Brad Paulsen wrote:
Richard Loosemore wrote:
Brad Paulsen wrote:
Richard Loosemore wrote:
Brad Paulsen wrote:
Richard Loosemore wrote:
Brad Paulsen wrote:
All,
Here's a question for you:
What does fomlepung mean?
If your immediate (mental) response was "I don't know." it means
you're not a slang-slinging Norwegian. But, how did your brain
produce that "feeling of not knowing"? And, how did it produce
that feeling so fast?
Your brain may have been able to do a massively-parallel search
of your entire memory and come up "empty." But, if it does this,
it's subconscious. No one to whom I've presented the above
question has reported a conscious "feeling of searching" before
having the conscious feeling of not knowing.
It could be that your brain keeps a "list of things I don't
know." I tend to think this is the case, but it doesn't explain
why your brain can react so quickly with the feeling of not
knowing when it doesn't know it doesn't know (e.g., the very
first time it encounters the word "fomlepung").
My intuition tells me the feeling of not knowing when presented
with a completely novel concept or event is a product of the
"Danger, Will Robinson!", reptilian part of our brain. When we
don't know we don't know something we react with a feeling of not
knowing as a survival response. Then, having survived, we put
the thing not known at the head of our list of "things I don't
know." As long as that thing is in this list it explains how we
can come to the feeling of not knowing it so quickly.
Of course, keeping a large list of "things I don't know" around
is probably not a good idea. I suspect such a list will
naturally get smaller through atrophy. You will probably never
encounter the fomlepung question again, so the fact that you
don't know what it means will become less and less important and
eventually it will drop off the end of the list. And...
Another intuition tells me that the list of "things I don't
know", might generate a certain amount of cognitive dissonance
the resolution of which can only be accomplished by seeking out
new information (i.e., "learning")? If so, does this mean that
such a list in an AGI could be an important element of that AGI's
"desire" to learn? From a functional point of view, this could
be something as simple as a scheduled background task that checks
the "things I don't know" list occasionally and, under the right
circumstances, "pings" the AGI with a pang of cognitive
dissonance from time to time.
So, what say ye?
Isn't this a bit of a no-brainer? Why would the human brain need
to keep lists of things it did not know, when it can simply break
the word down into components, then have mechanisms that watch for
the rate at which candidate lexical items become activated ....
when this mechanism notices that the rate of activation is well
below the usual threshold, it is a fairly simple thing for it to
announce that the item is not known.
Keeping lists of "things not known" is wildly, outrageously
impossible, for any system! Would we really expect that the word
"ikrwfheuigjsjboweonwjebgowinwkjbcewijcniwecwoicmuwbpiwjdncwjkdncowk-
owejwenowuycgxnjwiiweudnpwieudnwheudxiweidhuxehwuixwefgyjsdhxeiowudx-
hwieuhyxweipudxhnweduiweodiuweydnxiweudhcnhweduweiducyenwhuwiepixuwe-
dpiuwezpiweudnzpwieumzweuipweiuzmwepoidumw" is represented
somewhere as a "word that I do not know"? :-)
I note that even in the simplest word-recognition neural nets that
I built and studied in the 1990s, activation of a nonword
proceeded in a very different way than activation of a word: it
would have been easy to build something to trigger a "this is a
nonword" neuron.
Is there some type of AI formalism where nonword recognition would
be problematic?
Richard Loosemore
Richard,
You seem to have decided my request for comment was about word
(mis)recognition. It wasn't. Unfortunately, I included a
misleading example in my initial post. A couple of list members
called me on it immediately (I'd expect nothing less from this
group -- and this was a valid criticism duly noted). So far, three
people have pointed out that a query containing an un-common
(foreign, slang or both) word is one way to quickly generate the
"feeling of not knowing." But, it is just that: only one way. Not
all "feelings of not knowing" are produced by linguistic analysis
of surface features. In fact, I would guess that the vast majority
of them are not so generated. Still, some are and pointing this
out was a valid contribution (perhaps that example was fortunately
bad).
I don't think my query is a no-brainer to answer (unless you want
to make it one) and your response, since it contained only another
"flavor" of the previous two responses, gives me no reason
whatsoever to change my opinion.
Please take a look at the revised example in this thread. I don't
think it has the same problems (as an example) as did the initial
example. In particular, all of the words are common (American
English) and the syntax is valid.
Well, no, I did understand that your point was a general one (I just
focused on the example because it was there, if you see what I mean).
