Re: How much intelligence?

Curt Welch said:
Tony Orlow <aeo6@xxxxxxxxxxx> wrote:
Curt Welch said:

Curt, this isn't a matter of elevating us above animal status. I am
perfectly happy to admit that I am nothing but a smart ape. I agree that
associative leads to rational thinking, but as far as I can tell, the
mechanism is somewhat different. Rather than varying degrees of positive
or negative associations between sets of stimuli built up through
experience, which lead to emotional responses,

I don't think of walking to the kitchen to answer the phone as an emotional
response, yet we have no trouble explaining that behavior in terms of
reinforcement.

Inasmuch as it is the result of reinforcement, there is emotional response
involved, as far as I can tell. If you are expecting an important phone call,
you may jump to the phone, hoping to hear "good" news, or you may hesitate, if
you are afraid of hearing something "bad". Whether you are expecting a call or
not, one answers the phone due to reinforcement, based on the effect of one's
own responsiveness on the responsiveness of other people, and one's need for
response and interaction from those around them. On a basic level, simple human
contact is an inherent need anyway, and doesn't even rely on association with
other stimuli, since it has its own hardwired positive emotional value.

If you do something, it's because you are hoping for something good, or trying
to avoid something bad, and on that level, good and bad are emotional
evaluations.


natural language and other
apparently distinctly human abilities stem from a different type of
processing.

Yeah, many people like to believe that. I just don't happen to be one of
them. There's far too much in common with all behaviors to waste much
serious time trying to define it as different. It's mostly the people that
specialize in the study of language that seem to want to see their field of
study as being unique and special. Is that because they know more about
language than anyone (which we would expect to be true), or is it because
their focus has simply blinded them to what exists outside of their field?

That's always a point worth considering. People's areas of study often shape
their mentality, given the present state of the art and prevailing theories at
the time. Perhaps people become linguists because they feel language is
something unique and important, and maybe being around other people that
believe the same thing only reinforces that predilection for the field and its
way of thinking. Personally, I am not a linguist, and consider non-verbal
learning to be equally important to the processing of symbols and grammars, but
as a computer scientist, I have studied grammars and language parsing, and it
seems inherently different, algorithmically, from statistical association and
value assignment to sets of stimuli in real timespace.

Now, maybe a general purpose neuron can simply be rewired in such a way as to
produce this new effect. Maybe it doesn't require a new kind of cell. But I
would be hard pressed to accept, without some strong evidence, that the
emergent rational level of thought doesn't at least involve a novel sort of
neural arrangement, or an augmentation of a particular type of network that
exists in small part for some special purpose in other animals.


The people who focus on all behavior can't find a reason to justify
language behavior as somehow uniquely different from all the other
behaviors. If anyone should be an expert on the differences between
language behavior and other behaviors, shouldn't it be the Behaviorists and
not the Linguists?

Oh, I don't consider language to be a unique phenomenon at all. In my view, the
difference that makes us possible is a single tweak, either a new type of
connection that proved highly advantageous, or a pre-existing one that was
special purpose but became general purpose as a logical structure. In my
opinion, the mere capability of basic logical thought has given rise to all
sorts of "uniquely human" abilities, including language, but also including
music, art, mathematics, philosophy, religion, science, etc, etc, etc. They all
rely on the same basic ability to abstract from the here and now and apply to
the there and then. Language isn't paramount among all these abilities. Didn't
some used to consider music to be mankind's hughest achievement? I think it's
pointless for the heads of the hydra to compete that way. :)


When I propose a statement, it needs to be parsed for meaning
and correlated with concepts, but it doesn't need to be emotionally
evaluated.

Language has always struck me as one of the most emotional behaviors we
have. You see no emotion in a poem? In the words "Four score and seven
years ago .."? In the words, "I have a dream....", in the words "Get the
*** out of here you ass hole!". Language is not free from emotion by any
means. But how much emotion is there in bending down to pick up a pen when
you drop it? Some behaviors strike us as being emotional and some don't,
but the boundary doesn't seem to have anything to do with whether the
behavior is language behavior or some other type of behavior.

My point is that we have a basic emotional mechanism, associative learning,
which is well understood and easily modelable, and yet it doesn't give rise to
language. For language and other abstract logical activities, we need a
different basic mechanism, or a refinement of some type of the original
mechanism. I tend to think of it as another layer of processing, which derives
conclusions discretely by logical implication, rather than statistically by
causal implication.


Of course, the interplay between the emotional levels of mind
and the rational levels of mind are complex, and often confounding in the
search for a precise description of the relationship, but it's really not
that hard to characterize.

