how would a computer data base RE: visual audio communications tied in with biotech worch in Chinese as the foundation language s?
N
was chattig with some guy at nearly the last of the last fashionable
'set' in the hive and we said 'Dell!' but basically I cant imagine any
hot industy engineering technology dept not taking on and paying, ne,
nurturing the best minds to presuppose a product...? ....?...?...?...?
anyways.....sci,lan
Don Stockbauer
AI at its best.
pataphor
No, probably 4chan, and it proves my point. You know, the point Curt
claims to get but then completely ignores.
P.
Curt Welch
> On Thu, 17 Dec 2009 21:12:43 -0800 (PST)
> Don Stockbauer <don.sto...@gmail.com> wrote:
>
> > > so? =A0I have a basic engineering and arts foundation? Tell you what!
> > > I was chattig with some guy at nearly the last of the last
> > > fashionable 'set' in the hive and we said 'Dell!' but basically I
> > > cant imagine any hot industy engineering technology dept not taking
> > > on and paying, ne, nurturing the best minds to presuppose a
> > > product...? ....?...?...?...? anyways.....sci,lan
> > AI at its best.
>
> No, probably 4chan, and it proves my point. You know, the point Curt
> claims to get but then completely ignores.
What point did I completely ignore?
--
Curt Welch http://CurtWelch.Com/
cu...@kcwc.com http://NewsReader.Com/
pataphor
cu...@kcwc.com (Curt Welch) wrote:
> pataphor <pata...@gmail.com> wrote:
> > No, probably 4chan, and it proves my point. You know, the point Curt
> > claims to get but then completely ignores.
>
> What point did I completely ignore?
Well, you want to study the intelligence of a single human. But in case
you're dealing with 4chan this is clearly futile. They have herd
intelligence though.
I probably should have written "acknowledges and then carries on
as if nothing happened".
P.
Curt Welch
Oh, that point.
So you actually believe a single human has no intelligence or that there's
nothing to study there, or duplicate in AI?
You realize right that if you get the single human AI working, then all you
have to do to create herd intelligence is build 100 of the same thing
right?
So what exactly am I'm ignoring by trying to build the thing I have to
build 100 of to create herd intelligence?
pataphor
> So you actually believe a single human has no intelligence or that there's
> nothing to study there, or duplicate in AI?
No, just that academics are no more intelligent than other humans, even
though they conjure up intelligence tests, dress up for ritual
ceremonies, and produce outrageously complex scientific theories to set
them apart from the rest.
> You realize right that if you get the single human AI working, then all you
> have to do to create herd intelligence is build 100 of the same thing
> right?
This probably means you still do not take my theory for real.
You'd never find the right single human AI that way. You can't design
atmospheric molecules by some lucky break, so that it will result in it
snowing horizontally with very thin powder snow. It is something one
will never come up with if one only looks at the individual particles.
There are interactions that cannot be uniquely assigned to the
individual elements.
Maybe a metaphor would help, so, at the risk of making things even more
unclear:
If you'd generate all the permutations of some strings with repeated
elements, they wouldn't just be easily broken down into groups the way
one could do it if there were no repeats.
> So what exactly am I'm ignoring by trying to build the thing I have to
> build 100 of to create herd intelligence?
Please don't take it the wrong way. By writing very long posts you more
or less give me read access to your mental hard drive, and naturally I
will find places that are not updated to the latest theory. If one would
have to scan all data for consistency each time a new idea was added,
one would have no time to act. If I would write as much as you there
would be inconsistencies becoming visible too. Meanwhile, I very much
appreciate your posts.
Nevertheless, once one starts seeing human intelligence as a distributed
phenomenon, like individual neurons not holding the grandmother data,
one cannot but conclude that consciousness does not reside in the
individual brain. And the update to that info can take a long time, and
mine is certainly not finished yet.
For example, very speculative, what if learning is a process of neurons
starting to fire together in some sequence but this is not limited to a
single brain anymore? The same algorithm that causes our brains to
select one reality over another (it decides to 'see' things in a certain
way) also makes it want to enlist other brains in this view and suppress
conflicting info even if that info is another human's brain. We just
have leaped over the individual human brain barrier. Now we know why we
communicate and do politics and religion.
P.
Curt Welch
> Curt Welch wrote:
>
> > So you actually believe a single human has no intelligence or that
> > there's nothing to study there, or duplicate in AI?
>
> No, just that academics are no more intelligent than other humans, even
> though they conjure up intelligence tests, dress up for ritual
> ceremonies, and produce outrageously complex scientific theories to set
> them apart from the rest.
Well, maybe you just don't know enough about my views yet because I
certainly don't think of academics as being the example of intelligence we
should be looking to duplicate. Not that there is not plenty of
intelligence there, just that's not how I work on AI.
My approach to AI is a bottom up one. Some like to attack the problem from
the top down. In my direction, I try to figure out how to make very simple
signal processing systems perform highly simple acts of intelligence. I
design and build various neural networks and try to figure out how a neural
network should be built to make the entire network act intelligent - which
INCLUDES making multiple networks interact intelligently.
The more classic approach to AI is top down, where you take some aspect of
high level human behavior, and try to duplicate that with software - such
as chess playing, or reasoning, or car driving, etc.
If someone were to study what the academics do, and try to duplicate that
in AI, that would be a top down approach - not something I do at all. I
work on the very basic skills I believe is needed to create _ALL_ forms of
intelligence that exists in all humans, and even in most mammals, and try
to make networks do those things.
I believe the most fundamental low level behavior is behavior learning
itself. That is, how we learn new intelligent behaviors. I think without
getting learning working correctly, nothing else in AI is even worth
talking about. You have to figure out how to make machines that can learn
like humans before you can even begin to work on the rest of AI.
I work on trying to get my machines to learn the most trivial behaviors
currently. It's what the entire field of machine learning does, because
that's how far away from making AI we currently are.
> > You realize right that if you get the single human AI working, then all
> > you have to do to create herd intelligence is build 100 of the same
> > thing right?
>
> This probably means you still do not take my theory for real.
You haven't given us any theories yet to "take for real". You have only
thrown out a handful of highly vague comments without actually telling us
what your point is or what you believe is or why you are even here. Your
writing style is very odd that way. You are very indirect as is if you
were scared to commit to an idea and talk about it.
You are free to tell us, or not tell us, whatever you want to. But, are
you for example, trying to solve AI by actually building machines? Or, do
you just philosophy about what humans are, and what others how work in AI
should be doing? That's a common problem here. :)
Do you think you have theories that would actually help us build machines
that acted intelligent? Or do you just have words that you think have some
high degree of self importance in reveling the truth nature of humans? I
can't tell the answer to any of these questions yet based on how much you
have written so far.
You have made multiple references to computers which makes be think you are
a software developer. If so, are you using your skills to actually try to
create an intelligent machine? Ot just talking about it?
This is the philosophy of AI so it is the place to just talk about, but if
you are going to suggest _we_ should "take your theory for real" (meaning
you think I should change the AI work I play with and my approach) just
because you have given me some new insight I've never thought of, you will
have to come up with something far more clever and useful than "humans are
more intelligent in a herd" - which is all I can grasp you have said so
far.
> You'd never find the right single human AI that way.
The aproach to AI that _everyone_ has used since the beginging is
1) theorize
2) build
3) test
4) repeat
I'm sorry, but that approach will _always_ find the right answer in the end
no matter how you approach step one. That's because in step three, if your
machines are not creating an intelligent herd, then you failed to get steps
one and two right so you refine your approach and try again.
This technique will always converge on a good solution, if there's a good
solution available to be converged on, that the we are able to understand.
> You can't design
> atmospheric molecules by some lucky break, so that it will result in it
> snowing horizontally with very thin powder snow. It is something one
> will never come up with if one only looks at the individual particles.
> There are interactions that cannot be uniquely assigned to the
> individual elements.
That's just not true.
You start with some "best guess", (step one), build it, and test it, see if
it produces your "snowing horizontally with very thin powder snow" and when
not, you itterativly improve you "best guess" to try and make it produce
better results. You keep doing that until you get the right result.
What you don't seem to grasp, is that the design choices are not random
guesses. They use the data collected from real test results, to _improve_
the last design. If improvements are not getting better results, then you
thrown out the previous approach, and look for a new one.
It's a search guided by studying the results of your tests. And the
results of those tests very much do allow you to quickly converge on the
right answer by quickly eliminating large amount of "wrong" answer and
"wrong" approaches.
> Maybe a metaphor would help, so, at the risk of making things even more
> unclear:
>
> If you'd generate all the permutations of some strings with repeated
> elements, they wouldn't just be easily broken down into groups the way
> one could do it if there were no repeats.
Yes, you are being very unclear. You need to learn to communicate better.
You are not helping the herd intelligence here. :) You clearly have a
sharp mind that likes to think about things in deep ways (which is good),
but to allow others to benefit from your thoughts, you have to find ways to
communicate with others.
> > So what exactly am I'm ignoring by trying to build the thing I have to
> > build 100 of to create herd intelligence?
>
> Please don't take it the wrong way.
We take each other the wrong way all the time time here. Don't worry about
that.
> By writing very long posts you more
> or less give me read access to your mental hard drive, and naturally I
> will find places that are not updated to the latest theory. If one would
> have to scan all data for consistency each time a new idea was added,
> one would have no time to act. If I would write as much as you there
> would be inconsistencies becoming visible too. Meanwhile, I very much
> appreciate your posts.
>
> Nevertheless, once one starts seeing human intelligence as a distributed
> phenomenon, like individual neurons not holding the grandmother data,
> one cannot but conclude that consciousness does not reside in the
> individual brain. And the update to that info can take a long time, and
> mine is certainly not finished yet.
Then develop it further by writing long posts here and letting people pick
them apart. The more you share your ideas, and interact with the other
people here, the more intelligent the distributed intelligence here
becomes. The less you share, the less the distributed intelligence has the
ability to grow.
We just bull shit here. This is Usenet. This is not the scientific
community. Very few of us here are actually part of the formal distributed
academic community and their distributed intelligence efforts to solve AI.
There have been various people here in the past that are well known parts
of the academic efforts at solving AI, but there are less and less of them
as time goes on. They don't have time to bull shit here, if they are
actually trying to get paid for producing papers and teaching.
The subject of distributed intelligence is nothing new here. We have
debated it constantly for years. Where as I am "Mr Reinforcement
learning", Don is "Mr. Global Brain". - which is the point that our society
(and all human organizations like clubs, and religious groups, and
corporations, and countries) should be seen as more than just a society,
but as a real itnelil9gence in it's own right, separate, and generally more
intelligent, and more powerful, than any of it's members.
The ideas of group intelligence you mention are nothing new. It's an old
and well debated topic.
I don't just think that distributed intelligence exists outside the brain,
I think it also exists _inside_ the brain. I think individual intelligence
is created by a distributed process of neurons interacting with each other
in a society of neurons. I think the only way to solve the problems of
scale in learning, is to attack the problem as a distributed society.
Marvin Minsky who is one of the very well known names in AI, who used to
post here on a regular basis many yeas back, wrote a very popular book
called "Society of Mind" where he lays out his tries of how the brain is a
collective society of agents working together to create the end result of
human intelligence.
Your idea that we need to look at a society to explain intelligence is
really nothing new. And so far, that's all you have said about what your
idea is.
> For example, very speculative, what if learning is a process of neurons
> starting to fire together in some sequence but this is not limited to a
> single brain anymore? The same algorithm that causes our brains to
> select one reality over another (it decides to 'see' things in a certain
> way) also makes it want to enlist other brains in this view and suppress
> conflicting info even if that info is another human's brain. We just
> have leaped over the individual human brain barrier. Now we know why we
> communicate and do politics and religion.
>
> P.
Yes, exactly. All well known idea. Nothing new yet. Keep trying.
When we work together, and share a goal, we have created a larger
distributed intelligence. If we interact with each other here, we form a
distributed intelligence in the group with a goal of solving AI, or at
least, of better understanding the problem of AI and better understanding
ourselves.
But again, you seem to want to learn to the idea that intelligence doesn't
exist in a single brain - that it really only forms when you get multiple
brains interacting. I just don't believe that's true in any sense. I
think the intelligence starts at a far lower level, and by the time you get
a single human, you already have a huge distributed intelligence at work
inside that one humans. And when that one human interacts with other
humans, the span of that distributed actions just grows even wider.
How neurons interact with each other to create the single intelligence is
of course very different than how humans interact with each other. Neurons
use pulse signals and the link, humans use language, and physical
interaction, etc. And how the goals are specified and communicated are
different. NO on fully understand how it works in the brain yet, but it no
doubt makes heavy use of chemical messengers. Outside, we use tokens like
money, and many other abstract techniques for communicating and sharing
goals. But the result is still the same - that as, when the systems are
able to work together in some aspect of a common goal, it's creating a
greater intelligence.
One example I like as a thought demonstration of this greater intelligence
is to think from the perspective of an alien watching the earth from a
distance. They see a large asteroid flying towards the planet, and the
plant reaches out (with missiles) and knocks the asteroid out of he way.
That's an act of the earth's global intelligence doing at the level of the
earth, the same the our brain does, when someone throws something at us,
and we knock it out of the way with our arm.
How the parts "inside the earth" work to make the earth intelligent like
that, you can't see from a distance. But just watching the earth knock the
asteroid out of the way, is a clear sign that the earth itself has a fairly
high level of intelligence. Likewise, when we see a human knock a rock
thrown at it out of the way, we can't see the small parts inside that make
that happen, but we know the human has some high level of intelligence
because of their ability to do that.
Bot the earth as a whole, is already one large intelligence, and a single
human, is one large intelligence.
So now that we have multiple examples of intelligent systems, how do we
make a computer intelligent? I believe we do it by building a
reinforcement learning machine (or society if you would like) out of a
society of interacting signal units that share a common reward signal that
allows them all to play a small part in the bigger action. There is no end
of different ways you could try to create such a system. I have a basic
architecture I find attractive I've been playing with the past few years
based on the concept of pulse sorting.
