December 6th, 2019

2019

lytdybr

Завтра веду "Машинный интеллект" (https://system-school.ru/machine), опять переделывал слайды -- там ведь продолжаются по два прорыва в неделю. Мои мысли в последнюю неделю довольно плотно занимает текст François Chollet "On the Measure of Intelligence", last revised 25 Nov 2019, https://arxiv.org/abs/1911.01547. Текст задаёт какой-то набор понятий (framework) по сравнению AI и человеческого интеллекта, при этом позволяет начинать осторожно говорить об AGI без немедленного прихода любителей порассуждать о SkyNet, "роботы заменят людей" и прочих робо-апокалипсисах. Мы продолжаем говорить про интеллект в терминах решения задач, не переходя на личности (pun intended), то есть не втягивая в разговор "сознание", "свободу воли" и прочие сомнительные конструкты. И это не единственная работа, из-за которой пришлось менять слайды.

John Sowa жжёт в ответ на реплику Jack Ring в Ontolog Forum (https://groups.google.com/d/msgid/ontology-summit/910a0b7fca29e91e0401a15c43659778.squirrel%40webmail2.bestweb.net):
Jack R> The only way to resolve different perspectives is through dialog.

I certainly agree. For interoperability among independently developed and developing systems, there must be a dialog among all the people and groups of people who need to collaborate and cooperate.

That implies several points: (1) The best language(s) for carrying on a dialog are our familiar natural languages. (2) The main reason why NLs are so useful is that their words and patterns of words are flexible -- they can shift their meanings and patterns of meanings to adapt to any imaginable situations. (3) After various parties to the dialog have reached agreement, the informal NL can be formalized in a precisely defined language (natural or artificial) in which the details are specified to whatever degree of precision is appropriate for interoperation.

Note that precision is only achieved *after* a lengthy process of design, analysis, discussion, refinement, negotiation, and agreement. The degree of precision is determined by the nature of the applications and the problems that must be solved during the design and development process.

This is another reason why a precise, rigidly specified top-level ontology is a terrible starting point for independently developed, but interoperating systems. But it also explains why precisely defined mid-level ontologies (AKA microtheories) are valuable for the various components within a single project or family of projects. The precise ontology determines a silo for a single purpose. But interoperability among silos requires a looser overall framework.
John

А вот ещё про динамические онтологии:

In fhe Ontology Summit telecon last week, there was some talk about the dynamic nature of knowledge graphs. It implies that the subject matter described by typical KGs is constantly changing. That makes it impossible to have a fixed schema -- no fixed, completely specified formal ontology.

I made the point that natural languages have a similar nature. In a typical dialog or debate, people have different points of view. They sometimes agree, sometimes disagree. They ask questions. They may revise, clarify, or change their opinions. Sometimes there may be more points of view than there are participants in the discussion. Whether or not they agree on a final outcome, the senses of many words and phrases may change during the discussion.

This afternoon, I received an anouncement of a new journal that will be devoted to this topic: "Interactional Linguistics". Its goal is to analyze evidence of "language as locally contingent, temporal, and ever-adaptive."

The reason why language changes dynamically is that it expresses the dynamic interactions among the people who use it.

But some uses of language are static, and they can be translated to some version of formal logic. For example, consider the owner's manual for your car. That is a static description of a finished product. It's possible to specify a fixed ontology for all the parts, their interconnections, and their functions during the operation of the car. But cars from different manufacturers or from the same manufacturer in different years will require ontologies with different definitions for many of the same terms.

In fact, automobile designs undergo a major revolution every 20 years. Just compare the horseless carriages of 1900 and the Model T Fords of the 1920s. In the 1980s, no cars had computers. But by 2000, they all had multiple computers. Today, you can see Teslas going down the highway with the so-called driver sound asleep. Just imagine the cars of the 2040s. Any ontology for cars would require a major revision every 5 years.
John

Там дальше было описание журнала интерактивной лингвистики, но оно не так интересно. Потом Джон в ответе на реплику Matthew West написал про dynamic theory of ontology:

JFS> the dynamic nature of knowledge graphs... implies that the subject matter described by typical KGs is constantly changing. That makes it impossible to have a fixed schema -- no fixed, completely specified formal ontology.

MW> No it doesn't. It just means your schema needs to take change into account and be extensible.

We agree in principle, but the time scale makes an enormous difference in practice. If the time scale is measured in years or months (as in the example about cars), it's possible to update the ontology on a rgular basis. But when the time scale is measured in seconds (as in a discussion among two or more people), the practical problems require a major change in strategy.

Anybody who took a course in philosophy probably read a couple of dialogs by Plato. In each one, Socrates asks leading questions that cause the other participants to make major changes in their ontology from one sentence to the next. For another example, just listen to any telecon for the Ontology Summit.

For more about interactional linguistics, see the Wikipedia article and the references cited there: https://en.wikipedia.org/wiki/Interactional_linguistics

That article distinguishes written languge (the source data for most theories of language) from rapidly changing perspectives in any dialog among two or more people. But similar changes occur from one article to the next on every subject -- ranging from scientific publications to news stories to political commentary.

Articles on the same subject by different authors or by the same author in response to the others will have different senses for the same words and phrases. Any task that requires an analysis and comparison of multiple inputs from different sources must accommodate changes in the ontology for each text (including each text message or tweet).

These comments don't mean that formal ontologies are impossible, but they must accommodate an open-ended system of microtheories. And the choice of which microtheory to use at any moment must take the context into account. See "A dynamic theory of ontology", http://jfsowa.com/pubs/dynonto.pdf .
John
Поскольку у нас в курсе онтологики много про коммуникацию, нам нужно как-то особо выпятить моменты изменения схемы в ходе коммуникации, как это на это специально обратить внимание. Так сказать, "непрерывная схематизация, непрерывный рендеринг" — чтобы не было ощущения одноактности этой процедуры (статичная неизменяемая с трудном полученная схема и потом долгая коммуникация по её поводу, не меняющая эту схему никак, только передающая её содержание для других ролей).

Одна из студенток ВШЭ начала в качестве курсовой делать приложение, которое будет сочетать две функции: давать читать учебник "Системное мышление 2019" и решать (с проверкой) к нему задачи, но ещё и напоминать, что нужно таки читать, и нужно решать -- что-то типа приложения Keep Yoga, только это будет Keep Systems Thinking Learning. Интересно, что из этого проекта получится, но процесс пошёл.

С техпредами в МФТИ всё пока плохо: оценки по экзамену поставлены, из тридцати человек половина получила "неуд". Но есть и несколько "отлично", что хоть как-то утешает. Ещё не встречал группы с таким отношением к предмету. Например, были письма в личку с предложением мне "пойти на компромисс" (от автора одной из самых плохих работ). Что у этих людей в мозгах?! Повторюсь: когда говорят, что "жизнь их научит", я и есть та самая "жизнь, которая учит". Но мои возможности тут скромны, им для вправления мозгов нужны какие-то более сильнодействующие варианты учащей жизни.

Опубликовано видео одного урока из курса Церена Церенова по системному мышлению для школьников -- https://www.facebook.com/tseren.tserenov/posts/2443930382371570 (само видео -- https://www.youtube.com/watch?v=lMcHvw5Q2mU).

UPDATE: обсуждение в фейсбуке -- https://www.facebook.com/ailevenchuk/posts/10216944199931289