It's not really hard at all. It's all easily explained in terms of
reinforcement of behaviors which are contingent on the current
environmental context (where the internal state of the brain is part of the
environment).

Yes, but that's all very vague and doesn't lead to any kind of mechanical
explanation, or a clue as to how to implement such a beast.


The only answer to the type of argument stated above by Tony is to
actually build a reinforcement leaning machine that can learn to use
language the way humans use language. Nobody has done that.

Has anyone tried in the last 50 years?

You bet. I've spent over 30 years trying (off and on).

How's that going?


If it is possible simply by brute
force of associative learning to develop natural language, and the answer
is so Pavlovianly simple, why have efforts not yet succeeded along these
lines?

Understanding what a machine is doing, can be far easier than understanding
how it is doing it. We understand the what. We don't yet fully understand
the how.

Are you claiming that we have mapped out al the neural pathways and their
interactions in the brain? Uh, I rather think not. When you talk about the
"what", what are you talking about? There are a wide variety of neurons with
different kinds of synaptic interactions, in a number of different subsystems,
which we really don't understand fully at this point.


Have you used a compression program to reduce the size of a computer data
file? Seems almost like magic that it can make it smaller, and then later
return it to its full size. Yet, with a simple example, we can understand
how it can do this. We could see how it could change a string of zeros
into a special marker, followed by a single zero, and then a count of the
number of times to repeat that zero. If the file had enough long strings
of repeating characters, this simple technique would allow us to translate
the file to this encoded version which was smaller.

Yes, that's about the simplest compression routine, and there are many others.


But have you ever tried to write a compression program? If you have, you
would understand how hard it is to write a good compression program that
could take most files, and make them smaller. Understanding what it is
doing, is easy. Understanding how it can do it because we have an example
of one algorithm it could use, is easy. Understand how to make it work as
well as the ones we have on the shelf is very very hard. Every few years,
someone comes up with an improved algorithm for compression programs which
no one had thought of for the past 50 years. These things are really
interesting because once the algorithm is discovered, it's trivial to
understand it, and understand why it works so well - but yet it's something
a large collection of smart people never thought about, until that one guy
figured it out.

I learned a little bit about compression. What works depends largely on what
you're working with. If you have some characters that simply appear much more
often than some others, a Morse code type of compression might be a good
approach. If you are dealing with simple graphics, with large areas of a single
color, your approach (what is that, RLL, Run Length Limited, encoding?) would
be adequate. The tricky part is figuring out something that handles all
situations well.

So what is your point, that a single algorithm might be able to handle lots of
different cases? Sure, it might be able to do a specific thing pretty well with
most situations. But, does this apply to a kind of processing that is doing a
different specific thing?

If you want to claim that language is the result of associative learning, then
why don't you lay down the rules of associative learning through reinforcement,
and derive from those the rules of grammar and translation into linguistic
meaning. Then you will have something to work with, but I don't see an
efficient implementation of natural language coming about that way without some
serious tweaks to the underlying mechainsms involved in associative learning.


There are many algorithms like this which are easy to understand in theory,
but very hard to find a good implementation for. Sorting is another
classic example. It's a trivial problem, yet new sorting algorithms are
stilling be discovered:

http://en.wikipedia.org/wiki/Sort_algorithm

"library sort was first published in 2004"

Sure, sorting is a classic computer science topic. Is there an analog hardwired
in our brains?


For example, image recognition software that works as well as the brain has
never been duplicated. It's trivial to understand what the task is. And
it's trivial to write programs to demonstrate that it can be done. But
doing it as well as the brain, has not yet been duplicated. No one has
build software that can look a pictures, and tell you what's in the
picture. Is it a cat? Or a dog? Or a car? Or a house on a lake?
Nothing we have produced comes close to what humans can do but yet there's
no real magic about what the task is. It's just the "how" that continues
to elude us. But slowly but surely, the algorithms and the understanding
of how it's done, is improving.

Yeah, well, you may consider that ability to be a simple emergent phenomenon
arising from increased scale of associative learning, but it just ain't the
case. There is a very specific area of the visual cortex devoted largely to
facial recognition, although in certain types of visual experts, such as
botanists or birdwatchers, a large portion of this area may be devoted to
recognizing plants or birds. In other animals, it's used for recognizing each
other as well as other visually important things. There is a specific structure
involved in this kind of processing, and attempts to create some blanket theory
of mind that ignores specific details of the processing of specific types of
information is doomed to failure.