We just keep searching by taking an educated guess, explore where it leads
us, when it doesn't lead us to human level machine intelligence, we adjust
and try again. Repeat until done.
I believe it's highly important to understand society is a higher
intelligence. Not that it just acts intelligent, but that IT IS a higher
intelligence which is separate from the individual intelligence of the
humans (and machines) that make it up. You are preaching to the choir here
with you ideas that social intelligence is important.
BTW, in my "brain dump" work here by writing so much - it took me about 1
1/2 hours to think up, and write this message - just to give you an idea of
how much I waste with this stuff. Now I've got to out out and shovel show.
We had about 20" of crap dump on us yesterday - a hug amount of snow for
this area (Washington DC metro area).
pataphor
> Well, maybe you just don't know enough about my views yet because I
> certainly don't think of academics as being the example of intelligence we
> should be looking to duplicate. Not that there is not plenty of
> intelligence there, just that's not how I work on AI.
It's not just academics, but all people are having more or less the same
intelligence level. Like cars, there are some that accelerate from zero
to a hundred kilometers per hour in 5 seconds, and there are those that
take a bit longer, but they are all cars, just vessels. In the coming
era we will have more or less well equipped avatars and it will be gross
to discriminate against someone only because her vessel or avatar is
operating in low resolution or maybe just in black and white.
What I am arguing against is people trying to use intelligence claims as
a status tool. This can happen in the seemingly most open communities.
For example, people arguing in favor of ending all animal suffering,
elevating animals to a higher level, but at the same time making
inappropriate distinctions between humans based on tiny intelligence
differences (surely, if seen from an interspecies level viewpoint) which
are for the most part fabricated because of status drive. I believe it
is more prevalent in younger people, with higher testosterone levels
making them want to appear bigger than they are (physically and
intellectually) so as to increase their chances of finding a mate.
Of course there are also those who do it for acquiring more resources.
But the whole point is that all our culture is the result of a
collective effort, where for the most part the people with the ideas are
not the ones who get the credit. Instead, the credit goes to their
supervisors, or to the music industry if they are musicians, or to the
shrewd businessmen exploiting someone else's ideas.
The sincere innocence with which people take ownership of the ideas in
their heads is often just self serving stupidity. They think, 'because I
am the one who thought of this first, it is *my* idea'. Not so though,
because a lot of times ideas are adaptations of ideas that other people
had earlier. Then there are people who accuse other people of stealing
when they are stealing themselves and their self importance doesn't
allow them to look back at the 2000 year old sources of their ideas.
In my opinion it is a convoluted mess and therefore it's very hard to
attribute ideas to persons even when the specific person is yourself and
the idea is a new one in your own head that you are sure you never heard
about before.
Moreover, the whole attribution thing is counterproductive because the
mentality fits in a world view we (humanity) are trying to leave behind
us, the idea that it is shameful not to work and that greed is good.
Instead, we as humanity are migrating towards the idea that people work
just because they like it, and that job hunting has bad consequences if
it means increased social inequality because a master, or owner, class
is exploiting the basic physical needs of a lower class by controlling
the job market by means of owning the production tools. There are signs
this is all going to change, solar cells are a one time purchase,
freeing one from the energy overlords, the rep rap project promises
local manufacturing, and in general there is a trend away from centrally
controlled resource markets.
> My approach to AI is a bottom up one. Some like to attack the problem from
> the top down. In my direction, I try to figure out how to make very simple
> signal processing systems perform highly simple acts of intelligence. I
> design and build various neural networks and try to figure out how a neural
> network should be built to make the entire network act intelligent - which
> INCLUDES making multiple networks interact intelligently.
Sorry but I still suspect you are too proud to look at the statistical
properties of the data because of some ill advised idea that your
algorithm, if it is any good, will fend for itself. This is exactly the
wrong kind of approaching the problem. AI is a 'win anyway you can'
problem not some aesthetic sword duel between gentlemen.
> The more classic approach to AI is top down, where you take some aspect of
> high level human behavior, and try to duplicate that with software - such
> as chess playing, or reasoning, or car driving, etc.
What I see in Go programming for example is people using a very low
level language like C because 'it is fast' and then never switching to a
higher programming language after the basic libraries are finished. In
fact I was just planning to start programming a go playing program (I
had been writing only GUI's for gnugo, some web based GUI even, and game
browsing tools before) because someone finally had ported a few things
to python (if only to write up the algorithms in a decent language so
that I could later write my own low level libraries in C and then import
them into python) when I ran out of money for my personal budget so I
ended up without electricity and all my computers went dead. That was
half a year ago, and now I have hopes to be reconnected to the grid
sometime and start programming again.
> I work on trying to get my machines to learn the most trivial behaviors
> currently. It's what the entire field of machine learning does, because
> that's how far away from making AI we currently are.
I was in machine learning a long time ago. It still makes me sick to my
stomach to see over payed academic assholes issuing calls for papers
where they get free vacations while their research assistants stay
working late at the institute. And the same people that get the flying
trips also get ten times the pay. And you say this is about learning? I
say it is about keeping people stupid.
>>> You realize right that if you get the single human AI working, then all
>>> you have to do to create herd intelligence is build 100 of the same
>>> thing right?
>> This probably means you still do not take my theory for real.
>
> You haven't given us any theories yet to "take for real". You have only
> thrown out a handful of highly vague comments without actually telling us
> what your point is or what you believe is or why you are even here. Your
> writing style is very odd that way. You are very indirect as is if you
> were scared to commit to an idea and talk about it.
I'm interested in combinatorics, I use that to make nice pictures and
music and just for the fun of it. If it leads to AI, all the better, but
it's not what drives me.
> You are free to tell us, or not tell us, whatever you want to. But, are
> you for example, trying to solve AI by actually building machines? Or, do
> you just philosophy about what humans are, and what others how work in AI
> should be doing? That's a common problem here. :)
Don't go territorial on me just because you wrote more here recently. I
have been here longer and haven seen people come and go, including
myself for a few times.
> Do you think you have theories that would actually help us build machines
> that acted intelligent? Or do you just have words that you think have some
> high degree of self importance in reveling the truth nature of humans? I
> can't tell the answer to any of these questions yet based on how much you
> have written so far.
I thought going ad hominem was just reestablished here as a bad practice?
> You have made multiple references to computers which makes be think you are
> a software developer. If so, are you using your skills to actually try to
> create an intelligent machine? Ot just talking about it?
That is for you to decide, I just offer my view and I appreciate you
giving yours.
> This is the philosophy of AI so it is the place to just talk about, but if
> you are going to suggest _we_ should "take your theory for real" (meaning
> you think I should change the AI work I play with and my approach) just
> because you have given me some new insight I've never thought of, you will
> have to come up with something far more clever and useful than "humans are
> more intelligent in a herd" - which is all I can grasp you have said so
> far.
I suggest you not speak for the herd.
>> You'd never find the right single human AI that way.
>
> The aproach to AI that _everyone_ has used since the beginging is
>
> 1) theorize
> 2) build
> 3) test
> 4) repeat
Nope. What happens is they find something and then write a research
proposal trying to prove their result using statistic manipulations. But
statistics lie.
> I'm sorry, but that approach will _always_ find the right answer in the end
> no matter how you approach step one. That's because in step three, if your
> machines are not creating an intelligent herd, then you failed to get steps
> one and two right so you refine your approach and try again.
I agree it would be good if that approach would be followed,
unfortunately we'd all be dead of starvation or some other worldly
problem before we'd be advanced enough to make it work to our advantage.
> This technique will always converge on a good solution, if there's a good
> solution available to be converged on, that the we are able to understand.
Yes, yes, but we don't have the time.
>> You can't design
>> atmospheric molecules by some lucky break, so that it will result in it
>> snowing horizontally with very thin powder snow. It is something one
>> will never come up with if one only looks at the individual particles.
>> There are interactions that cannot be uniquely assigned to the
>> individual elements.
>
> That's just not true.
>
> You start with some "best guess", (step one), build it, and test it, see if
> it produces your "snowing horizontally with very thin powder snow" and when
> not, you itterativly improve you "best guess" to try and make it produce
> better results. You keep doing that until you get the right result.
No, you listen to what the other guys did because doing everything from
scratch would take one to cro magnon level during a lifetime, if one is
lucky.
> What you don't seem to grasp, is that the design choices are not random
> guesses. They use the data collected from real test results, to _improve_
> the last design. If improvements are not getting better results, then you
> thrown out the previous approach, and look for a new one.
But humans are allowed to run for help to their colleagues whereas your
poor robots aren't allowed to do that, and they are also not allowed to
ask their bosses (*you*) to change the problem into something they just
have found a solution for. You are expecting more from your machines
than what you'd expect from employees. Maybe your robots are female and
feel discriminated because that way they'll never be equal to you.
> It's a search guided by studying the results of your tests. And the
> results of those tests very much do allow you to quickly converge on the
> right answer by quickly eliminating large amount of "wrong" answer and
> "wrong" approaches.
Not eliminating, suppressing, until other data makes that viewpoint more
relevant.
>> Maybe a metaphor would help, so, at the risk of making things even more
>> unclear:
>>
>> If you'd generate all the permutations of some strings with repeated
>> elements, they wouldn't just be easily broken down into groups the way
>> one could do it if there were no repeats.
>
> Yes, you are being very unclear. You need to learn to communicate better.
> You are not helping the herd intelligence here. :) You clearly have a
> sharp mind that likes to think about things in deep ways (which is good),
> but to allow others to benefit from your thoughts, you have to find ways to
> communicate with others.
Hey, I was expecting you to immediately start writing algorithms that
generate permutations by index (quick, what's the n'th permutation
of some longish string with repeating elements, if the permutations are
ordered lexicographically and are all unique?)
>>> So what exactly am I'm ignoring by trying to build the thing I have to
>>> build 100 of to create herd intelligence?
>> Please don't take it the wrong way.
>
> We take each other the wrong way all the time time here. Don't worry about
> that.
Yeah, I just go away for a few years till things have cleared up. But I
liked Bloxy.
> Then develop it further by writing long posts here and letting people pick
> them apart. The more you share your ideas, and interact with the other
> people here, the more intelligent the distributed intelligence here
> becomes. The less you share, the less the distributed intelligence has the
> ability to grow.
What I do is sample the group from time to time, see if there aren't any
boring battles and then try to throw in some new material under the
grinders, if successful, I collect the dust and bake some bread.
> We just bull shit here. This is Usenet. This is not the scientific
> community. Very few of us here are actually part of the formal distributed
> academic community and their distributed intelligence efforts to solve AI.
> There have been various people here in the past that are well known parts
> of the academic efforts at solving AI, but there are less and less of them
> as time goes on. They don't have time to bull shit here, if they are
> actually trying to get paid for producing papers and teaching.
Ah, but the way I see it the academic world is producing the bullshit
while the actual work is done here, or even on 4chan.
> The subject of distributed intelligence is nothing new here. We have
> debated it constantly for years. Where as I am "Mr Reinforcement
> learning", Don is "Mr. Global Brain". - which is the point that our society
> (and all human organizations like clubs, and religious groups, and
> corporations, and countries) should be seen as more than just a society,
> but as a real itnelil9gence in it's own right, separate, and generally more
> intelligent, and more powerful, than any of it's members.
No, currently a group of people is more stupid than a single individual.
That doesn't preclude the individual consciousnesses being produced by
different groups though.
> The ideas of group intelligence you mention are nothing new. It's an old
> and well debated topic.
And that attitude problem will prevent you from seeing my point. After
all, you can just pick it up and throw it in with the other items of the
same category.
> I don't just think that distributed intelligence exists outside the brain,
> I think it also exists _inside_ the brain. I think individual intelligence
> is created by a distributed process of neurons interacting with each other
> in a society of neurons. I think the only way to solve the problems of
> scale in learning, is to attack the problem as a distributed society.
Yes, more humans means more hands to type away madly at the typewriter.
> Marvin Minsky who is one of the very well known names in AI, who used to
> post here on a regular basis many yeas back, wrote a very popular book
> called "Society of Mind" where he lays out his tries of how the brain is a
> collective society of agents working together to create the end result of
> human intelligence.
What happened, was he put off by the battle of the behaviorists years
ago? I somewhat believe the resulting bad atmosphere pushed a lot of
good people away. And this is the right Usenet group to discuss the
issues. Now all we have is this remnant and a few proprietary blogs. I
hate moderation of any kind, even if it rids us of some trolls the
damage it does is greater, because of mistaken uniformity, than even
newsgroups like comp.lang.lisp could inflict on a person. When I grow
up I will start to program lisp and post there.
> Your idea that we need to look at a society to explain intelligence is
> really nothing new. And so far, that's all you have said about what your
> idea is.
No, I said a lot more than that and you know it.
>> For example, very speculative, what if learning is a process of neurons
>> starting to fire together in some sequence but this is not limited to a
>> single brain anymore? The same algorithm that causes our brains to
>> select one reality over another (it decides to 'see' things in a certain
>> way) also makes it want to enlist other brains in this view and suppress
>> conflicting info even if that info is another human's brain. We just
>> have leaped over the individual human brain barrier. Now we know why we
>> communicate and do politics and religion.
>>
>> P.
>
> Yes, exactly. All well known idea. Nothing new yet. Keep trying.
For you?
> When we work together, and share a goal, we have created a larger
> distributed intelligence. If we interact with each other here, we form a
> distributed intelligence in the group with a goal of solving AI, or at
> least, of better understanding the problem of AI and better understanding
> ourselves.
>
> But again, you seem to want to learn to the idea that intelligence doesn't
> exist in a single brain - that it really only forms when you get multiple
> brains interacting. I just don't believe that's true in any sense. I
> think the intelligence starts at a far lower level, and by the time you get
> a single human, you already have a huge distributed intelligence at work
> inside that one humans. And when that one human interacts with other
> humans, the span of that distributed actions just grows even wider.
