Re: natural intelligence

Alpha <omegazero2003@xxxxxxxxx> wrote:
On Feb 18, 4:45=A0pm, c...@xxxxxxxx (Curt Welch) wrote:
Alpha <omegazero2...@xxxxxxxxx> wrote:

That is key; just what does it mean to understand something. I think
it means beig able to refer to an object or concept or precept or
recept *as* the thing a word or phrase or sentence referes to. And
further, to be able to reason (using language, intuition, analogy,
fuzzy logic etc.) about htose concept/percepts/recepts etc.

Tying the idea of understanding to the ability to communicate about it with
language certainly seems to fit some common usage of the word. I like to
think of it in broader terms, so I can say things like a dog understands
the food is behind the door (when he tries to get the door open).

Can a dog understand if he doesn't have any ability to talk about what he
understands?

Certainly my personal understanding seems to be closely connected with my
language skills, but at the same time, there is much I feel I can
understand without any reference to language. I can say I understand how
water pours out of a glass by visualizing what happens when I turn the
glass of water sideways and watch the water pour out. Though I can attempt
to describe my vision with words, no amount of words would in fact fully
describe what I'm visualizing in my head - and as such, I would have to
claim that no amount of words would explain my understanding of such an
event.


=A0So
people look for ways to "store knowledge" in their computer using a
simil=
ar
system, but which is easy for a computer to understand.

I don't thin the computer understand something like its 0s/1s the same
way we cognize words and images and memories etc.

Sure, certainly we have a level of understanding that's created by our
ability to talk about these things that the computer doesn't (yet) have.
So it's quite easy to argue that our level of understanding is far greater
because of that.

But when we try to figure out just what understanding is, it's hard to draw
the line at any one spot.

One aspect of understanding is that it's a power to predict something about
the future - such as my example of understanding how water pours out of a
glass. My understanding of that allows me to make predictions about what
will happen if I pick up that glass of water and turn it sideways.

Maybe using the power to predict future events would be a better way to
define understanding. In which case, computers have some understanding
because we can give them some powers to predict some things about the
future.

=A0It's basically a
problem that the computer can't use such a list until it can first
understand the language the list is created in. =A0So we have explored
simpler more precise languages which we can use in our computers to
store all possible knowledge.

Storing is not understanding.

How about if that storing is used to help predict the future?

I've not studied such language systems (like CYC), but the bottom line
seems to be that these language systems are good at representing high
lev=
el
information, but not the low level details that support it.

Well, domain knowledge (ontologies/subontologies) is based on theories
which are in turn based on microtheories which are in turn based on
sentences that confer a TPC-like logical status to the concept
represented by the sentendce. Like, "water above 200 degrees will
burn a human's skin." Theat is the sentence/concept. The mocrotheory
in which it fits might be a sub-ontology or domain of physiology and
water thermodynamics while the theories on play might be physics and
biology.

=A0That is, they
only capture the tip of the iceberg, and never the much larger lower
foundation which supports it. =A0Attempts to fill in the lower level
foundation seems to have always failed - no amount of "filling in
details=
"
seems to have worked. =A0No matter how many details get filled in, the
am=
ount
still missing always seems larger than the amount stored.

I donl't think so with large ontologies . Drug design & development
companies fill in the details from top to bottom- at least to
molecular protein folding all the way to inter-drug interactions and
human physiological responses.


I believe approaching the problem of representation from this directly
wi=
ll
never work. =A0Though it "seems" as if we are using language based
system=
s
externally to store knowledge, we aren't. =A0Most the knowledge is
still =
in
us, and not in the book, or in the list. =A0The language words just
trigg=
er
the activation of the true information which is stored in our brain as
we read the words. =A0The words we write in the book are more like keys
that unlock the knowledge that is already in us. =A0As such, trying to
build computer systems that are "word based" in nature, fails to
capture the re=
al
knowledge - because the knowledge never was in the words to start with.

