Date: Mon, 8 May 2006 14:22:25 -0400
To: "depei bao" <baodepei@gmail.com>
From: Marvin Minsky <minsky@media.mit.edu>
Subject: Re: About common sense computing-some suspicions
Cc: lieber@media.mit.edu, Ian Eslick <eslick@ai.mit.edu>,
Dustin Smith <dustin@media.mit.edu>
　

Hello,Prof Minsky:

　

I am a master candidate in CS department of Tsinghua University, China. After reading your paper "An Architecture for Cognitive Diversity", I am attracted by your architecture of common sense computing. It is more flexible than most current AI systems. However, I think your architecture may still have some problems.

What I see "human-level intelligence" lies in its flexibility. The difference between a person and an "automatic device" is how they can adapt to new environment flexibly. Likewise, the difference of intelligence between people has the same meaning. We say someone is "stupid" because he may follows some rigid rules without adapting himself to new situations. The traditional way of evaluating intelligence by rationality may conceal the real meaning of intelligence. Hence, I think the degree of flexibility should be regarded as the criteria of evaluating intelligence and also, as the most important principle of designing a human-level-intelligence machine.


　
Take a look at the chapters on my home page, in the entry

　
     The Emotion Machine (draft) ( 1 2 3 4 5 6 7 8 9 )

　
I used the word "resourcefulness" for what you call flexibility.  Anyway, see if you can think of even more ways to make a flexible machine.  I agree that logic-representations are too limited, and we need more versatile representations, such as those in Chapter 8.   And I'm sure that we will soon invent more and better representations.

　
I do not understand what you mean when you say, "Unlike semantic network, all human concepts should be represented by node and down to percepts."  You seem to assume that those networks are separate--but the nodes of each semantic network are actually links to other semantic networks, and these may reach down to other levels.   But they also can reach up to higher-level ones.

　
Also, I think you put too much emphasis on 'percepts.'   As infants, we use these to learn and develop, but  I suggest in http://web.media.mit.edu/~minsky/papers/Internal%20Grounding.html that we also have other ways to learn--ways that are based on experiments and observations inside the brain.

　
Finally, most people are not very flexible -- perhaps because they think too hierarchically -- so we should have even larger-scale goals than to imitate human thinking.  So our 'layers' must not be too clearly defined.  The system needs room to move, not only up and down, but sideways (for example, by using difference-networks and panalogies).

　
In summary, I do agree that 'rationality' is a false goal, and flexibility is a much better direction.  I very much like your ambitious thinking!

　
-- Marvin

Date: Fri, 8 Dec 2006 16:18:24
To: "depei bao" <baodepei@gmail.com>
From: Marvin Minsky <minsky@media.mit.edu>
Subject: Re: Some confusions about common sense systems
　

Dear, Prof. Minsky:

Long time no see. How are you? I am considering building a common sense system. But some problems confuse me. When we solve some  problems, we usually want to build a generalized model to handle most of the situations.

I would suggest, instead, that we construct *simplified* models -- and hope that they are general enough.

    

For example, we modify the raw model to include some exceptions in the problem domain. If it can be understood as a process of searching in a certain knowledge network. How do people terminate before they can verify that the model is really the goal (contain all the instances)? (actually, we believe we have achieved the most general model by "intuition" before we can verify it and in many cases, it is really the goal!).

Again, I think you are using "verify" too much. Try using the concept of "Difference Engine" instead; then the process stops when you don't recognize any more differences to be important.

　

Then, what is the internal mechanism of this "intuition" (it is really hard to grab, to what extent our inner mechanism stops the search?). Could you provide some suggestions and comments?

　

Thank you very much

Best regards

Bao Depei
from Tsinghua University

Dear Prof.Doyle:
I am a Chinese student from Tsinghua university. I have been admitted to NCSU with full support this year.
The major reason I go to the US for my Phd study is that I am very interested in realizing general-purpose AI, or human-level intelligence. I found your research attracts me a lot. I hope you can help me on my way to the goal.
In my opinion, there are two major trends in general AI researches. One is led by Dr.MacCarthy who claims that AI can be realized by formalism, specifically by logic. The other trend follows the opinion by Prof. Minsky who believes that there is no single model for AI and we need to combine various representation models.
I think, both approaches have their limitations. Human mind can not be described just by logic. There are many other "ways of thinking" in our mind. For example, I think there are (at least) two parts of our mind: consciousness and unconsciousness. For consciousness, it is perhaps governed mainly by the logical pattern of thinking (I also have doubt about this because, in fact, "logic" is just a concept in our mind), but for unconscious part (more specifically, the "underlying" mechanism of our brain which is not controlled by deliberation and reflection), I suspect that it is organized like in a logical system. In a word, our mind has very complex structure which can not be explained just by logic.
The method of Prof. Minsky which sees AI from cognition science, behaviorism and psychology is more realistic, I think. But his method (like in his book "the emotional machine") is too specific. For example, he will combine all kinds of representation models, statistics model, ANN, logic, etc together and have an agent select the model at appropriate time. I think, however, are there several general models in our mind which can describe all particular representation models? For example, declarative knowledge representation that we represent "concepts" by symbols seems the only way we adopt to solve the AI problem until now, but I prefer to dig into what is the underlying mechanism of our mind that bring us such kind of "symbolic pattern" of thinking.
If we can find several general models to explain all ways of thinking, and then combining them with motivations, the general AI problem is solved.