You have seen reports about feral kids, growing up amidst wild animals?
I suppose it would be hard to even talk to them. So we learn from other
humans, I just don't generally agree with the people taking credit for
it, because it is a collective effort.
> How neurons interact with each other to create the single intelligence is
> of course very different than how humans interact with each other. Neurons
> use pulse signals and the link, humans use language, and physical
> interaction, etc. And how the goals are specified and communicated are
> different. NO on fully understand how it works in the brain yet, but it no
> doubt makes heavy use of chemical messengers. Outside, we use tokens like
> money, and many other abstract techniques for communicating and sharing
> goals. But the result is still the same - that as, when the systems are
> able to work together in some aspect of a common goal, it's creating a
> greater intelligence.
Only that now the goal was switched for some other goal that was
achievable, while nobody was looking.
> One example I like as a thought demonstration of this greater intelligence
> is to think from the perspective of an alien watching the earth from a
> distance. They see a large asteroid flying towards the planet, and the
> plant reaches out (with missiles) and knocks the asteroid out of he way.
> That's an act of the earth's global intelligence doing at the level of the
> earth, the same the our brain does, when someone throws something at us,
> and we knock it out of the way with our arm.
>
> How the parts "inside the earth" work to make the earth intelligent like
> that, you can't see from a distance. But just watching the earth knock the
> asteroid out of the way, is a clear sign that the earth itself has a fairly
> high level of intelligence. Likewise, when we see a human knock a rock
> thrown at it out of the way, we can't see the small parts inside that make
> that happen, but we know the human has some high level of intelligence
> because of their ability to do that.
>
> Bot the earth as a whole, is already one large intelligence, and a single
> human, is one large intelligence.
>
> So now that we have multiple examples of intelligent systems, how do we
> make a computer intelligent? I believe we do it by building a
> reinforcement learning machine (or society if you would like) out of a
> society of interacting signal units that share a common reward signal that
> allows them all to play a small part in the bigger action. There is no end
> of different ways you could try to create such a system. I have a basic
> architecture I find attractive I've been playing with the past few years
> based on the concept of pulse sorting.
Humans not only have complicated problems, they also have many ways to
switch the problems for some other problem.
> We just keep searching by taking an educated guess, explore where it leads
> us, when it doesn't lead us to human level machine intelligence, we adjust
> and try again. Repeat until done.
I still disagree we do it that way, even if you claim it is nothing new
what I say. What we do is claim it isn't really AI if it does something
intelligent. Because it supposedly only works in a narrow domain, while
we humans are still superior because we can solve real world problems.
The only thing is, we can't either, when we have to operate under
unfamiliar circumstances we fail miserably, no matter how much we'd like
it to be true.
> I believe it's highly important to understand society is a higher
> intelligence. Not that it just acts intelligent, but that IT IS a higher
> intelligence which is separate from the individual intelligence of the
> humans (and machines) that make it up. You are preaching to the choir here
> with you ideas that social intelligence is important.
But in your desire to prematurely categorize me as soon as possible -- a
standard heuristic which I applaud if seen from your standpoint -- you
have mis categorized me as someone who thinks society is intelligent. It
is not intelligent yet, but it could be. Meanwhile, we humans derive our
mid level cluster starting points from that general 'reality'. We then
fill in the details and think all is well and that we have done it
ourselves. But we are wrong on both counts.
> BTW, in my "brain dump" work here by writing so much - it took me about 1
> 1/2 hours to think up, and write this message - just to give you an idea of
> how much I waste with this stuff. Now I've got to out out and shovel show.
> We had about 20" of crap dump on us yesterday - a hug amount of snow for
> this area (Washington DC metro area).
Yeah, it took a looong time to answer your post, on Mondays I have no
library access and almost all I have done today is answer your crap. An
they still haven't called for my reconnection to the electricity grid.
At least my laptop battery is full now.
P.
Curt Welch
> Curt Welch wrote:
>
> > Well, maybe you just don't know enough about my views yet because I
> > certainly don't think of academics as being the example of intelligence
> > we should be looking to duplicate. Not that there is not plenty of
> > intelligence there, just that's not how I work on AI.
>
> It's not just academics, but all people are having more or less the same
> intelligence level.
Well, it's all relative. There are some fairly wide extremes on both ends
with people that stand out with mental powers far above and far below the
average (some times both in the same person). But yes, as a whole, the
bulk of humanity has a fairly similar set of skills.
> Like cars, there are some that accelerate from zero
> to a hundred kilometers per hour in 5 seconds, and there are those that
> take a bit longer, but they are all cars, just vessels. In the coming
> era we will have more or less well equipped avatars
Huh? What type of avatars are you talking about? Our bodies, or something
Sci-fi like remote controlled robot bodies? Or some virtual avatars on
computer networks?
> and it will be gross
> to discriminate against someone only because her vessel or avatar is
> operating in low resolution or maybe just in black and white.
Well, not sure how these issues of discrimination plays into this debate
about how to solve AI.
You do realize that AI is a field of computer science right? It's a debate
about software development at it's core. Though since this group is the
philosophy of AI, we often wander far off track in debate what exactly the
I in AI is.
> What I am arguing against is people trying to use intelligence claims as
> a status tool.
Ok, that's fine. But again, not sure how that relates to the question of
how we create machine intelligence?
Though I'm all for working to tone down discrimination in society to make
it a nicer place for everyone, it's also going to continue to be a fact of
life forever. That is, life is not fair, and life is ultimately, a fight
for survival. It's not something we get to change just because we don't
like the game we born into. The forces of evolution are the high level
forces that created us, and there is nothing you can do to remove that high
level force.
If you choose to act as if there was no need for survival - that being nice
was more important than being alive, you and those like you will die, and
those that struggled to survive, will be the ones that live, so your effort
to change society will fail in the long run. The society that survives
into the future will always be full of those that struggle to survive.
And when you struggle against any force, there will _always_ be some people
that mange to do better than others. And the people that do better, will
always have a naturally high status in the eyes of the rest of us. It's a
discrimination you can't remove from society. The Tiger Woods of the world
will always (by luck, fait, or hard work), will always have lots of money,
and power, lots of beautify woman willing to sleep with them. The people
that manage to gain such positions will always be favored, and the rest of
us, will always end up being discriminated against.
Whether it's skill with a golf club, or skill creating and running
businesses, or skill as a comedian, or musical skills, or math skills, or
computer programming skills, the ones that stand out will always have a
higher status in society, and the rest will just have to cope.
> This can happen in the seemingly most open communities.
> For example, people arguing in favor of ending all animal suffering,
> elevating animals to a higher level, but at the same time making
> inappropriate distinctions between humans based on tiny intelligence
> differences (surely, if seen from an interspecies level viewpoint) which
> are for the most part fabricated because of status drive. I believe it
> is more prevalent in younger people, with higher testosterone levels
> making them want to appear bigger than they are (physically and
> intellectually) so as to increase their chances of finding a mate.
Well, calling someone dumb when they are dumb, is not discrimination. It's
just the truth. Assuming someone is dumb, (or smart) based on their name,
or size, or skin color, that's unfair discrimination. But it's also,
again, a fact of life. That is, it's how our brain works. It's hard to
ignore which is why it's hard not to unfairly discriminate. We use all the
environmentalism clues available to us though our senses to estimate the
future. So when we are trying to estimate how someone we have just met
will act in the future (such as how they will perform in a job, or as a
member of some group), we use all the clues available to us, including crap
like their height, and weight, and eye color, and hair color, and teeth,
and their voice, and skin color, and on and one. All these factors effect
our judgment whether we want them to or not. We have to work hard to not
overly trust our instincts because they are so full of personal bias in
important issues of creating a fair society.
> Of course there are also those who do it for acquiring more resources.
Everyone attempts to acquire more resources. Whether it's a large bank
account and retirement account for your personal safety, or a large number
of friends to make you feel better, or a more knowledge, everyone is always
collecting and managing resources. It's what we do to survive.
Some people go overboard on a status symbol competition, but again, this
stuff is a natural part of life. You can't escape it.
> But the whole point is that all our culture is the result of a
> collective effort, where for the most part the people with the ideas are
> not the ones who get the credit. Instead, the credit goes to their
> supervisors, or to the music industry if they are musicians, or to the
> shrewd businessmen exploiting someone else's ideas.
I don't know if there's really enough evidence to conclude that. Many
people with good ideas do get credit for their ideas.
It's true that good ideas are always the work of standing on the shoulders
of giants, and it's true that often the same good idea is found by multiple
people, and only one got credit for it in the history books (often not the
first to find it). But to suggest that it's the norm that the people with
the idea never get credit for it is just unjustified in my view.
> The sincere innocence with which people take ownership of the ideas in
> their heads is often just self serving stupidity. They think, 'because I
> am the one who thought of this first, it is *my* idea'. Not so though,
> because a lot of times ideas are adaptations of ideas that other people
> had earlier.
Adaptations of old ideas are new ideas. That's what most "ideas" are.
We call it "our idea" because it's how we get credit in society for doing
something useful to help society. There's nothing wrong with any of this.
> Then there are people who accuse other people of stealing
> when they are stealing themselves and their self importance doesn't
> allow them to look back at the 2000 year old sources of their ideas.
Well, again, I don't really know where you are going with all this. Do you
think someone where (me?) is calming "new idea" when nothing I write is a
new idea? Or that someone here is claiming that someone else stole their
ideas.
In the scientific community, people's careers live and die by the ideas
their create (or fail to create). In that community, there are well
established rules of competition and judgment to determine the worth of the
ideas generated and to determine who gets "credit" for the ideas. The
"game" is played by the publication of papers and the score is kept by a
rough weighted measure of the citation links. If you have lots of papers,
that are often sited by lots of the the other most cited papers, your
status in the scientific community is raised, and the odds of you getting
grant money goes up. As always, it's an imperfect system, but it mostly
works. The people that have the most of the best ideas in general get more
money.
> In my opinion it is a convoluted mess and therefore it's very hard to
> attribute ideas to persons even when the specific person is yourself and
> the idea is a new one in your own head that you are sure you never heard
> about before.
Yes, it's hard. But the formal system of publishing your ideas in research
papers makes easy to track down who actually said what when retroactively
if you wanted to try and actually understand the formation of ideas over
time.
> Moreover, the whole attribution thing is counterproductive because the
> mentality fits in a world view we (humanity) are trying to leave behind
> us, the idea that it is shameful not to work and that greed is good.
It will always be shameful not to work. That's just a direct fall out from
the high level forces of evolution that created us. Society will always
look down on the members that are unproductive - that are not helping to
make society better.
Even when we get to the point that we can't work at what is today a
"normal" job because the machines are doing all that work for us, there
will evolve some odd and complex set of "tasks" for humans to work on to
make live better for themselves, and other humans. Like maybe we will be
judged by the type of parties we throw, or the type and number of presents
we create and give to others. Or maybe human occupation will become art -
and everyone will be judged by the quality of the art they produce? Or
maybe it will be how hard we work at making sure the machines are doing the
best job possible for us? Whatever it turns out to be, it will be still be
seen as important work for the humans to make society better and society as
a whole will always look down on the people that fail to be productive
members of society.
It's up to you whether you want to ignore that pressure and choose to be a
bum, but the pressure will always be there.
> Instead, we as humanity are migrating towards the idea that people work
> just because they like it,
As we become more powerful as a society, and as the threats to our survival
die down, we become lazy. The more we can control the threats, the more
lazy we will become. But we will _never_ give up the pressure to work and
become productive. There are always threats to our survival (starvation,
disease, etc) that will always keep us honest. The cover of living will
never drop to zero because the threats to our lives will never drop to zero
which means we will always have to keep working to survive - which means we
can never move, as a society, to "letting only those that want to work
work". It just won't happen.
> and that job hunting has bad consequences if
> it means increased social inequality because a master, or owner, class
> is exploiting the basic physical needs of a lower class by controlling
> the job market by means of owning the production tools. There are signs
> this is all going to change, solar cells are a one time purchase,
> freeing one from the energy overlords, the rep rap project promises
> local manufacturing, and in general there is a trend away from centrally
> controlled resource markets.
You're funny. :) That's just nonsense.
If you want freedom from society, you can have it today. No need to wait
for solar energy or high tech rapid prototyping to kick in. Just go back
to what people did when they were forced to live on their own off the land.
Use plants to collect solar energy for you and burn the trees to get the
sun energy when you need it. Learn to make you own metal and become a
blacksmith and use those same trees to turn iron ore into metal farming and
construction objects like plows and axes using a hammer and a fire. Use
those tools to "rapid prototype" yourself a house and a barn and some
fences to hold your live stock.
We are not moving _towads_ such times. We are moving _away_ from them.
Life is far easier for us when we work together as a large coordinated
society instead of when we try to live on our own. It's why such
independent life styles are becoming a think of the past. We don't live
off the land anymore, we live off the tit of society by finding ourselves a
job.
Anyone that wants to go back to living alone, or living in a small group
(like a commune) are free to do so. They never, in general, have as easy a
life as those of us dealing with the complexity of living as a small part
of a huge society.
> > My approach to AI is a bottom up one. Some like to attack the problem
> > from the top down. In my direction, I try to figure out how to make
> > very simple signal processing systems perform highly simple acts of
> > intelligence. I design and build various neural networks and try to
> > figure out how a neural network should be built to make the entire
> > network act intelligent - which INCLUDES making multiple networks
> > interact intelligently.
>
> Sorry but I still suspect you are too proud to look at the statistical
> properties of the data
The statistical properties of WHAT data????
And yes, I'm a fairly arrogant prod selfish ass hole at times. But what
does that have to do with AI? I can't tell if you are trying to talk about
my short comings, or whether you are talking about my ideas? They aren't
the same you know.