I agree with that; there is something about "understanding" that
elicits an emotional and consciousness-based reation in humans. We
experience the understanding; we become one with it.


So, what is the knowledge in us?

I believe the correct way to understand it all, is to understand that
humans are simply complex reaction machines. =A0We produce complex
tempor=
al
reactions in response to complex temporal stimulus signals.

If someone asks me the question, "what did you do today", and I respond
with, "I went to get my car fixed and wrote this Usenet post using the
computers in the lounge at the dealer while I was waiting for the work
to get done", then I've just produced a complex temporal reaction to a
compl=
ex
temporal stimulus.

The only "knowledge" stored "in me" was the fact that my brain was
wired =
so
as to produce that reaction, to that stimulus. =A0Everything a human
does=
,
and everything we think about, can be explained as the behavior of a
complex reaction machine.

But that is too high level to get us anywhere. Simple being a
reatcion cannot confer understanding of a concept. I can react to W&P
with understanding or ambivalence or boredom or throwing the book in a
fire. It is the why of the understanding that gets us to the next
step in a sequnce of steps an agent undergoes after "reading"
something.

Yeah, maybe it's better to say a reaction that shows a purpose based on
reasonable predictions about the future? A machine that learns to run away
from the guy that keeps trying to hit is "pain" button would be showing an
understanding that running away is likely to prevent the thing it doesn't
like from happening in the future?

So, the first question in my view, is how do you build a reaction
machine that can produce these types of complex reactions to complex
stimulus events? =A0And the second question, is how can you make such a
machine, s=
o
that the way it will react, is constantly changing based on experience?

I see.


If we look at a chat-bot type problem, we often try to code it by
making the computer receive some input from a human (they type a
question, or statement), and then the chat-bot produces a response, and
then waits for the next input.

That's a fun toy to explore, but it's not the same class problem the
brai=
n
deals with. =A0The brain produces a continuous flow of output
behaviors, =
in
response to a continuous flow of stimulus signals. =A0There's none of
thi=
s
back and forth taking turns going on. =A0As such, the "reaction
machine" =
I
believe we need to build, must work the same way. =A0It must produce a
continuous flow of output behaviors in response to continuous flow of
inp=
ut
behaviors. =A0It must be a real time system - the data flows are real
tim=
e
data flows.

Well control systems that I have worked on - very complex hard real-
time systems, have that quality, yet nothing in the system
undertstands that it is part of a system that is controlling part of a
petro-chemical plant. There are upper-level representation of abstract
aspects of the system. For example, waste is an abstract concept that
the system is programmed to calculate. Or production quotas and
accouting parameters.

Yeah, with such a system, we might say it understands something low level
like turning a valve will regulate the temperature, but we can't say it
understands the higher level idea of what temperature is or why it it's
there, or all the many other higher level ideas the humans that designed
the system understand.

I think people need to stop thinking about human behavior from the high
level (which is how we mostly like to think about humans), and revert
to the better defined engineering problem of building high quality
reinforcement trained reaction machines. =A0The representation problem
is solved by finding clever ways to represent the way the machine needs
to react at the lowest of levels, and not by attempting to represent
the desired high-level behavior we want to emerge from the machine.

So everything interesting will just emerge? Is that how it happens in
humans from birth?

I think so.

The reinforcement learning machine must be designed in a way to allow
complex behaviors to slowly emerge by first learning simple behaviors and
then building on them. And, to equal human powers of behavior (language
and all), it's got to have enough innate learning power to learn all those
types of behaviors. - just like a computer is limited by how much memory
and processing power it has, the learning machine we want to human level
behaviors emerge from it must have enough innate "memory" and "speed" to
produce all those behaviors. This might include special configurations to
support the processing required for skills like language.

But other than these things. I think the basic way humans work, is that
they are large high power generic learning systems and that all our
interesting complex behaviors emerge from us a result of us trying to seek
rewards from interacting with our complex environment.

--
Curt Welch http://CurtWelch.Com/
curt@xxxxxxxx http://NewsReader.Com/
.

Loading