But the same response applies exactly to any other version of the
"absence of a feeling of knowing" question. Thus: the system tries
to assemble a set of elements (call them symbols, nodes, or
whatever) that together form a consistent interpretation of the
input, but the progress of the assembly operation is monitored, and
it is quite easy to tell if the assembly of a consistent
interpretation is going well or not. When the monitor detects that
there are no significant elements becoming activated in a strong
way, it can reliably say that this input is something that the
system does not know.
So that would apply to a question like "Who won the 1945 World
Series" as much as it would to the appearance of a simple non-word.
In principle, there is nothing terribly difficult about such a
mechanism.
Richard Loosemore
The question was about the "feeling of not knowing" not the "absence
of a feeling of knowing." These could be quite different feelings.
I prefer to stick to the more positive version commonly evinced by
the statement "I don't know." Characterizing this as "absence of a
feeling of knowing" implies there is a "normal feeling of knowing"
typically present that is felt as "absent" in the presence of a query
such as the "World Series" example query. I would not be prepared to
argue that position, but I can argue there is a "feeling of not
knowing" since an "I don't know" response is a definitive declaration
of that mental state.
Your general description of "monitoring" during "the progress of the
assembly" contains implicit within it some form of search. There is
nothing in the query statement, itself, that would engender the
"feeling of not knowing" (as would a phonological surface feature
anomaly) without performing a semantic (meaning) analysis. Such an
analysis will require a search (e.g., of concepts). It, therefore,
falls under one of the mechanisms posited in my initial post.
If you think the analysis in the last paragraph is incorrect, I would
appreciate it if you could provide a concrete example of monitoring
the process of the assembly.
Ah, but I am trying to suggest, quite deliberately, that a "feeling of
not knowing" something can be exactly explained by a mechanism that
tracks the progress in a recognition episode, and does a time out. So
this is not the absence of knowing, this is a mechanism that knows
that it does not know.
Now, you refer to something that you label "an absence of a feeling of
knowing", and you suggest that this might be different from a feeling
of not knowing" ..... but I honestly do not think that the first thing
is well defined. It sounds like something not happening, whereas what
I am suggesting is that there is a monitoring mechanism that, at a
specific time, kicks in and says "I declare that, since nothing has
happened in this recognition episode, we KNOW positively that this
piece of information is not in the system". This would indeed, as you
put it, be a definitive declaration of that mental state.
No. You brought up the "absence of a feeling of knowing." I agree it's
not well defined, which is why I brought it up in my reply. I
understand your point about a "time out." That would, roughly,
correspond to the point in search (i.e., recognition) where the search
mechanism feels justified in declaring it has no knowledge about the
thing it's trying to recognize. The criteria will probably depend on
the type of search mechanism.
(To be precise, though, I should say that the system *believes* that
it does not know: there is a whole research literature regarding
people thinking that they do not know something, where in fact they do
know it, but often implicitly. But let's not go in that direction).
Actually, I believe all knowledge is belief. It's the distance one must
jump from one's conclusions to one's premises that makes an knowledge
feel less like belief and more like "hard fact." Either way, the result
would be more or less the same (although false "not knowing" would
probably resolve itself quicker).
As for the role of "search" in this: yes, you could construe it as a
form of search. However, there are many circumstances in parallel
constraint satisfaction (e.g. neural nets) where it is stretching a
point to call it "search". However, I do not really want to dispute
that point either: I think it is a matter of convention whether a
neural net activation of a large number of microfeatures, for example,
should be called search or not.
What I did disagree with was the idea that there might be a search for
(in the case of your first, lexical example) a specific nonword, which
then was eventually *successful* because it found a lexical item
corresponding to the target nonword, together with a label attached to
it saying "not known". I can make sense of a search (massively
parallel or otherwise) that ends in a timeout, which then causes
something to say "Because we did not find a match within the timeout
period, the item or fact is not known", but I cannot make sense of a
mechanism that must find the target, together with a label attached to
it saying that the item is not known. Searches are not a problem,
only searches of that peculiar sort.
Sorry. That first example was not well thought out. In my defense, I
speak Norwegian so I missed the phonological surface feature "problem"
with that example. I still shouldn't have missed it, but I did.
What I wanted to emphasize was the fact that if you are happy that the
feeling of knowing could be explained by a search-plus-timeout
mechanism (or, as I would prefer to phrase it, a
constraint-satisfaction-plus- timeout mechanism), then your initial
question was quite easy to answer, and not so very mysterious.