It's hard most the time to grasp where your ideas are going. To grasp
what point you are trying to make.
> because of some ill advised idea that your
> algorithm, if it is any good, will fend for itself. This is exactly the
> wrong kind of approaching the problem. AI is a 'win anyway you can'
> problem not some aesthetic sword duel between gentlemen.
What on earth do you mean by "win anyway you can"???
The human brain performs some very specific functions andin AI we are
attempting to make computers perform the same functions. There is no "win
anyway you can" about it. Either you have duplicated those functions or
you haven't.
So far, all anyone has done, is duplicated small parts of what the brain
can do. No one has made any great progress on duplicating large parts of
what the brain can do. Here in the philosophy of AI, we mostly debate how
to approach the problem of making machines duplicate most the function of
the brain.
I can't tell where your points fit into this. Are you saying that am
Fine. But do you have an opinion of how it does work, or is "I don't
agree" the limit of your idea?
> even if you claim it is nothing new
> what I say. What we do is claim it isn't really AI if it does something
> intelligent. Because it supposedly only works in a narrow domain, while
> we humans are still superior because we can solve real world problems.
> The only thing is, we can't either, when we have to operate under
> unfamiliar circumstances we fail miserably, no matter how much we'd like
> it to be true.
Well, some people believe that full AI has to be created by using a lot of
narrow-AI projects glued together. I don't think that's even possible, let
alone practical. Most projects are so different they just can't be glued
together in any meaningful way. Solving how different programs like a
chess program and a 2D navigation program can be used together is as hard
as solving the entire AI problem in my view.
> > I believe it's highly important to understand society is a higher
> > intelligence. Not that it just acts intelligent, but that IT IS a
> > higher intelligence which is separate from the individual intelligence
> > of the humans (and machines) that make it up. You are preaching to the
> > choir here with you ideas that social intelligence is important.
>
> But in your desire to prematurely categorize me as soon as possible
Only because I'm trying to grasp what the point of your messages are! :)
Maybe that's my problem. Maybe you are just talking and have no point or
goal or direction?
> -- a
> standard heuristic which I applaud if seen from your standpoint -- you
> have mis categorized me as someone who thinks society is intelligent. It
> is not intelligent yet, but it could be.
So there is no intelligence in the individual or society???
Again, we normally talk about how to duplicate human behavior in a machine
here. Whether you want to call it intelligence or not is just word games.
Whatever humans do, that is what we are trying to duplicate in our
computers (or other man made machines). Do you have _any_ ideas on what it
is humans DO, or how to make a computer do it? Or do you just have lots of
idea about what everyone else is doing wrong and what everyone else fails
to understand without you actually telling us what's right or how to solve
the problem of AI?
It's easy to walk in and claim everyone is wrong, it's much harder to
actually walk in, and tell us what's right.
> Meanwhile, we humans derive our
> mid level cluster starting points from that general 'reality'. We then
> fill in the details and think all is well and that we have done it
> ourselves. But we are wrong on both counts.
Talking and thinking in the abstract is good. But if you don't every once
in a while, tie your abstract thoughts down to reality so we know what you
are talking about, your words will mean almost nothing to us.
> > BTW, in my "brain dump" work here by writing so much - it took me about
> > 1 1/2 hours to think up, and write this message - just to give you an
> > idea of how much I waste with this stuff. Now I've got to out out and
> > shovel show. We had about 20" of crap dump on us yesterday - a huge
> > amount of snow for this area (Washington DC metro area).
>
> Yeah, it took a looong time to answer your post, on Mondays I have no
> library access and almost all I have done today is answer your crap. An
> they still haven't called for my reconnection to the electricity grid.
> At least my laptop battery is full now.
>
> P.
You of course don't have to answer my post. You are free to ignore it.
Most do. If there's a point you want to address, do so, and just ignore
the rest. Just because I ask a question in a post doesn't mean I expect it
always to be answered. I won't think less of you because you failed to
spend hours addressing silly points I've wasted a few hours making. Social
conventions on Usenet are somewhat different than normal face to face
social conventions.
casey
> ...
> Of course there are also those who do it for acquiring
> more resources.
It is all about acquiring resources. We come into the world,
compete by means fair and foul to get as much as we can, and
then we leave. This is true of all animals. It is all about
maximizing reproductive success.
JC
Yevgen Barsukov
You are missing here that humans propagate not only genetic
programming, but also cultural programming. And success in which of
this two is more important for our decision making
is highly debatable. We can look at two extremes:
A) purely instinct driven animal, which is optimized
to propagate his genes
OR
B) purely cultural programing driven animal. Basically monkey
infested with cultural virus that cares not at all about propagating
of monkey genes, but just about propagating itself
onto other monkeys.
Real humans are somewhere between A and B and there are probably many
variations with different weights. In overall during history of human
civilization there have been a gradual drift from A to B.
There have been some critical breaking points, where people
would clearly sacrifice the monkey for propagating the cultural
program. Being a Christmas time, it is proper to mention for example
J. Christ.
Such events have been duly noted and caused further acceleration in A
to B drift.
At this point it went so far that we can start focusing mostly on A,
and furthermore discuss how compatible is the cultural programing with
different hardware, that are not monkeys.
Some of such "cross-species" propagation already started. The example
I like are some part of cultural programing is Symbolic mathematics,
such as Maple or Mathematica. It is complex logic (for example how to
solve analytically systems
of differential equations) that can be executed on silicon hardware
the same way as it was before executed on human brains.
So far it is just a small subset of overall cultural programming
that has been "ported" to different hardware, but I don't see any
limit to this process. Basically all parts of cultural programming
that will be eventually ported, as long as it will be cost effective
(e.g. silicon will cost less then monkeys) and energy efficient. Note
that propagation into
silicon rather than monkeys is not going to happen just for the
fun of it, but only if it will help future propagation.
It is just like water will more through thick pipe
than through thin pipe, propagation of cultural programing
will be going more into hardware that is more useful for further
propagation. As long as this hardware is monkey brain, we will see a
lot of intelligent "programmed monkeys" e.g. people. If silicon will
become more suitable for future propagation, we will see more
"programmed silicon".
Regards,
Yevgen
pataphor
[...]
> We can look at two extremes:
>
> A) purely instinct driven animal, which is optimized
> to propagate his genes
>
> OR
>
> B) purely cultural programing driven animal. Basically monkey
> infested with cultural virus that cares not at all about propagating
> of monkey genes, but just about propagating itself
> onto other monkeys.
>
> Real humans are somewhere between A and B and there are probably many
> variations with different weights. In overall during history of human
> civilization there have been a gradual drift from A to B.
A very interesting view. And on the surface it seems highly credible.
But what I am missing here is a sense of identity, or self. It is like
this view is completely from the outside and not giving us any meaning
or personal space. That's why it scares the hell out of me, however
logical and credible it sounds. Are you sure you are not missing some
third component you forgot to include in your model?
P.
Yevgen Barsukov
It only sounds scary if we assume that sense of identity can only be
associated with
the animal genetic programming, but not with the cultural programming.
If we look closely at this assumption and see what are its
justifications,
it will stop looking that convincing.
Lets start with sense of identity in an animal. How complex an animal
has to
be to have this function? Does it have to be a mamal, say a social
animal like a wolf?
Or is it an amoebae that needs to keep multiple conflicting functions
in optimal compromise state in order to improve survival chances?
Lets try to start with the simplest cell imaginable. It has to have
energy to function,
so it has to use whatever changes it can make to its operation to
maximize amount
of energy available while minimizing amount of energy used. Lets say
it has a sensor protein that causes its flagella to spin faster and
move it forward if sugar concentration is increasing (function A), and
it has another sensor that detect that internal resources become
depleted (function B). Obviously function B will try to stop flagella
from spinning to conserve resources.
How to resolve conflict between A and B?
There has to be some program of arbitration, that assigns weights for
relative importance of conflicting actions and ultimately decides what
will happen. In this simple case arbitration can be handled in a
centralized manner, say be a single molecule.
This molecule will be the "sense of self" for amoebae.
What if a more complex animal, say a fox, has 100000000 various
conflicting genetically pre-programmed priorities, each one driven by
a separate program? Arbitration between all these programs becomes
extremely complex computational problem that itself will require a lot
of computations. Considering extremely low frequency of animal
computing systems (just 10 Hz neurons firing rate), it is clear that
such huge amount of computations can not be run in just one
"processor", it has to be a distributed parallel computations.
Clearly that could not possibly be controlled hierarchically,
by just one processor connected to all of them - communication speed
would not be enough. They only way how such system can be convergent
is by "voting" of individual subunits with large amount of
independence. The final vote of all of parallel sub-processes will
decide which of the conflicting actions will be taken. This state of
agreement of multiple sub-processes will be the sense of identity for
complex animal.
How about arbitration between multiple processes that are not
genetically programmed, but programmed from external sources (say from
100 000 years of transmission between generations of an animal)? These
system will have exactly the same problem as system with genetically
programmed decision making processes. In fact there is no difference
whatsoever if it is genetically or "externally" programmed - since is
it is running on a very similar neural
hardware, arbitration is most likely to work in a very similar way and
reach a common ground
in a way of "voting" of individual computing subunits. The state of
consensus of all computing subunits will be available as a signal to
the executing units, and this signal is the "sense of self".
Note that it contains contribution from both genetically programmed
and culturally programmed decision making units. I venture to assume
that later dominate now, so your sense of identity is more that of
cultural virus rather than that of a monkey. If so, it is fair to say
to "I" am really the eternal and death defying cultural program.
Can the same mechanism work also on non-neural hardware bases (say on
silicon computer)?
I don't see why not. It has not been implemented yet because all
cultural programming that has been so far been ported to silicon did
not have a function of decision making. It all had function of "pre-
processing" information for decision making of humans. So existing
computers were artificially made as part of human arbitration process
but do not contain or need arbitration ability in themselves.
This is acceptable now as processing speed in silicon computer is
comparable (at least for parallel processing) to that of monkey
brains, so slow "silicon-human" communication link does not hinder
overall efficiency too much so it is worth it to communication only
the intermediate results and do final processing that leads to
decision in monkey brain.
But at some point this silicon-monkey communication link might become
slowest part of the system and it will make sense to get rid of it and
implement arbitration between conflicting priorities (e.g. sense of
self) and final decision making inside silicon itself.
Regards,
Yevgen
Curt Welch
Well, for my 2 cents on this, I agree that we are the result of both a)
nature, and b) nurture. But to suggest that the way we change in response
to nurture is even mostly under the control of the "will of the cultural
virus" is a huge failure to understand how the propagation of memes works.
The genes have total control over what memes it will allow the body to
except. The memes have very little control over the genes (currently).
There is no doubt that currently, the genes are in control here - not the
memes. The memes are emergent behaviors that the genes allow to happen.
When the genes create emergent meme behaviors in our society, and those
memes fail to do a good job of keeping the genes alive, then those genes
(and the memes they allows to emerge) all die off together.
Humans can't be forced to accept any meme. They only accept the memes that
are proven to useful to prevent harm to the body or harm to our
reproductive success - which is a requirement created by the genes. They
are the gate keepers of what memes are allowed to take hold in us and
always have the upper hand.
We don't accept just any meme we hear from society. Our brain judges the
value of the different memes it is exposed to, and accepts the ones it
evaluates as being the best - the most useful. But the evaluation system
the brain uses, is specified by our genes. Which is why the genes have
ultimate control over the memes - not the other way around in any sense.
Learning systems are flexible in that they have the power to learn a wide
range of very odd behaviors - like swinging a stick at a little white ball
in order to maximize our odds of survival (Tiger Woods). But they are not
so flexible that they will learn just anything they are exposed to. They
don't work like a camera just capturing and duplicating _anything_ it sees.
It's nothing like that. All changes to our behavior happen for very
specific reasons - because the change is shown by the internal behavior
worth estimation system to be better than the previous behavior.
Looking at the odd things humans do, you might not understand how on earth
some behaviors could exist. But they all tie back to the experience that
person has had in their life relative to their internal drives defined by
their genes.
Learned behavior are just as instinctive as fixed behaviors. They are just
more complex. That is, the instinctive behavior of learning is operant
conditioning.
With a simple instinct, we might sense heat, and move away from it. The
hardware is pre-build to always perform that action to that type of
stimulus. With operant conditioning, we see some fire, move our hand
towards it, get burned, and then the internal behavior system adjusts the
map so we are less likely to produce that same behavior in the future.
Though the "move away from heat" is pre-wired in one case, the "adjust our
behavior to prevent sensing too much heat" is innate in the other. It's an
innate behavior created by the genes in both cases. The first is just
stimulus response, the second is stimulus, response, evaluation, stimulus,
different response. The sequence is innate and unchangeable in both and
fully specified by the genes - who always have the upper hand (for now).
As long as humans continue to be grown as they are now, under the control
of our genes, the genes will continue to have the upper hand here. But the
more we use our intelligence, to change our bodies, and our genes, the more
complex the entire process of evolution becomes and the less we will be
able to say the genes are in control.
However, the upper hand of evolution will never lose control. That is,
whatever form we evolve to, it will only exist, if it proves to be good at
survival. At the moment, our genes are still the prime controller of our
evolution, and the memes really control almost nothing - they were created
by and allowed to exist, by our genes.
Curt Welch
That's a stretch. :)
Yes, your example is a control function that regulates the behavior of the
system, but to call that a "sense of self" is very odd in my view.
Any control system that can sense something about the condition of the
system, instead of sensing only external conditions can be called a "sense
of self"
> What if a more complex animal, say a fox, has 100000000 various
> conflicting genetically pre-programmed priorities, each one driven by
> a separate program? Arbitration between all these programs becomes
> extremely complex computational problem that itself will require a lot
> of computations. Considering extremely low frequency of animal
> computing systems (just 10 Hz neurons firing rate), it is clear that
> such huge amount of computations can not be run in just one
> "processor", it has to be a distributed parallel computations.
