[Fis] Paradigm shift: from "I know" to "I don't know"!

[钟义信]

Dear Krassimir,


What you proposed is an interesting point of view, related to paradigm shift. This is, of course, worth of discussing.


What I would like to remind is that the term of  "paradigm" in Kuhn can be referred to many things like pattern, mode, model, case, exemplar, etc. However, in <Paradigm Change in AI>, the term of "paradigm" can only be referred to the "scientific worldview and its methodology". This is because of the fact that only the <scientific worldview and its methodology> can be considered as the criterion of whether a scientific discipline needs a revolution whereas the others (like pattern, model, ...) cannot.


Comments are welcome.




My best wishes,





 Prof. Yixin ZHONG
AI School, BUPT
Beijing 100876, China
 

 
 
 
------------------ Original ------------------
From:  "Krassimir Markov"<itheaiss at gmail.com>;
Date:  Mon, Dec 4, 2023 08:52 PM
To:  "fis"<fis at listas.unizar.es>; 

Subject:  [Fis] Paradigm shift: from "I know" to "I don't know"!

 

Dear Yixin, Eric, Joseph, and FIS Colleagues,
The conference in London is over and now I have little time to meet Joseph's expectations.
Let's remember Shannon's Theory. 
Information is one bit if the answer to the question is "Yes", 
If the answer is "No" then entropy is one bit. 
In other words, we count the ones and zeros in the answers. 


At the same time, no question is allowed without an answer!

At the level of signals, this means that a signal with modulation "1" or with modulation "0" must be received. If no signal is received with one of these modulations, it is assumed that an error has occurred and the signal should be retransmitted. 
If this is impossible, the system is helpless.


In conclusion, the motto that emerges is that "I must know" to be able to act.


This is the motto of Artificial Intelligence at the moment. 
The system must be trained, it must know everything in advance to operate. 
Even Large Language Models are based on prior learning and respond only based on what has already been learned. 
This is a feature of existing data structures that are used in artificial intelligence systems, for example - neural networks. 
For each new knowledge, the neural network must be trained again.

At the same time, we live in conditions of incomplete data. 
I have to create models for what "I don't know".


Going back to Shannon's Theory, this means allowing for unanswered questions. 
In other words, it is no longer possible to count and evaluate the answers, but to count and evaluate the questions. 
And the quantity of information and entropy to be calculated based on the number of solved and, accordingly, unsolved problems. 


Now the motto is "I don't know, and I have to create an answer myself!". 


This is the next level of artificial intelligence, where the system should determine by itself what it "doesn't know" and generate new knowledge.

In other words, a paradigm shift from "I know" to "I don't know!" is required.


Finally, let us remember antiquity and the words of the great philosopher: "I know that I know nothing!"


With friendly greetings,
Krassimir

PS: In the next letter I will comment on the relationship between the concepts of "matter", "energy" and "information".
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://listas.unizar.es/pipermail/fis/attachments/20231206/ccf90fc6/attachment.html>