Actually, one of the benefits of having a list of things not known comes
from being able to search it for a positive result without having to
exhaustively (however one defines that) search the (presumably much
larger) list of things known. Again, I'm using "list" here generically
-- I'm not concerned about the memory storage mechanism (not here, at
least).
In the case of the semantic example (trying to answer the question
"Who won the World Series in 1954?") the mechanism would look the same
as in the lexical example, provided that you used constraint
satisfaction to get there. Initially, the words would activate
associated concepts (baseball, United States, sports, historical
events, people's names, team names, city names ...), and it would also
activated concepts that that captured self-knowledge, such as the fact
that (in my case) I know almost nothing of baseball, being that I am
English. All of those concepts would gel into a framework
representing the question itself and also the knowledge that I have
about the context of the question, and out of that interacting melange
of concepts would emerge a stable state, among which be the
meta-knowledge about what happened as a result of allowing all this to
percolate through the system. And one part of all that would be the
knowledge that the network of concepts had settled into a steady state
after the question, but the "slot" for "name of team that won in 1954"
would have nothing in it. It is the combination of the steady state
and the lack of a filler in the relevant slot that would lead the
system to conclude that it did not know the answer.
That sounds like a *lot* of work. And it sounds like it would have to
be repeated every time you ran into that query (which may be never
again, but humor me for a sec...). If at the end of the process the
first time (i.e., when you didn't know you didn't know the answer), you
put that query (or the meta-knowledge about what happened) on the things
I don't know list, wouldn't that act to keep the subsequent percolation
time to a minimum? And, couldn't it also act as an on-going
subconscious means to activate the desire for conscious learning?
Well, if one is using a constraint satisfaction system, this is not a
lot of work, but quite natural. However it might take a lot of work for
me to elaborate on that point.
As for your comment about the second time you encounter the question,
yes, absolutely: everything is changed at that point. The knowledge of
the episode is going to play a role, and the conclusion happens quicker.
This is consistent with psychological evidence also, because people
answer questions more quickly if they have done the same recently.
Now, sometimes the lack of knowledge comes quickly, and sometimes it
comes slowly. In the case of your question, my brain came to the
conclusion very quickly. But to someone else it might be that they
try for a while, before realizing that they do not really know the
answer. The speed at which the conclusion appears is a function of how
quickly the concepts settle down into a stable configuration that does
not include a slot filler.
Again, it is not that I dispute the role of "search". I only dispute
that the problem is particularly difficult to imagine a solution for
(at least, it is *not* difficult in a constraint-satisfaction
formalism, but I would not be surprised if in some other formalism it
turned out to be difficult.... I did ask you about that in my first
message, because I quite open to the possibility that some formalism
might make it into more of a problem than I think it is), and
secondly, I would dispute that the system keeps lists of stuff that it
does not know, as a general policy.
Well, now, that was a very satisfying response! You took at least part
of my initial post seriously and provided some very good arguments for a
paradigm in which the "feeling of not knowing" would be generated. And,
your scenario makes as much sense to me as the one I proposed. You have
taught me something and, for that, I thank you very much! We continue
to disagree on the "list" business. I should point out, once again,
that I am using that term as a conversational convenience, not
literally. Perhaps the same functionality could be provided in the form
of feedback to a constraint mechanism.
Well, thanks for saying so.
The way I would interpret the role of explicit knowledge that I do not
know something (as in the second time I am asked about the 1954 world
series) is that there is now a positive piece of new knowledge in the
system which should really be considered as just one more piece of the
puzzle that comes together when the system thinks about the question on
the second occasion. In other words, I would not see this piece of "I
do not know this fact" knowledge as a fundamental part of the mechanism
that goes about making judgments of knowledge ... it is not (for me at
least) part of the mechanism, but only a piece of data that is fed into
the mechanism. The mechanism proper is something that would observe the
relaxation process that occurs when the system is trying to recognize or
understand anything, and when it observes the relaxation process
behaving in a certain way - e.g. not resolving on the desired target -
it makes its decision and puts an "I do not know" element into the mix.
I do find these mechanisms interesting, because they play quite a
powerful role in the normal functioning of the system. In particular, I
find the role of the "not" element to be crucial and often
underestimated .... this is similar to but not quite the same as the
thig you have been talking about. In regular (simple) neural nets,
representing "not" is a big problem.
But enough for now.
Richard Loosemore
-------------------------------------------
agi
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/
Modify Your Subscription:
https://www.listbox.com/member/?member_id=8660244&id_secret=108809214-a0d121
Powered by Listbox: http://www.listbox.com