> Clearly that could not possibly be controlled hierarchically,
> by just one processor connected to all of them - communication speed
> would not be enough. They only way how such system can be convergent
> is by "voting" of individual subunits with large amount of
> independence. The final vote of all of parallel sub-processes will
> decide which of the conflicting actions will be taken. This state of
> agreement of multiple sub-processes will be the sense of identity for
> complex animal.
Well, trying to define "sense of identity" that way is again, very odd.
> How about arbitration between multiple processes that are not
> genetically programmed, but programmed from external sources (say from
> 100 000 years of transmission between generations of an animal)? These
> system will have exactly the same problem as system with genetically
> programmed decision making processes. In fact there is no difference
> whatsoever if it is genetically or "externally" programmed - since is
> it is running on a very similar neural
> hardware, arbitration is most likely to work in a very similar way and
> reach a common ground
> in a way of "voting" of individual computing subunits. The state of
> consensus of all computing subunits will be available as a signal to
> the executing units, and this signal is the "sense of self".
>
> Note that it contains contribution from both genetically programmed
> and culturally programmed decision making units. I venture to assume
> that later dominate now, so your sense of identity is more that of
> cultural virus rather than that of a monkey. If so, it is fair to say
> to "I" am really the eternal and death defying cultural program.
>
> Can the same mechanism work also on non-neural hardware bases (say on
> silicon computer)?
> I don't see why not. It has not been implemented yet because all
> cultural programming that has been so far been ported to silicon did
> not have a function of decision making.
Huh? Every robot controller ever built had decision making as its prime
function.
> It all had function of "pre-
> processing" information for decision making of humans.
Well, yes, lots of IT systems work that way. But still, to suggest that
they don't make their own decisions is like trying to pretend no computers
or programing languages implement IF statements.
> So existing
> computers were artificially made as part of human arbitration process
> but do not contain or need arbitration ability in themselves.
That's just not true. Every one of them is full of self arbitration
ability.
We have created and selected the arbitration systems they use so as to
maximize our own needs - but they still have their own arbitration systems,
just like we do.
And likewise, our arbitration systems were selected by forces beyond our
control - the forces of natural selection. But just because some other
system created/selected them, does mean we don't have the power to make our
own decisions.
> This is acceptable now as processing speed in silicon computer is
> comparable (at least for parallel processing) to that of monkey
> brains, so slow "silicon-human" communication link does not hinder
> overall efficiency too much so it is worth it to communication only
> the intermediate results and do final processing that leads to
> decision in monkey brain.
>
> But at some point this silicon-monkey communication link might become
> slowest part of the system and it will make sense to get rid of it and
> implement arbitration between conflicting priorities (e.g. sense of
> self) and final decision making inside silicon itself.
>
> Regards,
> Yevgek
Well, I actually agree with most everything you said above, except your
very odd definition of what "sense of self" means. Pataphor brought up the
idea of a "sense of self" seeming to be missing, and I have no real idea
what he things that means. However, when humans talk like that, it's often
a second hand reference to the illusion of consciousness - that is, the
illusion that humans seem to sense a soul-like think existing in them which
is somehow separate from their body. That aspect of a sense of self
certainly does not exist in lower animals - it's a complex effect created
by a more complex adaptive learning system. Does it exist in wolfs?
Maybe? But probably not. It's probably not justified to say it exists
unless the animal has the ability to talk about the concept.
The only real sense of self that exists separate from the illusion is the
ability of a system to draw the line between self and the rest of the
environment. But that's not all that special or hard to understand.
What Pataphor might have been making a reference to is just the fact that
in our normal langauge we use a different set of words to describe the
actions of humans than we do to describe the actions of all the non-human
stuff and in doing so, it creates a large default implication that humans
are special (conscious feeling beings with a soul created as the Son of God
type of special) vs the rocks and trees which are just "things without a
soul". If you use the non-human descriptive words to talk about humans, it
takes away from us that illusion of being unique and special - which if
course it should because we aren't any more special than the rocks or the
trees.
casey
> On Dec 22, 2:14 pm, casey <jgkjca...@yahoo.com.au> wrote:
>> It is all about acquiring resources. We come into the world,
>> compete by means fair and foul to get as much as we can, and
>> then we leave. This is true of all animals. It is all about
>> maximizing reproductive success.
>
>
> programming, but also cultural programming. And success in
> which of this two is more important for our decision making
> is highly debatable. We can look at two extremes:
>
>
> A) purely instinct driven animal, which is optimized to
> propagate his genes
>
> OR
>
> B) purely cultural programing driven animal.
Cultural programming can only take hold only to the extent that
the individual finds a particular program (meme) rewarding and
our primary rewards are genetically determined.
This has made me think again about what makes humans different
from most other animals? Perhaps it is our ability to imitate
and understand the actions of others rather than learn by
trial and error alone. Without a human social environment no
amount of trial and error as suggested by Curt will make an
individual show much intelligence. It is the accumulated trial
and errors of many people over many generations that has given
the individual, with the ability to imitate, to acquire human
level intelligence in a lifetime.
As Richard Dawkins wrote in his book "Unweaving the Rainbow",
"Memes could not spread but for the biologically valuable
tendency of individuals to imitate.".. "Individuals that are
genetically predisposed to imitate enjoy a fast track to skills
that may have taken others a long time to build up."
I take "genetically predisposed" to be innate and "skills"
to be what we call intelligence when executed. This would make
Curt's generic learning in humans to be the ability to imitate.
JC
Curt Welch
> On Dec 23, 10:07=A0am, Yevgen Barsukov <evgen...@gmail.com> wrote:
It's certainly very true that a huge amount of our current abilities came
to us from our society. So our ability to learn from others is a key
feature of our intelligence. If we couldn't learn from others so easily,
we would be fairly stupid individuals by comparison.
What I don't agree with however, is that we need to add yet another
hardware module to explain this skill. One module, that implements strong
operant conditioning, explains all our abilities, including our ability to
mimic. You don't need to add modules for 10 types of learning and 5 types
of memory just because you can't grasp that they are all just different
ways to describe the emergent behaviors of our operant conditioning module.
casey
> casey <jgkjca...@yahoo.com.au> wrote:
Newsgroups: comp.ai.philosophy
From: c...@kcwc.com (Curt Welch)
Date: 24 Dec 2009 23:10:04 GMT
Local: Fri, Dec 25 2009 10:10 am
Subject: Re: No easy road to AI
casey <jgkjca...@yahoo.com.au> wrote:
> On Dec 24, 9:27=A0am, c...@kcwc.com (Curt Welch) wrote:
> > ...
> > Most notably, they will be conscious, and they will hate being
> > hurt as much as we do.
>> Some people don't feel pain and yet they are conscious.
>
>
> Cool, what does that mean? What exactly are those people not feeling?
http://www.cnn.com/2006/HEALTH/conditions/01/27/rare.conditions/index.html
Newsgroups: comp.ai.philosophy
From: c...@kcwc.com (Curt Welch)
Date: 25 Dec 2009 00:22:30 GMT
Local: Fri, Dec 25 2009 11:22 am
Subject: Re: New kid on the block
> One module, that implements strong operant conditioning,
> explains all our abilities, including our ability to mimic.
Operant conditioning the theory of ALL brain behaviours!
So what is different about our module compared with the
modules of animals with less or no ability to mimic?
You say operant conditioning explains all our behaviors
how about some memory behaviors?
If you are able to explain how various manipulations of the
inputs to the human brain produce various kinds of memory
behaviors due to an operant conditioning mechanism that would
be interesting.
One input for example was a glowing light in a dark room
attached to the rim of a rotating wheel. When you increase
the rotational rate of the wheel starting at zero a speed
can be found where you see a complete circle. It turns out
that it needs to make a complete turn every 1/10 of a second
for a complete circle to be seen.
How do you explain this emergent behavior of 1/10 of a second
in terms of operant conditioning?
If you read out a series of digits in chunks (pause between
groups of digits) they will be easier to recall than if they
are read out at a fixed rate for each digit.
---> time
820938947108
820 938 947 108
8209 3894 7108
How does operant conditioning predict this memory behavior?
It has been found that the faster a person can speak a set
of words the more words they could hold in memory. It turns
out that the subjects could only remember as much as they
could say in 1.5 seconds. Why does this time limit emerge
from operant conditioning?
You say the brain is one big single module. Ok why does
disabling one of its components, the hippocampus, result
in the inability to remember what it did five minutes ago
but still enables it to remember new skills? How does this
emerge as a result of operant conditioning?
JC
Curt Welch
> On Dec 25, 11:22=A0am, c...@kcwc.com (Curt Welch) wrote:
> > casey <jgkjca...@yahoo.com.au> wrote:
> Newsgroups: comp.ai.philosophy
> From: c...@kcwc.com (Curt Welch)
> Date: 24 Dec 2009 23:10:04 GMT
> Local: Fri, Dec 25 2009 10:10 am
> Subject: Re: No easy road to AI
>
> casey <jgkjca...@yahoo.com.au> wrote:
> > > ...
> > > Most notably, they will be conscious, and they will hate being
> > > hurt as much as we do.
>
> >> Some people don't feel pain and yet they are conscious.
> >
> >
> > Cool, what does that mean? What exactly are those people not feeling?
>
> http://www.cnn.com/2006/HEALTH/conditions/01/27/rare.conditions/index.htm
> l
That's neat. Not for them of course, but in what it might revel about the
brain.
> Newsgroups: comp.ai.philosophy
> From: c...@kcwc.com (Curt Welch)
> Date: 25 Dec 2009 00:22:30 GMT
> Local: Fri, Dec 25 2009 11:22 am
> Subject: Re: New kid on the block
>
> > One module, that implements strong operant conditioning,
> > explains all our abilities, including our ability to mimic.
>
> Operant conditioning the theory of ALL brain behaviours!
No John it never was and never will be a theory of all _brain_ behavior.
It's a theory of all _intelligent_ behavior.
> So what is different about our module compared with the
> modules of animals with less or no ability to mimic?
We, I've explained this to you before in private emails about this same
subject, but let me try again....
An operant conditioning system must have an inherent power of perception.
If it can't sense a stimulus, it can't learn how to respond to it. No
sensory system can sense everything, or anything. If it doesn't have eyes,
it can't sense light. But after it gets the sensory data to work with, it
also needs a pattern decoding system to recognize the required patterns. If
it doesn't have the right hardware to "see" the needed abstract patterns,
it can't learn by operant conditioning to respond to it.
What patterns do we need to recognize in order to _learn_ to mimic? We
have to recognize a lot of things. We have to recognize the person's
behavior. We have to recognize the result it produces so we can judge the
value of their behavior, and we have to recognize our own behavior. And,
we have to have the correct abstractions in our perception system to
recognize the similarities between their actions and our own so we can see
them as being the same type of action. Once you have all that ability in
the perception system, then you can _learn_ to mimic.
The limit of what you can learn to mimic by operant conditioning, is the
limit of what you can perceive is there to be mimicked. If someone opens a
door, and gets food, we can mimic the behavior by opening the door. But
the high level abstract concept of "open the door" requires a very large
and high quality perception system.
You have given the exmaple MANY TIMES of the ape which could not mimic the
action of washing dishes, but could mimic the action of wiping the plate
with a wet cloth. clearly, the ape could mimic. But what was missing?
Very simple. His percpetion system didn't have enough resolution to fully
abstract out the result of the actions - clean plates. The action of
moving his arms to wipe the plate he could "see" and mimic but the action
of "removing all the stuff from the plate" he could not see. He didn't
have a strong enough perception system to abstract out that feature of the
action. And if he can't see the difference, he can't mimic it.
Lower animals that have operant conditioning systems obviously don't have a
strong enough perception system to recognize the required features of the
environment so as to learn the value of the "mimic" behaviors.
> You say operant conditioning explains all our behaviors
> how about some memory behaviors?
Yeah, pretty much. Some of the detailed memory tests we can do on humans
are probably exposing more of the implementation details of the brain than
they are exposing the specifics of how we produce intelligent actions.
> If you are able to explain how various manipulations of the
> inputs to the human brain produce various kinds of memory
> behaviors due to an operant conditioning mechanism that would
> be interesting.
I strongly suspect Skinner already did that 60 years ago.
> One input for example was a glowing light in a dark room
> attached to the rim of a rotating wheel. When you increase
> the rotational rate of the wheel starting at zero a speed
> can be found where you see a complete circle. It turns out
> that it needs to make a complete turn every 1/10 of a second
> for a complete circle to be seen.
>
> How do you explain this emergent behavior of 1/10 of a second
> in terms of operant conditioning?
It's got nothing to do with operant conditioning. You are simply testing
the limits of the perception system which has nothing to do with operant
conditioning. Why isn't this obvious to you?
If you take a picture of a circle with a digital camera, what do you get in
the image? If the circle is large enough, the pixel pattern indicates a
circle and you could build a pattern detector that would indicate "circle"
when it "sees" that pattern of pixels. But when you move back from the
circle, it becomes smaller. And at some point, it stops being a black
circle with a white inside and simple becomes a black dot - because it
reduces to the size of 4 pixels.
At that point, the perception system could no long correctly identify it as
"circle" and instead would identify it as "black dot". How do you explain
that the circle starts to look like a dot by operant conditioning? You
don't. You explain it by simple basic trivial to understand limits of the
sensory and perception system.
> If you read out a series of digits in chunks (pause between
> groups of digits) they will be easier to recall than if they
> are read out at a fixed rate for each digit.
>
> ---> time
>
> 820938947108
> 820 938 947 108
> 8209 3894 7108
>
> How does operant conditioning predict this memory behavior?
Again, you aren't testing operant conditioning at that point. You are
testing the limits of the perception system which drives the behavior.
