in the general scope of the conference.
- There is a clear focus on a specific technology, problem or application.
- There is an anticipation of a sufficiently large community
interested in the topic.
- There is a concrete plan for the tutorial format.
- The presenters have reasonable experience in providing engaging lectures.
- The material for the tutorial is made available to the FOIS 2025
community before the tutorial date.
=============
Important dates
=============
Submission deadline: 28 February 2025
Notification: 15 March 2025
Tutorials take place on-site in conjunction with FOIS 2025 in Catania,
Italy (08-09 September 2025)
================
Terms and conditions
================
The presenters/instructors of accepted tutorials are expected to:
- participate in the tutorial in person;
- have an active role in the tutorial;
- submit the material for attendees (slide sets, additional teaching
material, software installation, question answering, and usage guides
for practical hands-on sessions, e.g., Jupyter Notebooks.) to the
Tutorial Chairs and make it available on the FOIS 2025 website at
least one week prior to the date of the tutorial.
============
Tutorial chairs
============
Lucía Gómez Álvarez, University Grenoble Alpes, Inria, CNRS, Grenoble
INPG, France
Gilles Kassel, Jules Verne University of Picardie, France
Please direct all further questions to tutorials...@easychair.org
--
Professeur émérite à l'Université de Picardie Jules Verne
Chercheur au laboratoire MIS
33 rue Saint-Leu
80039 Amiens Cedex 1
https://philpeople.org/profiles/gilles-kassel
gregsharp73
Michael Denny
Greg, I am not trying to dismiss either machine learning or ontology as offering predictive power. I believe they both should be able to, but your allusion to explanation as a differentiator I would claim runs in the opposite direction you suggest when you say: ”I suspect many an ontologist has experienced that sinking dissolution of power once all has been described but nothing has been explained”. Computational ontology is readily able to trace the reasoning path that leads to its conclusions. Machine learning and neural networks are notoriously unable to do that at all.
To view this discussion visit https://groups.google.com/d/msgid/ontolog-forum/a05a3d9f-cfb1-4485-93a5-b84dcf6b8b07n%40googlegroups.com.
Michael Denny
FWD of reply:
John,
Post-training techniques to enhance LLM output, of course, can vary across models and applications in order to refine parameter weightings and augment prompting, but I do not see how that capability touches on any distinction between computational ontology and LLMs. DeepSeek apparently allows more or easier fine-tuning (post-training) than some other LLMs. Also, I do believe there are ways during inference (post-training) in which LLMs can apply some forms of logical or probabilistic reasoning to adjust the context, but none of that is retained or ever affects the underlying model. What would be nice is if generative AI could reliably induct semantically coherent logical axioms and rules from the global corpus that could be processed under a panel of inference rules within ontology or other symbolic AI systems to build theories and classify propositions (extract knowledge) from a subset or specialized corpus. Then things would become explainable and validatable.
Mike
From: John Bottoms [mailto:jo...@firststarsystems.com]
Sent: Saturday, February 8, 2025 1:38 PM
To: Michael Denny
Subject: Re: [ontolog-forum] Why Michael's talk disappointed me.
Mike,
The difference between ChatGPT and DeepSeek appears to be in the training architecture. ChatGPT, as I read it, use human in the loop for critique, while DeepSeek is using digital agents to do the checking. There are design elements that have not been revealed in both cases but the MIT Review article sounds reasonable.
"How DeepSeek ripped up the AI playbook-and why everyone's going to follow lead" -MIT TECHNOLOGY REVIEW
John Bottoms
* * * * * *
To view this discussion visit https://groups.google.com/d/msgid/ontolog-forum/000f01db799b%2481b89110%248529b330%24%40com.
John Bottoms
Mike,w
I didn't word my view clearly enough. The brute fact is that the companies that participate in the $00B project are going to define how the ontologies will be constructed and used. Most of the time when writing specifications and standard there are fields that attempt to meet the needs of users with parameterized approaches.
My views on DeepSeek and related architectures is that they have showed specific efficiencies that drive their architectures. It is fortunate that China has shared the basic approaches. In my view the work of ontology science will be driven by the economics. And we should look to what roles this group can provide in facilitating identifying and covering the needs of all participants.
-John Bottoms
To view this discussion visit https://groups.google.com/d/msgid/ontolog-forum/001801db7a84%24b20b2c30%2416218490%24%40com.
gregsharp73
Greg
Michael Denny
Greg,
True. I expect (or hope) such dissolution may be curbed by building ontologies to exploit the inference power their logic language offers in order to attack well defined problems rather than simply serve as catalog systems or knowledge repositories.
To view this discussion visit https://groups.google.com/d/msgid/ontolog-forum/1772ce10-c8a5-426e-b16d-d6576311989dn%40googlegroups.com.
Alexandre Rademaker
Have you ever try to explain a nontrivial long tableaux proof? Back in my PhD, the motivation I had to implement natural deduction and sequent calculus deduction systems for some DL fragments were precisely the lack of explainability in the refutation-based tableaux provers.
As far as I know. The SOTA OWL provers are still tableaux-based, right?
Sure, ontologies are not necessarily in DL/OWL…
Best,
—-
Alexandre Rademaker
http://arademaker.github.io
Alex Shkotin
--
All contributions to this forum are covered by an open-source license.
For information about the wiki, the license, and how to subscribe or
unsubscribe to the forum, see http://ontologforum.org/info
---
You received this message because you are subscribed to the Google Groups "ontolog-forum" group.
To unsubscribe from this group and stop receiving emails from it, send an email to ontolog-foru...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/ontolog-forum/AF276A17-5879-4D42-899F-1950202D64BF%40gmail.com.