Reinforcement learning is the same thing. The principle about how it works
is easy to demonstrate, and easy to implement for simple cases, but it's
been a long slow road to find strong implementations that work well for the
hard problems. Image recognition is tied up in the same problem, because
if you can't do a good job of classifying the sensory data (be it vision,
or sound, or touch or whatever), then you can't do a good job of
associating behaviors (actions) with sensory events.

Very true. One of the hardest thing, I think, is processing raw sensory input
and producing mental objects that are easy and fruitful to process.


AI is an easy problem to understand, but a hard problem to implement. The
Behaviorists told us what human behavior is 50 years ago, but far too few
people believed what they were telling us. Surely man, the greatest thing
on the face of the planet, was more than a trial and error learning
machine? Nope.

Trial and error is not restricted to associative mechanisms. Explain how
associative mechanisms produce aggregate linguistic mechanisms if you want a
theory worth considering.


Lack of computing power or memory? Nope, that's not it.

Lack of the correct algorithm. If we had the right algorithms, we could
build human level AI today. It might cost a small fortune and take up a
warehouse, but it would probably be in our power to build it if we had the
right algorithms. I suspect however, that it wouldn't even be that
expensive - that a rack or two of special processors could do it. We can
already build very small, very complex, and very fast hardware into a very
small space these days.

Today's desktop could probably put us to shame. The basic algorithm for
associative reinforcement is simple. It's been done. It didn't produce the
expected emergent behaviors, because the theory that these emergent behaviors
are built from that algorithm is off the mark. There are several levels to
mental processing which are built in a hierarchy starting with love and
extending to logical deduction. Without a good map of how these levels
interact, you ain't got a model to work with.


Everyone understands that brains are learning machines of some type but
nobody knows with what type of learning machine we are. Is it lots of
different types of technology's welded together, is it one general
technology, or something between the two? This is what the field of AI
has been trying to get a handle on for 50 years and though the field
has produced a large amount of knowledge and new technologies, we
really don't know anything more today than we did 50 years ago about
what is the right technology because nothing yet produced has managed
to act like a human. There have been lots of promising results, but no
answers.

And what aspects of language have such attempts failed to produce, if not
all?

The correct question is what behavior has such attempts failed to produce.
And the answer is that a ton of behaviors far simpler than natural
language, have still failed to be reproduced. We haven't yet seen a
reinforcement learning machine learn to walk on two legs for example (as
far as I know).

That's because that kind of behavior requires a whole specialized feedback
network that operates automatically, like our cerebellum. So, the only robots
that walk on two legs have specific programming built in for that purpose.
Expecting a super-simple model like that to be able to control a physical
system moving in space is kind of unreasonable.


All language is just behavior, and all behavior is language.

And it's all just stuff. Very deep.

We don't call
it body language for the fun of it. Seeing somebody smile at us has
meaning just like when someone speaks to us. It's not a different type of
meaning, it's the same type of meaning. We understand all sensory data
using the same systems for extracting "meaning" out of the sensory data.

That kind of communication is not natural language. I can have a fine verbal
conversation with a fox or a cat, but that's not quite natural language either.
One can have very meaningful facial interactions with a dog, but that's not
natural language either, and neither is gently handling a snake to convey
friendly intentions. Natural language involves words that have meaning
generally unrelated to their form.

We know the meaning of the phone ringing as much as we know the meaning of
thunder as much as we know the meaning of a water bottle. All these things
allow us to predict things about the future, and allows us to select
behaviors to deal with what we expect the future to hold. We see a water
bottle and we know that we can fix our thirst by using the correct arm
motions to drink the water hidden in the bottle even though we can't see
the water.

Any dog can do that. Why can't they read a book?


We see a message written on the note pad that says "I left the water bottle
in in the fridge" and we extract the same meaning from the vision of that
note that we extracted from the vision of the glass of water. It regulates
the behaviors we will select in the future to respond to thirst.

Do you leave notes for your dog? Does the dog not have enough associative power
to associate those chicken scratches on paper with relief from thirst?


The entire world is constantly speaking to us. The fact that some of the
things we "hear" in the environment we choose to call "English" doesn't
mean that that sensory data was processed with a different type of hardware
than what was used to process what the water bottle "said" to us.

Curt, you are blurring the meaning of natural language so grossly that we're
not even on topic any more. Sure, associative learning is involved in learning
all those concepts. Linguistic communication is another thing entirely.


The brain is a language machine.

Uh, I thought you were just complaining about people that elevate language
above other behaviors. Now you say that's what the brain does? I don't think
you know what language is.