Again, to implement operant conditioning you have to build a perception
system that drives behavior. Operant conditioning then adjusts the mapping
from perception to behavior. To test how well the operant conditioning
system is working, you have to test how behavior _changes_ over time.
A test like the one above, show the simple limits of the perception system.
If you pause, you are changing the patterns that you are asking the
perception system to recognize because the perception system is a temporal
pattern perception system.
Asking it to recognize the temporal pattern 820 is withing the limits of
what it can recognize. Asking it to recognize "820938947" is beyond it's
limit to recognize a pattern that large. It can recognize 4 short patterns
of about .5 seconds in length but it can't recognize a 2 second long
pattern of that complexity. Tests like that show what the perception
system is able to do, and what it's limits are. And they give insights
into how it might have been implemented.
> It has been found that the faster a person can speak a set
> of words the more words they could hold in memory. It turns
> out that the subjects could only remember as much as they
> could say in 1.5 seconds. Why does this time limit emerge
> from operant conditioning?
PERCEPTION SYSTEM LIMTIS.
If we built an operand conditioning system that used a digital camera as
it's sensory input, you could understand why the system couldn't see a
circle when it got to small right? Well temporal pattern recognition
systems have limits as well. It's a time vs resolution trade off that's
inherent in the systems design. This tests give you insights into the
limits of the perception system and in turn, give us insights into how it's
implemented.
> You say the brain is one big single module. Ok why does
> disabling one of its components, the hippocampus, result
> in the inability to remember what it did five minutes ago
> but still enables it to remember new skills? How does this
> emerge as a result of operant conditioning?
I don't know.
It's an implementation detail of the brain - not of operant conditioning.
It's an implementation detail THAT IS NOT IMPORTANT TO MAKING MACHINES
INTELLIGENT because the be intelligent, you don't have to implement the
operant conditioning system exactly like it's implemented in the brain. How
you implement it however will determine what range of skills it can learn,
and what it can't learn, and how fast it can learn. The closer you get the
implementation to the brain's implementation the more human-like the skill
set will become, but it can still be highly intelligent even without a
close match to human skills.
I also can't even guess at your question because you aren't being specific
enough about what humans do. What is "remember a new skill" and what is
"not remember what it did five minuets ago"? If you can be more specific
by giving a real example of what one of these people do in both cases I'll
be happy to speculate about what role the hippopotamus might be playing in
the brain's implementation of operant condition.
The reason our ability to mimic can't be separated from our ability to be
conditioned is because they both control the same behaviors. How would you
build a system that mimics? How would it implement the complex task of
mimicking by watching what someone else did? How would it decide what
aspect of the behavior was to be mimicked? Humans can't perfectly clone a
behavior so we really don't just mimic actions. What do we, is mimic
results. But we don't mimic every result, we mimic the results that _we_
determine to be important. If IO watch someone open a cookie jar, I might
open the same cookie jar so I too can get a cookie. But they might do it
with their left hand, and I my right hand. Did we fail to mimic the action
if we don't get the hand right? No, because we weren't attempting to
exactly clone the full behavior. We were trying to clone the results of 1)
getting the lid of using our hands, and 2) get the cookie in our mouth.
But how does a "mimic" module abstract out what result or action to mimic,
and what not to mimic? This is a problem of worth - of value. And mimic
hardware doesn't explain value. Operant conditioning however does explain
where the concept of value comes from and how one action can be evaluated
for worth over another. All the problems you have to solve to implement
operant conditioning are the same problems you have to solve to mimic
results of importance. In order for it to make _intelligent_ choices about
what, and how, it mimic, it has to implement all the perception, and action
selection hardware needed to implement full operant conditioning.
The problem is that ALL INTELLIGENT ACTIONS must be evaluated for worth.
We will stop trying to mimic the "get a cookie" behavior the second a
stimulus of higher importance comes along and triggers an action of high
importance. All the actions that we learn, have to selected by a _single_
coherent, behavior priority system that determines at every instant in
time, what behavior is the current most important to be doing. Operant
conditioning explains how all our behaviors become prioritized under a
single global rating system. Without a single global behavior rating
system, we would not be able to instantly interrupt any one behavior, and
start producing a different behavior in it's place. You can't make such a
system work very well if some of the behaviors are produced by a hard coded
"mimic" module and others are produce by other modules. It's just not a
workable engineering solution to the problem of how you prioritize all
behaviors, and _adjust_ the priorities of all those behaviors by a single
global conditioning system.
casey
> casey <jgkjca...@yahoo.com.au> wrote:
>> If you read out a series of digits in chunks (pause between
>> groups of digits) they will be easier to recall than if they
>> are read out at a fixed rate for each digit.
>>
>> ---> time
>>
>>
>> 820938947108
>> 820 938 947 108
>> 8209 3894 7108
>
>
> You are testing the limits of the perception system which drives
> the behavior.
But it is still an _ability_ to retain a set of digits in memory
and you made the statement that "strong operant conditioning
explains ALL our abilities". Now you are saying that the ability
to recall a string of digits is due to a perception system??
And how do we know what abilities is an innate product of this
perception system and what is due to operant conditioning?
> Again, to implement operant conditioning you have to build a
> perception system that drives behavior. Operant conditioning
> then adjusts the mapping from perception to behavior. To test
> how well the operant conditioning system is working, you have
> to test how behavior _changes_ over time.
How do you separate changes in behavior due to maturation from
those due to conditioning?
>> It has been found that the faster a person can speak a set
>> of words the more words they could hold in memory. It turns
>> out that the subjects could only remember as much as they
>> could say in 1.5 seconds. Why does this time limit emerge
>> from operant conditioning?
>
>
> These tests give you insights into the limits of the perception
> system and in turn, give us insights into how it's implemented.
And the explanation given for this 1.5 second limit is called the
phonological loop which experiments suggest is required for the
acquisition of language. We can disingage the system by preventing
the person using internal vocalization.
It may also be telling us how it deals with its limitations and
all systems have limits. You use the throw away explanation of
chunking being due to a temporal pattern perception system limited
to 3 or 4 items at a time however considering the size of the brain
it appears to be an usually small limitation. A computer can remember
digits of any length. Maybe chunking serves a purpose?
> ... you don't have to implement the operant conditioning system
> exactly like it's implemented in the brain.
From what you have written above it appears to me you do however
have to start with a working "perceptual system" which will limit
any changes possible with operant conditioning.
So perhaps when you are talking about operant conditioning I am
talking about the perceptual system?
> I also can't even guess at your question because you aren't being
> specific enough about what humans do. What is "remember a new
> skill" and what is "not remember what it did five minutes ago"?
> If you can be more specific by giving a real example of what one
> of these people do in both cases I'll be happy to speculate about
> what role the hippopotamus might be playing in the brain's
> implementation of operant condition.
http://en.wikipedia.org/wiki/Clive_Wearing
http://en.wikipedia.org/wiki/HM_(patient)
> Humans can't perfectly clone a behavior so we really don't just
> mimic actions. What do we, is mimic results.
I agree with that. In the case of the Ape you seem to be suggesting
it doesn't have the perceptual system to "see clean dishes" or maybe
to value "clean dishes"? So this seems to be all about building a
perceptual system so you have things to associate, something I have
suggested at the beginning. The confusion is you kept saying it is
"nothing but" operant conditioning perhaps referring to how it is
learned not how things are found to learn about? Association doesn't
strike me as a difficult thing to implement but finding things worth
associating seems to me to be the hard part. How do we perceive
"clean dishes"?
JC
Curt Welch
> On Dec 26, 6:23=A0am, c...@kcwc.com (Curt Welch) wrote:
> > casey <jgkjca...@yahoo.com.au> wrote:
> >> If you read out a series of digits in chunks (pause between
> >> groups of digits) they will be easier to recall than if they
> >> are read out at a fixed rate for each digit.
> >>
> >> ---> time
> >>
> >>
> >> 820938947108
> >> 820 938 947 108
> >> 8209 3894 7108
> >
> >
> > Again, you aren't testing operant conditioning at that point.
> > You are testing the limits of the perception system which drives
> > the behavior.
>
> But it is still an _ability_ to retain a set of digits in memory
> and you made the statement that "strong operant conditioning
> explains ALL our abilities". Now you are saying that the ability
> to recall a string of digits is due to a perception system??
I have ALWAYS said that to BUILD an operant conditioning system, you must
include a perception system. Operant conditioning can't work without
innate perception abilities. How many times have I talked about the fact
that I once tried to do it in two modules which I called the inet and the
onet, with my inet "module" performing the perception function and the onet
"module" performing the reinforcement learning (aka operant conditioning)?
I've also explained how I concluded it could never work when divided that
way and that both perception and conditioning had to exist in each module
of a distributed network.
This is no more complex than arguing that we need to build a car, and
then when I talk about what type of engine it has, you claim "BUT YOU SAID
IT NEEDED TO BE A CAR, NOT AN ENGINE!". When I say it's a car, that
implies the fact that it needs an engine.
When I say it's all operant conditioning, that implies the fact that it
needs innate perceptional and action mechanisms.
> And how do we know what abilities is an innate product of this
> perception system and what is due to operant conditioning?
Operant conditioning is a learning function. It explains how behavior
changes over time, and why. It explains how a given perceived stimulus
condition becomes associated with a given behavior. The ability to
perceive the stimulus and the ability to perform the action is always
innate.
> > Again, to implement operant conditioning you have to build a
> > perception system that drives behavior. Operant conditioning
> > then adjusts the mapping from perception to behavior. To test
> > how well the operant conditioning system is working, you have
> > to test how behavior _changes_ over time.
>
> How do you separate changes in behavior due to maturation from
> those due to conditioning?
You can't. It's almost impossible without a full and completely
understanding of the internal implemetnation of the body - which no one has
yet. Much of what people like to claim as the body "finishing the growing
function" I'm sure is actually just learning at work. But without fully
understanding how the body and brain develops, we can't really answer what
development is really just learning, vs some innate growth and wiring
function.
I suspect careful testing could actually revel much, but it would risk the
correct development of a child so it's not going to get done as part of any
controlled experiment.
> >> It has been found that the faster a person can speak a set
> >> of words the more words they could hold in memory. It turns
> >> out that the subjects could only remember as much as they
> >> could say in 1.5 seconds. Why does this time limit emerge
> >> from operant conditioning?
> >
> >
> > These tests give you insights into the limits of the perception
> > system and in turn, give us insights into how it's implemented.
>
> And the explanation given for this 1.5 second limit is called the
> phonological loop which experiments suggest is required for the
> acquisition of language. We can disingage the system by preventing
> the person using internal vocalization.
Yes, they make up all sorts of silly names for this data don't they.
> It may also be telling us how it deals with its limitations and
> all systems have limits. You use the throw away explanation of
> chunking being due to a temporal pattern perception system limited
> to 3 or 4 items at a time however considering the size of the brain
> it appears to be an usually small limitation. A computer can remember
> digits of any length. Maybe chunking serves a purpose?
Well, no, I think the entire concept of "chunking" is highly misguided.
The little I've read about the theories of memory that use those sorts of
terms seems to be quite far off base on what's happening internally in my
view.
What it's showing us however, is limitations of the specific
_implementation_ the brain is using. There are no doubt other ways to
implement a similarly strong system that will have a fairly different set
of limitations. However, it's all potential clues into how it can be
implemented.
> > ... you don't have to implement the operant conditioning system
> > exactly like it's implemented in the brain.
>
> From what you have written above it appears to me you do however
> have to start with a working "perceptual system" which will limit
> any changes possible with operant conditioning.
Well, yes and no. I believe to create strong conditioning, the operant
conditioning must be applied to the perception system itself. It can't be
done as two separate modules. So the conditioning needs to actually change
the systems perception. Again, this is something I've tried to talk to you
about many times over the past 5 or so years.
However, it's Yes, in the extent that the limits of the system's ability to
perceive are also the limit of the systems ability to learn.
My pulse sorting networks - the ones you have written code for - are
working examples of how you can build a perception system that can be tuned
by reinforcement. If you have a 10 node network, then the limits of what
that network can be configured to recognize (perceive) are also the limits
of what the network can learn.
> So perhaps when you are talking about operant conditioning I am
> talking about the perceptual system?
At times, maybe. But again, I don't think it's possible to separate the
two as separate problems to be solved. We must solve them both at the same
time because any design for a perception system that isn't structured so it
can be adjusted by a reward signal has no hope of being directly useful in
the solution to AI.
> > I also can't even guess at your question because you aren't being
> > specific enough about what humans do. What is "remember a new
> > skill" and what is "not remember what it did five minutes ago"?
> > If you can be more specific by giving a real example of what one
> > of these people do in both cases I'll be happy to speculate about
> > what role the hippopotamus might be playing in the brain's
> > implementation of operant condition.
>
> http://en.wikipedia.org/wiki/Clive_Wearing
>
> http://en.wikipedia.org/wiki/HM_(patient)
I see almost no ability for these people to "remember a new skill". What
were you talking about? Clive Wearing seems to be forever stuck in a life
before the brain damage - which just indicates some key part of the
learning system has been disabled.
> > Humans can't perfectly clone a behavior so we really don't just
> > mimic actions. What do we, is mimic results.
>
> I agree with that. In the case of the Ape you seem to be suggesting
> it doesn't have the perceptual system to "see clean dishes" or maybe
> to value "clean dishes"?
Yes, it could be a little of both at work.
> So this seems to be all about building a
> perceptual system so you have things to associate, something I have
> suggested at the beginning.
And something I've been creating networks that actually solve and trying to
explain to you how they solve it for 5 years here and trying to get you to
come up with useful ideas on how to do better for 5 years.
> The confusion is you kept saying it is
> "nothing but" operant conditioning perhaps referring to how it is
> learned not how things are found to learn about? Association doesn't
> strike me as a difficult thing to implement
In low dimension problems, it's trivial. That's why RL algorithms are
trivial to write for low dimension problems. In high dimension problems
the number of possible associations are so large, the problem becomes very
hard. Far too hard to solve with the overly simplistic "try things until
you find stuff that works" approach. It has to be solved using a different
approach - one that is able to converge on good associations over time
without trying a billion bad ones first. No one has yet found a good
generic solution to this problem yet - which shows how hard it is.
