Richard,
I was distinguishing between two different attitudes that people take to
the problem of making a definition. One attitude (the one you adopt
here, and the one I would also wholeheartedly adopt) is to look for a
useful *descriptive* definition: something that takes the commonsense
definition and sharpens it in a way that makes it useful for guiding
research. It is perfectly okay for descriptive definitions to be
circular, so long as they are not egregiously so. In that respect,
using the word "goal" in the definition is not really a problem.
But there is a second type of definition that tries to *formalize* what
the subject is, and that is where my challenge was really directed.
So the issue is not whether intelligence can be defined, but whether
it can be formally define.
A formal definition can also be "descriptive", in the common sense of
the term. I guess what you preferred as an informal definition, or a
definition in a natural language.
In a formal definition we try to capture the thing in such a way that,
from that point on, the process of working with the theoretical idea
becomes much more automated, rigorous and equation-bound. This is what
Newton did to concepts like "inertia" and "gravity", both of which were
quite vague and subjective before he came along, but which were simply
numbers in equations after he had done his formalization.
Yes. formalization is an attempt to make the description more clear
and certain. Other things being equal, it is indeed desired. However,
when a formalization actually makes the idea even less clear (this
kind of thing does happen), or the clearness is achieved under the
cost of the other desired properties of a concept (e.g., faithfulness
and fruitfulness), then it is wrong.
But now here is my problem: in the case of this formalization of the
commonsense term "learning", the original meaning has been distorted so
as to make it possible to get a mathematical handle on it. By trying to
make the definition formal, we have actually corrupted and reduced the
original meaning. This kind of formal definition has an agenda behind
it -- the goal is to make "learning" mean something rigorous because the
people doing the formalization want to be able to study "learning" using
the tools of mathematics.
I agree with you here. I also think the "machine learning" community
has make the concept "learning" too poor when formalizing it.
If you just say that "a formal treatment of a subject is not
necessarily better than an informal treatment of the same subject",
then I fully agree with you. Indeed, in mainstream AI people tend to
give too much credit to the formalness of a work, compared to the
other desired features.
However, I hope people won't go to the other extreme to see
formalization as harmful or undesired. Just like everything else,
there are "good formalization" and "bad formalization", as well as
"necessary formalization" and "excessive formalization".
Lastly: you may say to me that everyone should "formalize" their
definitions at some point or they will just be writing descriptions
forever. Very true. Hutter formalizes at step one. You formalize much
later.
Yes. Also it is a matter of degree. Since AI is not math, nobody can
fully formalize their definitions. As far as formalization increase
the clearness of a theory without hurting it much in the other
aspects, it is a good idea.
My strategy is to be extremely circumspect about any kind of
formalization that closes off options, so my initial goal is not a
theory as such (in the sense that NARS is a theory), but to produce a
"framework" within which theories can be couched.
I see your point. However, I'm sure you know that such a framework
cannot be really "theory-neutral". Instead, it more likely to be
within the freedom given by a certain theory.
For example, in NARS there are many system parameters to be
determined, and a testing platform/framework will surely make the job
much efficient. On the other hand, I don't think it is like to build a
framework that allow all kinds of system to be couched together.
You need to convince me (later) with more details. For example, I and
Ben have been tried for years to find concrete ways to compare NARS
and Novamente, and we haven't found them yet.
Thanks for the explanation --- the issue is more clear to me now, and
no formalization is needed here. ;-)
Pei
-----
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?member_id=231415&user_secret=fabd7936