It receives a constant flow of language
from the environment in the form of sensory data, and in response, it sends
out a constant flow of language to our muscles to make them move. The
syntax, grammar, and semantics of the raw sensory data that the brain deals
with is far more complex than the fairly trivial syntax and semantics of
English. Natural language like English is just a trivial small subset of
all the language the brain has to understand, and produce just to stay
alive. Natural language is just the top of the iceberg in terms of the
language that the brain has to deal with. The reason we have had so much
trouble building specialized language machines (chat bots for example) is
because that without the foundation of language processing that allows us
to understand what the raw sensory data that represents a bottle of water
means, we have no hope of understanding what a few written words mean.

Well, that's all very wonderfully mysterious and everything, but I don't think
that's going to help you implement anything. Good luck, anyway.


Doesn't grammatical structure seem like an inherent ingredient in
the recipe for language?

Of course it is. But the brain is dealing with that same problem in a far
more complex form as it learns to do even the most basic things in life,
like recognizing a water bottle and not mistaking a picture of a water
bottle for a real water bottle because one can be used to solve a thirst
problem and the other can't.

That's an entirely different function. Any camel can do that.


Do you know that when a child is raised in
isolation, like a closet or attic, if they don't learn natural language
by a certain age they never can?

Are you aware that if a baby is born with a cataract in their eye, it must
be fixed very quickly, because if you wait too long, the brain will never
learn to use the bad eye even though the eye is fixed and working
perfectly? And when they do fix these problems in babies, they must tape
the good eye closed in order to allow the bad eye time to "catch up" with
the good eye.

Yes, I've heard that.


This happens because the brain is very plastic, and the neurons that would
normally have been used for the bad eye get co-opted for use processing
data from the good eye.

Yes, the net is malleable, within each subsystem. The visual center requires
training, rather than being hardwired to recognize particular types of
environmental features, so that an individual can develop according to its
environment. Eskimos have something like 19 words for snow. Saharans probably
have as many for different type of sand. It all depends what your world is
like.


The cortex is very plastic, and if you don't use it for one type of data,
it will be consumed and allocated to processes other data. That's why if
the language center isn't allowed to develop for language use, it will be
taken over for some other purpose, and once allocated, it never goes back.

Actually, it kind of atrophies, once it is finalized without having been used.
It's not like you can use the visual cortex for auditory data if you are born
without eyes, and with temporal lobe damage. Since both eyes feed into the same
subsystem, it's possible to coopt the other half if it's not in use, but that's
not always the case.


Do you know that there are cases where
more than one child has been hidden away like this and that, given the
opportunity, they will develop their own language spontaneously? If you
want to talk about Skinner's Box, you might want to keep these other
boxes in mind. They clearly indicate that there are innate mechanisms in
the human mind meant to handle abstraction and language

Actually, your example strongly suggests it's not natural language
hardware. If it were natural language hardware, it wouldn't be taken over
for another purpose, such as vision or other aspects of sound besides
language. Your example suggests that the hardware we use for language can
easily be used for other functions and isn't specialized only for the use
of language.

No it doesn't. These children are not using the unused verbal part of the brain
for other purposes. They don't have any other purpose. It just reaches a point
where the brain is supposed to have learned all the different basic sound types
it's supposed to recognize, and fixes it as permanent so that it can move on
with the next level of development. Ever read about Piaget and the stages of
childhood development? There are all sorts of areas where one learns what they
learn by a certain age.

The point is that, without anyone teaching them language, they develop it
spontaneously. What are they creating these extra associations about, which
lead to this enhanced scale of association which you say results in language,
when they are locked in a dark closet? It's not the result of a lot of
associations, but of a distinct inherent ability to assign meaning to otherwise
meaningless inputs.

Okay, so that sounds like assigning meaning through association, and maybe the
link mechanism is the same, but the mechainsm creating the link is different.


But, as I've suggested, a think a better way to look at the brain is as a
language machine. Everything it does is language processing, not just the
subset set of behaviors we call English.

I think your definition of language is too broad. What was it, specifically?


which are greatly
reduced in those other animals which have them, and distinct in nature
from emotional processing.

All the people that carefully study behavior, find no justification for
using the term "emotional processing".

That's because behaviorists look at the animal as a black box which produces
behavior in response to stimuli. It's not my fault you entirely ignore the
inner structure of the mind, except for the Pavlovian associative mechanism,
and fail to explain behavior based on this mechanism alone. Until you can
explain love in mechanical terms, you don't have even the semblance of a mind
theory.


I am not saying other animals don't have
logic. That border collie in Germany and that Afrrican Grey parrot both
exhibit clear simple logic. That's what really separates us from other
animals, the degree of logic, not of emotional processing.