> but finding things worth
> associating seems to me to be the hard part.
Yes, well, it's not hard for low dimension problems. It's trial. But it's
very hard for high dimension problems because there are so many
associations which are not very valuable.
However, everything we perceive has _some_ value. So finding perceptions
that have some value is trivial. Finding lots of perceptions with really
good value is what is not so easy. But the way the high dimension problem
is solved, is to start with some default associations, and adjust them so
as to keep increasing their value - this sort of approach allows the system
to converge on far better associations over time.
> How do we perceive "clean dishes"?
Well, can a kid that hasn't yet developed language correctly learn to mimic
dish washing? I suspect not. I suspect we "perceive" "clean dishes" with
the help of our langauge system. But that just makes it a more complex
problem to explain because you have to at the same time explain how we are
taught to talk about clean dishes.
The more general question of how does perception work in an high dimension
reinforcement trained learning system is the general problem I've been
talking about for years here - and which I've written some rather long
posts explaining my best understanding of how to solve it yet again
recently.
I don't think there's anything special about the perception of clean dishes
that isn't included in the general problem of being able to perceive
anything, from a face, to an ice cream cone. It's all the same problem.
Yevgen Barsukov
Lets stop right there. You use the word "meme" in many of
your statements, and that is not what I am talking about,
at least not in the sense how Dawkings defined "meme". He defined it
is some cultural content which is irrelevant for purpose of Life and
propagates just because it has effective reproduction mechanisms and
because the immune system of actual Life it parasites on did not yet
figure out how to stamp
it out.
Being a zoologist involved with mechanisms rather than reasons,
Dawkings is not focusing fundamental thermodynamical nature of Life
and so it is all a game of reproduction to him. In reality Life and
all its forms are just such a necessity in certain chemical
environments as a whirlpool in a stream when water flow exceeds
certain rate.
And it does not matter if this is Life which decision programming is
stored in DNA or it is stored in neural states or it is stored in
scrolls of papyrus - point is, Life will take whatever form is
thermodynamical optimal at this moment and will be most capable of
accelerate entropy increase.
When I talk about cultural programming, I mean it with the same level
of inherent belonging to Life as genetic programming. Humanity is just
a Life form, some of which is not stored in DNA, but stored in books,
CDs and Usenet archives. What we see around us - houses, cars, large
hadron collider and cruiseships are phenotypes of this Life-form.
>
> The genes have total control over what memes it will allow the body to
> except. The memes have very little control over the genes (currently).
> There is no doubt that currently, the genes are in control here - not the
> memes. The memes are emergent behaviors that the genes allow to happen.
> When the genes create emergent meme behaviors in our society, and those
> memes fail to do a good job of keeping the genes alive, then those genes
> (and the memes they allows to emerge) all die off together.
Genes are in control of physiological functioning of the body
(including its input/output system such as hearing, vision etc).
What it can not control is what these systems are hearing and seeing.
Here is one extreme example - A baby can be placed in a dark room
without any information input (choice A) or into a room full of toys,
parents, aunts and other siblings (choice B). What it will learn is
completely determined by these choices. But - neither of these two
choices in controlled by the genes in the slightest. What is it
controlled by?
It is controlled by cultural programming of the parents and other
programmed monkeys this baby is in charge of.
So is everything else what baby will learn, which in turn will define
what it will teach to its own babies and so on and so forth. By
controlling the environment from the time when brain is starting to
learn, cultural program is the master of the content, while genes are
just responsible for the execution of whatever content is presented by
obediently delivering sensory inputs carefully selected and controlled
by cultural programmers down the throat of open and receptive blank
brain.
Note that such receptiveness itself has been evolutionary reinforced
because babies that would not learn
got eaten by hyenas, fall of the entrance of the cave, got burned by
the fire or otherwise died from being incompatible with culture-
defined phenotype of humanity and so did not propagate their genes.
Those the very genes that you think me cause opposition to
cultural programming have been eliminated during last 500 000 years of
culture-defined evolution of monkeys. Only the
"culturally receptive" monkeys have been selected and became
the backbone of our modern gene pool.
Admittedly, there are many behaviours (instincts) that have peen pre-
programmed genetically at the time that precedes formation of cultural
programming system and are still in many cases duplicating or
contradicting cultural programming. But honestly - make an experiment
and analyze yourself for say last 2 hours: how many of actions you did
were caused by instinct and how many by cultural programming?
....
I am sure you found that excluding purely physiological functions like
breathing, eating and urinating most of your actions were caused by
cultural programming and have no instinctive basis.
> Humans can't be forced to accept any meme. They only accept the memes that
> are proven to useful to prevent harm to the body or harm to our
> reproductive success - which is a requirement created by the genes. They
> are the gate keepers of what memes are allowed to take hold in us and
> always have the upper hand.
I think this statement is not even true to actual memes that Damkins
defined. The whole point of memes is that they posess mechanisms of
propagation that are unusually sucessful (such as deviation in a
lyrics of a hymn that is caused by typical accustics of a chappel).
As for applicability to cultural programming, as I have shown above,
baby's genetic make-up has no say whatsoever what programming it will
accept, it just defines the
structure of machinery that is being USED in this programming.
And this machinery is pretty much the same, except of some
pathological cases where it is broken.
>
> We don't accept just any meme we hear from society. Our brain judges the
> value of the different memes it is exposed to, and accepts the ones it
> evaluates as being the best - the most useful. But the evaluation system
> the brain uses, is specified by our genes. Which is why the genes have
> ultimate control over the memes - not the other way around in any sense.
>
> Learning systems are flexible in that they have the power to learn a wide
> range of very odd behaviors - like swinging a stick at a little white ball
> in order to maximize our odds of survival (Tiger Woods). But they are not
> so flexible that they will learn just anything they are exposed to. They
> don't work like a camera just capturing and duplicating _anything_ it sees.
> It's nothing like that.
I quite agree. It is not simple "mimicking" that is why I don't like
work meme. Cultural programming technology is just as complex and
incomprehensible as genetic programming in DNA - it has been evolving
ever since language was created or maybe even before, and so it has
countless tricks and optimizations both to store and to imprint useful
for humanity information which we are not nearly aware of. It is also
rapidly changing.
For example in the past some empirical optimizations found by deadly
trial and error experiments (for example that you need to wash hands
before eating) had to be stored in religious stories and fairy-tales,
now many of such optimization have well understood reasons (existence
of bacteria, known since very recent in historical therms discovery)
and so can be imprinted through direct causal explanation that is
based on education. Interestingly, in both cases end result is the
same - monkey is going to wash its hands and not die from dysentery,
even though methods to achieve this results appears opposite.
All changes to our behavior happen for very
> specific reasons - because the change is shown by the internal behavior
> worth estimation system to be better than the previous behavior.
>
> Looking at the odd things humans do, you might not understand how on earth
> some behaviors could exist. But they all tie back to the experience that
> person has had in their life relative to their internal drives defined by
> their genes.
In many cases experiences of a single person are not statistically
significant to verify that certain behavior is useful or not for
humanity. However, experimentation of humanity as a whole can span
over hundreds of years of maintaining certain behavior mandated by
cultural programming and find with absolute significance that say
washing hand or not marrying cousins is harmful or useful.
Individual human does not discover by himself a damn thing, and even
when he does in most cases it is illusion specially created by
cultural programming hidden methodics to keep things interesting and
to encourage the particular individual
monkey who was lucky to be the final step in a long chain of discovery
process to disseminate the resulting information. This even applies
for purely empirical and serendipitous discoveries, because the
definition of the problem and significance of the observation
themselves is created by cultural programming.
Real discovery processes span over many generations and
include huge loads of information created by hundreds or thousands
which admittedly can fit into a single monkey brain, but is mostly
loaded there ready to consume. IF you want to challenge this
assertion, give me an example of a discovery you made, and I will show
you how cultural programming has totally caused and defined it.
> Learned behavior are just as instinctive as fixed behaviors. They are just
> more complex. That is, the instinctive behavior of learning is operant
> conditioning.
>
> With a simple instinct, we might sense heat, and move away from it. The
> hardware is pre-build to always perform that action to that type of
> stimulus. With operant conditioning, we see some fire, move our hand
> towards it, get burned, and then the internal behavior system adjusts the
> map so we are less likely to produce that same behavior in the future.
> Though the "move away from heat" is pre-wired in one case, the "adjust our
> behavior to prevent sensing too much heat" is innate in the other. It's an
> innate behavior created by the genes in both cases. The first is just
> stimulus response, the second is stimulus, response, evaluation, stimulus,
> different response. The sequence is innate and unchangeable in both and
> fully specified by the genes - who always have the upper hand (for now).
>
> As long as humans continue to be grown as they are now, under the control
> of our genes, the genes will continue to have the upper hand here. But the
> more we use our intelligence, to change our bodies, and our genes, the more
> complex the entire process of evolution becomes and the less we will be
> able to say the genes are in control.
>
> However, the upper hand of evolution will never lose control. That is,
> whatever form we evolve to, it will only exist, if it proves to be good at
> survival. At the moment, our genes are still the prime controller of our
> evolution, and the memes really control almost nothing - they were created
> by and allowed to exist, by our genes.
Evolution of genes is already secondary for humanity, it is evolution
of ideas that is primary. Think about a thousand of wild aborigines
with perfectly fine genes, and handful of cultured man with AK-47
having the power of ideas evolution. Who will have survival advantage?
This is even more obvious if you look at thermodynamical
implications. One human is only accelerates about 100W worth of
entropy increase (result of genetic evolution) but the car he drives
accelerates 250 kW of entropy increase this car is a phenotype of
humanity that is caused by cultural programming and has no
representation in genetic programming at all.
It is not to say that genetic evolution has finished, it still
continues but in the environment created by cultural evolution
and to satisfy thermodynamical optimum behavior of Life of the
humanity as a super-organism defined by both cultural and genetic
programming rather that of an individual monkey.
DNA is still a storage medium used by this super-organism, along
with CDs and paper, and it still has extremely efficient and extremely
robust survival programming that will be used and reused by humanity
for thousands years to come. But it has lost its uniqueness as the
_only_ way how evolving programming can be stored.
Regards,
Yevgen
>
> --
> Curt Welch http://CurtWelch.Com/
casey
>
>>> I also can't even guess at your question because you aren't being
>>> specific enough about what humans do. What is "remember a new
>>> skill" and what is "not remember what it did five minutes ago"?
>>> If you can be more specific by giving a real example of what one
>>> of these people do in both cases I'll be happy to speculate about
>>> what role the hippopotamus might be playing in the brain's
>>> implementation of operant condition.
>>
>> http://en.wikipedia.org/wiki/Clive_Wearing
>>
>> http://en.wikipedia.org/wiki/HM_(patient)
>
>
> I see almost no ability for these people to "remember a new skill".
"Wearing can learn new practices and even a very few facts –not from
episodic memory or encoding, but by acquiring new procedural memories
through repetition."
http://en.wikipedia.org/wiki/Clive_Wearing
And read again the section on Motor skill learning at,
http://en.wikipedia.org/wiki/HM_(patient)
I read all about this in books about the brain. Remembering episodes
in your life (episodic memories) involves different parts of the brain
to learning motor skills. Learning problem solving procedures is also
a "motor skill". In rats the hippocampus is involved in knowing where
the rat is and the same in true in humans. Wearing and HM cannot find
their way around a spatial environment (although I suspect they could
be taught a stimulus response procedure to navigate with). I suspect
that the machinery we use make a spatial representation to find our
way around we also use to represent our life episodes.
> What were you talking about?
Is it now any clearer?
> Clive Wearing seems to be forever stuck in a life before the brain
> damage - which just indicates some key part of the learning system
> has been disabled.
But this "part" only effects the laying down of episodic memories,
it doesn't remove the ability to lay down memories as to how to do
something. If you never tested for episodic memories you would
never know it was lacking.
> It [perception] has to be solved using a different approach - one
> that is able to converge on good associations over time without
> trying a billion bad ones first. No one has yet found a good
> generic solution to this problem yet - which shows how hard it is.
Evolution I would suggest has found such generic solutions in the
form of heuristics which is exactly what we do in solving high
dimensional problems such as a game of chess. The question you need
to ask is how do we extract heuristics from our experiences?
>> finding things worth associating seems to me to be the hard part.
>
>
> ... it's very hard for high dimension problems because there are
> so many associations which are not very valuable.
Which is why I have suggested evolution found what amounts to the
same thing as heuristics in our software to decide what is valuable.
>> How do we perceive "clean dishes"?
>
>
> I suspect we "perceive" "clean dishes" with the help of our
> language system.
The recognition of a clean dish vs. a dirty dish I suspect would be
easy to implement. The question is why a system would care one way
or the other. Why would the state of the dish be rewarded? The
Ape it is not rewarded so it never twigs that we started with dirty
dishes and ended with clean dishes. It does however enjoy playing
with dishes and soapy water. If you were to hide food under clean
dishes and not under dirty dishes I suspect the Ape would learn to
discriminate between the two and maybe learn to wipe off the dirt
if it could associate the result, a clean dish, with a reward.
current state (dirty dish)
choose operator (wipe dish)
desired state (clean dish)
> I don't think there's anything special about the perception
> of clean dishes that isn't included in the general problem
> of being able to perceive anything, from a face, to an ice
> cream cone. It's all the same problem.
This is where GOFAI can give some insight, for in trying to duplicate
certain human skills, such as recognizing a face or inferring a goal
(clean dishes) in the actions of another you get some idea of what
the learning system has to achieve and thus a measure of its
difficulty
by how hard it is to duplicate. Calculus is easy, seeing is hard.