PS: I think that neural science has its usefulness to verify our theory. But it is not so helpful in finding the theory. I think our mind is produced due to some special structures of our brain. it is hard to find such kind of structure by measuring some physical properties of our brain. The only way to find it is from human behavior. It is like the gravity, without the theory of Newton's law which was obtained from observation of apple's "behavior", We can not hope to find "gravity" by measuring the property of the apple.

Could you have some comments on my ideas?

Thank you very much

Bao Depei

----- Original Message -----
From: "Jon Doyle"
To: "depei bao"
Cc:
Sent: Wednesday, May 02, 2007 3:51 AM
Subject: Re: A student for the academic help


Dear Bao Depei,
>
Thank you for your note, and congratulations on your admission to our
program.
>
You are correct in thinking that I am interested in general
human-level intelligence, and I certainly would be happy to talk with
you about this topic and some of the other issues you mention.
>
Although there were arguments between proponents of logic-based
deduction and other conceptions of thinking back in the 1970s, I think
it is fair to say today there are very few people who regard logical
deduction as a plausible method for making intelligent agents. Almost
everyone will agree that intelligence involves multiple sorts of
methods, and even multiple competing modules each of which uses the
same method.
>
That said, logic remains a very important tool for specification,
analysis, and expression of systems. The character of the mental
operations and representations being specified, analyzed, or expressed
does not depend on the properties of logic itself. As Minsky points
out, the structure of the mind is a lot more complicated than the
simple notion of logical representation and inference.
>
I might be misunderstanding what you have written, but I do not think
the general intelligence problem is solved just by finding a complete
set of general mechanisms. Intelligence depends on both knowledge and
skill (or inferential and learning capabilities). If by general
models you mean mental skills, those do not in themselves have the
knowledge (or at least all the knowledge) needed to exhibit
intelligence. Finding that knowledge, or finding convincing reasons
to think that the ignorant mind can learn it all from some
environment, is needed before one gets intelligence. Fortunately for
you and other graduate students, it looks like there is plenty of work
to be done to identify the needed mental skills, base knowledge, and
learning environments.
>
I agree completely that knowledge of neurons says very little about
the nature of thinking and the mind. We have virtually no means for
determining the properties of any common material (granite, oak wood,
Earth's atmosphere) from the properties of the subatomic particles
from which these are composed. And even were we to be able to derive
the properties of paper and ink from these, we would have no
information about the nature of the poem written in the ink on the
paper.
>
Sincerely yours,
>
Jon Doyle
>
=========================================================================
Jon Doyle http://www.csc.ncsu.edu/faculty/doyle
Department of Computer Science Jon_Doyle@ncsu.edu
North Carolina State University Voice: 919-513-0423
Box 8206, Raleigh, NC 27695-8206 Fax: 919-515-7896
=========================================================================

----- Original Message -----
From: "Jon Doyle"
To: "depei bao"
Cc:
Sent: Thursday, May 03, 2007 12:32 AM
Subject: Re: Re: A student for the academic help

Dear Prof.Dolye:
Thank you very much for your reply. I am very pleased to hear that you are focusing on realizing human-level Intelligence.
I completely agree that our mind consists of both knowledge representation and metal skills. In my word, knowledge is many concepts organized together and mental skills are our "ways of thinking".
But that is not my point (Because it is my first email to initiate the topic, so I don't dig into detail. I think that is where the misunderstanding comes from). I think that the methods taken by Prof. Minsky is on the right road to the goal. But in my opinion, his method of combining many representation models might be too specific. So it is very hard, if possible, to realize. Too many models should be realized explicitly in the system. There should be some general models both for knowledge representation and mental skills which can derive all kinds of specific models for our daily thinking (after learning, maybe some of these derived specific models are represented explicitly in our brain, but they are "software" of our mind, not hard-wired mechanism of our brain).
As for our concept system (our knowledge), I think declarative knowledge representation is doomed to be too rigid. How could we represent the "meaning" of a concept with explicit boundary? The "meaning" should be embraced in some models. For example, there should be a general model for concrete objects of the world, which specifies how to distinguish the object (shape, color, etc). (in fact, in our brain, those are perhaps several mechanisms which have some cognitive roots). For a specific concept like "tree", it includes a label "tree" and a group of the general models (as a concrete object, as an abstract object and so on). The "meaning" of the tree is embraced in a group of several such general models (different people may have different groups of general models for the same concept). Another example is a little bit abstract concept like "time". To find a general model of "time" is to answer why human have the concept of "time", what is the cognitive basis of "time". After all those primitive concepts have been constructed, the upper level (if knowledge is organized as a hierarchical structure, however, I think it should be built as a more complex network) can be built on primitive concepts in the same way. All our concepts and their meanings may be organized in this way. Furthermore, the primitive concepts like the concrete object and time are direct reflection to these general models.
As for mental skills, it is the same. For example, logical and probabilistic ways of thinking can be regarded as two general models. To find the general models are to answer why human can think logically and what is the cognitive basis of probabilistic thinking. Then upper level "ways of thinking" is the combination of lower level primitive models (that's why different people can have different ways of thinking). Also, we need a model for the combination which also has individual differences.
Of course, all of above statements are just part of functions of our mind. We need answers to what is the goal of human beings, what is the motivation of our activities, how we learn from external world and many more. But most of those mechanisms have two parts, one is hard-wired function of our brain, another is learned.

sincerely

Depei