Of course, I am always happy to hear specific counterarguments. :)

The low level hardware that allows us to all of this, seems to be very
generic, and very plastic. It reacts to whatever data is feed to it. It's
all generic signal (aka language) processing hardware.

That's a very generic and plastic statement, so much that it hardly says
anything, except that mind does all sorts of stuff.


Our ability to abstract is just as important when chasing a rabbit for food
as it is when we use language. We need to take the raw sensory data of
this furry grey object and abstract out the concept that it's food that we
can eat and not a rock. When we chase a rabbit, and we sense a vertical
brown band in our vision, we need to correct abstract out the fact that
it's a tree that will stop us from getting the food if we don't correctly
act to move around it before we run into it.

Poppy***. If it moves, it's food. That doesn't take any abstraction. Any skunk
can do that. It's not abstraction, and it's not language.


When the furry grey rabbit vanishes from sight behind the tree, we need to
correct abstract the idea that the rabbit is not gone, but simply hiding
behind the tree so that we can run around the tree to find it.

Only a chicken would be fooled by such a trick.


All behavior is nothing more than a problem of selecting the correct
behavior to produce in response to the current environment we find
ourselves in.

That is correct.

And that environment might at times include the memory of a
rabbit we just saw 5 seconds ago before it ran behind the tree, or might
include a sequence of sounds we call words. In all cases, we are able to
select, and produce, a very wide range of different behaviors in response
to a very complex and changing multidimensional environment. And it's all
done in parallel so that one behavior might be selected for the right leg
and another might be selected for the left arm that has nothing to do with
the elements of the environment triggered the right leg behavior.

I don't understand your implementation details.


The neocortex is playing a key role in processing the raw sensory data and
abstracting a valid description of the environment by the patterns of
neural activity it creates. We know the fuzzy grey thing zipping across
our field of vision is the same type of thing we just ate last week because
the neocortex has created the same patterns of activity for this grey patch
of light as it did for the one last week which we ate and really liked the
taste of even though that thing was very different in many ways. It had to
correctly define the concept of "good tasting rabbit" and correctly
associate that grey streak of light with that concept. It plays the same
role when it correctly manages to classify strings of sounds as a given
word and strings of words to have a collection of abstract meanings. We
learn how to respond to all these different complex combinations of
environmental conditions by a long slow processes of reinforcement
learning. That training tells the brain which conditions to ignore, and
which to use as triggers of different behaviors.

Do you honestly think all that is going on in a little kitten's head while she
chases a dangled string? I don't. That behavior is hardwired like so much into
the kitten's head. There isn't even associative learning involved in the
initial play, much less language or anything logical.


When it's all done, we do end up with natural language hardware. But that
language hardware was shaped out of the same context sensitive behavior
generating hardware we use to create all our behaviors.

Then it should be easy to implement. What have you been doiung these past 30
years? Get to it.


We know that our finger motions are under the control of reinforcement
learning. If the fingers are also used to type language, how can language
not be under the control of the same reinforcement learning system? The
brain only gave us one system for learning to move our fingers and that one
system is used for all finger motions we learn to produce. It makes no
sense to assume evolution did it any other way.

In fact, our fingers all have little globs of grey matter in them. They do a
lot of their own learning, without bugging the brain, or even the cerebellum,
much, from what I have heard.


The reason we have strong language skills when other animals don't is just
because we have a more complex motion controller system that has enough of
the right hardware in the right configuration to produce all the behaviors
needed for language. Learning to communicate in English is like learning
100,000 tricks. If your behavior learning hardware can only learn 1,000
tricks, then there's no way you can use it to learn language. Or if the
behavior learning hardware only produces a context which goes back 2
seconds, you can't learn to parse and correctly respond to a 10 second
string of sounds.

If it's as simple as all that, then do it.


So I agree that we have hardware which is configured in just the right way
that it allows us to learn all the tricks we call natural language skills,
but I don't agree that it's special hardware that works differently than
the hardware which allows us to learn all our other behavior tricks. It's
all one big reinforcement learning system built to allow us to adapt all
our behavior to the needs of our environment during our life time.



Well, I see that as a squishy subtheory. It doesn't explain much or predict
much. The only way to really support such a theory is to explain exactly how
the language abilities derive from the associative ones, or implement something
that works on those priciples. I think you are going to find that what
establishes logical implications is a distinctly different mechanism from what
establishes causal implications, and that a distinctly different type of
structure is required. But, that's just my opinion. What do I know? :)

--
Smiles,

Tony
.

Loading