JC
Curt Welch
> On Dec 27, 6:32=A0am, c...@kcwc.com (Curt Welch) wrote:
> >
> >>> I also can't even guess at your question because you aren't being
> >>> specific enough about what humans do. What is "remember a new
> >>> skill" and what is "not remember what it did five minutes ago"?
> >>> If you can be more specific by giving a real example of what one
> >>> of these people do in both cases I'll be happy to speculate about
> >>> what role the hippopotamus might be playing in the brain's
> >>> implementation of operant condition.
> >>
> >> http://en.wikipedia.org/wiki/Clive_Wearing
> >>
> >> http://en.wikipedia.org/wiki/HM_(patient)
> >
> >
> > I see almost no ability for these people to "remember a new skill".
>
> episodic memory or encoding, but by acquiring new procedural memories
> through repetition."
>
> http://en.wikipedia.org/wiki/Clive_Wearing
>
> And read again the section on Motor skill learning at,
>
> http://en.wikipedia.org/wiki/HM_(patient)
>
> I read all about this in books about the brain. Remembering episodes
> in your life (episodic memories) involves different parts of the brain
> to learning motor skills. Learning problem solving procedures is also
> a "motor skill". In rats the hippocampus is involved in knowing where
> the rat is and the same in true in humans. Wearing and HM cannot find
> their way around a spatial environment (although I suspect they could
> be taught a stimulus response procedure to navigate with).
That doesn't seem to be correct for HM. The article says he was able to
draw a detailed map of the home he moved to 5 years after the surgery.
Considering how little he could learn, that was surprising.
> I suspect
> that the machinery we use make a spatial representation to find our
> way around we also use to represent our life episodes.
>
> > What were you talking about?
>
> Is it now any clearer?
Well some.
> > Clive Wearing seems to be forever stuck in a life before the brain
> > damage - which just indicates some key part of the learning system
> > has been disabled.
>
> But this "part" only effects the laying down of episodic memories,
> it doesn't remove the ability to lay down memories as to how to do
> something. If you never tested for episodic memories you would
> never know it was lacking.
That doesn't seem to be very accurate for Wearing.
The rest of the quote from the wikipeda article (following your quote
above):
"For example, having watched a certain video recording multiple times on
successive days, he never had any memory of ever seeing the video or
knowing the contents, but he was able to anticipate certain parts of the
content without remembering how he learned them.[2]"
This just shows he has _some_ trivially small amount of learning still
working. Being able to anticipate events in a movie is learning. But it's
sure nothing like you describe as "doesn't remove the ability to
lay down memories as to how to do something". HM on the other hand seems
to have more learning ability left.
In my approach to trying to create operant condition in this problem space,
I've had to build in two types of learning to my networks. The first could
be described as classical conditioning - and is the same learning required
to create a good perception system because it's only driving by the
constrains of the sensory data and not by conditioning. So as you have
seen, I have used two learning parameters in each node of my network. The
first which is the gap_value is adjusted according to the sensory data.
And the second, the sorting ratio, is adjusted by the reward signal. So
each node has two very different types of learning at work. Externally,
they create classical and operant conditioning effects.
It could well be the brain has likewise implemented two parallel learning
systems to solve these same two basic problems. The first system would
likely be implemented in each node of the network in how each node adjusts
synapse weights and grows connections. It could be a totally parallel and
distributed type of learning with no central "controller" of any type
needed. Conditioning however requires a global reward signal acting as a
central controller and probably the use of chemical messengers to
communicate and change all the nodes in the network as required. There is
likely no way to globally disable the first type of learning because there
is probably no central control for it to disable. Though I guess there
could be some type of chemicals used to make it work - and if there was a
global way to block that chemical it might gobally shut down that type of
learning. But in the second, if you disable or harm the central reward
signal function, you would turn off all conditioning.
It sounds to me, with Wearing, that that is what happened to him. He's
lost all ability to be trained by his reward system. He can't learn
anything new based on rewards (primary or secondary). His brain is stuck
in the configuration it was in when his learning ways turned off.
But his classical conditioning system is still functioning. Which might
explain why, when watching a movie multiple times, he learns to anticipate
coming scenes even though he had never seen it before his damage. It could
just be his perception system re-tuning itself to adjust it's expectations
of what to expect next. For it to have a noticeable effect in his
behavior, it would have had to be something he was trained to understand
before he lost his ability to learn new things. But that's easy to
understand that any adult would have circuits to "understand" and "respond
to" major events in a movie. The classical conditioning didn't teach him
how to respond to something like the car crash in the movie, but it would
allow him to anticipate it was about to happen so that he would produce his
pre-learned reaction, before it actually happened in the move.
The wikipedia article only gave that one example of what Wearing could
learn, but from that, it seems a far stretch from your description of being
able to learn new motor events. It sounds to me like Wearing has lost all
(or almost all) of his ability to learn by reward (operant conditioning)
and only has left his ability to have his perception adjusted by classical
conditioning.
HM on the other hand is far more complex. It looks like he still has some
conditioning left, but that the surgery just disable large and important
parts of his brain.
Even with my belief that the cortex is fundamentally a homogeneous learning
network, what you end up learning does end up being distributed to
different parts of the network. Every part ends up performing different
functions. If you just go around and cut out random parts of the network,
you are going to lose random parts of your behavior, and lose various
associated abilities to re-lean even the type of behavior that was lost
because you no longer have the right toplogy network to learn those
behaviors.
At the same time, the implementation of operant conditioning isn't all that
simple because of the requirement that it must also implement secondary
reinforcement. Without secondary reinforcement, you could imagine an
implementation that used a chemical messenger to represent the reward
signal which was flooded out to the whole brain and effected each neuron
equally. But because secondary reinforcement is needed, each node in the
system must also have some ability to act as a reward predictor and
generate reward signals that train other nodes in the system. But yet, all
the secondary rewards must be slaves to the primary rewards - somehow.
It's a complex problem that in the case of the distributed networks I've
played with, I've never found a good implementation for. But the brain
does somehow manage it. And if you cut out random sections of the network,
you might end up blocking the flow of secondary rewards to some sections of
the network, but not others. Which might create odd and complex damage to
the function, such that secondary rewards still work in some sections of
the network, but not others, or that primary rewards don't correctly
function in all sections of the network. Which leaves an odd mix of
learning abilities in tact while striping out others.
My point however, is just because you call some behaviors "memory" and
other behaviors "motor skills" and make up different labels for the
different types of "memory' and different types of "motor skills" doesn't
in any way dis-prove the theory that the cortex operates as one large
generic operant conditioned learning system. All you have done, is labeled
the different parts of the generic network and the type of function that
happens there.
> > It [perception] has to be solved using a different approach - one
> > that is able to converge on good associations over time without
> > trying a billion bad ones first. No one has yet found a good
> > generic solution to this problem yet - which shows how hard it is.
>
> Evolution I would suggest has found such generic solutions in the
> form of heuristics which is exactly what we do in solving high
> dimensional problems such as a game of chess. The question you need
> to ask is how do we extract heuristics from our experiences?
That's the same question I've always been asking.
But what are you saying above? Evolution found the heuristics? Or
evolution found a system that can find heuristics for us? I agree with the
second, I don't agree with the first. Well, yes, evolution clearly has
found heuristics whenever it hard-codes a solution - but I mean that what
makes us intelligent, is the first. The ability of our brain to learn.
When you take away the ability to learn, you are left with a brain like
Wearing. He's got a huge collection of complex behaviors hard-wired now
but they can no longer change and adapt. He's like any other dumba
hard-wired AI project that are able to perform complex functions, but which
can't learn anything on their own.
Evolution made us stand out from the rest of the lower animals because it
gave us the very strong learning system that allows our behaviors to be
re-wired over time by conditioning to allow us to learn what at times feels
like no end of new complex behaviors (even though there are limits to our
learning).
> >> finding things worth associating seems to me to be the hard part.
> >
> >
> > ... it's very hard for high dimension problems because there are
> > so many associations which are not very valuable.
>
> Which is why I have suggested evolution found what amounts to the
> same thing as heuristics in our software to decide what is valuable.
Well again, you can not explain our behavior by assuming secondary
reinforcement was hard-code by evolution. It was not evolution that
figured out for us that dollar bills and coke bottles should be considered
valuable in our perception system.
Humans LEARN what is valuable by simple and well understood secondary
reinforcement. It's well understood and fully implemented in our
reinforcement learning algorithms already for the small scale problems.
The only thing we are missing, is a workable implementation for high
dimension problems.
Secondary reinforcement works in TD RL algorithms by using estimates of the
value of a current state to condition the states that came before. When
you make a move in a game like Backgammon, the value estimating system
gives you value for the board position. And that value is back-propagated
to the previous board position as a reward signal. When a move (an action)
made by the agent, causes the environment to change into a state that the
internal estimation system believes is higher, that acts as a reward for
that action. This is secondary reinforcement because3 it happens without
any primary reward being received at that time.
Again, how to make that work in low dimension problems if VERY WELL
understood. It's done by internally storing a separate reward value for
every state the environment can be in (or sometimes every state->action
pair). Over time, these values converge on accurate estimates of the
states true values. Over time, this system becomes a perfect future reward
predictor.
But that only works for low dimension problems where 1) every state of the
environment can be stored in memory, and 2) every action from every state
can be tried many times to allow the system to learn.
In high dimension problems (all problems of interaction with the real
world) we 1) don't know the true state from the sensory data 2) couldn't
store value estimates for each state even if we did known it, and 3) will
never get to visit most state->action pairs and won't get to visit any of
them even twice.
But this just means that first we need 1) to create an abstract model that
represents as much of the state of the environment in whatever space we
have to work with as possible (which is what classical conditioning is used
for) and 2) and used a distributed action system so that each distributed
state->action event gets visited many times, even though the total
state->action event may never be visited twice.
So we train a network of nodes, where each node deals with it's micro
state->action learning problem (the training of a single node in my
network), and allow the nodes to act together, to solve the global learning
problem. So though the network as a whole, may never be in the same state
twice, each node in the network is seeing the same states over and over
again.
So, yes, heuristics are needed to both allow us to act intelligently, and
to know what is valuable - both of which I believe are one and the same
problem. And both of which MUST BE LEARNED in order to explain advanced
human behavior.
A generic operant conditioned learning system is the only thing that can
fully explain where our intelligent behavior comes from. None of it can be
explain by evolution hard-coding it for us.
What makes us intelligent is that evolution figured out how to separate
itself from the problem of hard-coding all our intelligence into us, and
instead, only hard-code our primary motivations, while allowing the
behavior system to self adapt on it's own - which means learning not only
the behaviors, but also learning to estimate secondary rewards by learning
what is valuable in our perception system.
Yes, building such a generic system is not easy,. But you keep wanting to
try and explain how it's solved by claiming evolution "hard-coded" some
amount of the answers into us thereby making the problem somehow easier
(less for us to learn). But we know exactly how much humans can learn both
in the way of learned behaviors, and learned values so we know Evolution
didn't encode that in us.
And though this type of learning is not easy to implement, We already now
exactly how to solve the problems of dimension by using a distributed
learning system. TD-Gammon showed how it can be done for that one domain.
So there's nothing big missing there. There's only a few small
implementation details to be worked out.
And I keep trying to get other people to realize this, and work on those
problems, but what I spend all my time writing about, is just trying to get
you to see, and work on, the obvious, instead of making up bogus excuses
about how evolution hard-coded answers for us instead of implementing a
strong generic learning system. We KNOWN evolution implemented a strong
generic learning system because we can test human generic learning skills
and see it at work. There's no question as to whether it's there in
humans. It's there, and until we duplicate the strong generic learning
system, we won't have solved AI - we will at best create another Clive
Wearing in a machine.
> >> How do we perceive "clean dishes"?
> >
> >
> > I suspect we "perceive" "clean dishes" with the help of our
> > language system.
>
> The recognition of a clean dish vs. a dirty dish I suspect would be
> easy to implement. The question is why a system would care one way
> or the other.
Because it's a strong generic reinforcement learning systems and "caring"
is what these systems are designed to do.
> Why would the state of the dish be rewarded?
Because someone was trying to train the chimp using primary or secondary
rewards of some type.
> The
> Ape it is not rewarded so it never twigs that we started with dirty
> dishes and ended with clean dishes.
Well, that's the problem. The perception system has to divide up the world
into "things" (which is the internal abstract representation of the state
of the environment). That decoding of the environment, as I've talked
about many times, also has to be adjusted based on value so we allocate a
higher resolution of decoding to the areas that at more important to us.
But it still means that even elements that have little to know special
value, will still be decoded to some resolution. We are not totally blind
to any high level detail, we are just selectively more aware of the details
that are more important to us. So a larger part of the network could end
up being allocated to decoding faces, and a smaller part allocated to
decoding rocks. So we end up with a higher resolution decoding of face
details, and lower resolution decoding of rock details.
How much resolution we end up with, depends not only on how important a
feature us, but how large the decoding network is to start with.
A human, might have enough decoding network to distinguish clean from dirty
plates, even before the network is taught to learn the value of diry plates
though conditioning. A chimp brain might not have enough network to "see"
the difference in plates simply because it has a lot smaller cortex than
huamns do (1/5 the size I think?).
So even before being trained to allocate more network to the state of food
serving plate, humans might be able to mimic plate cleaning when a chimp
can't because it has a high enough resolution network to see that
difference by default.
I like to think of this effect of conditioning as if it were applied to a
digital camera. The default mode is to allocate the same amount of space
across the image to each pixel (the way we build all our digital camera).
But if it could be conditioned by a reward signal, it would learn to
allocate more pixels to the parts of the picture which were important. So
if faces were important, it would allocate more pixels to the faces, and
take away pixels from the rest of the image, so that we would end up with a
high resolution image of the face, and a low resolution image of the
background. So even though you only have a million pixels to work with,
you end up getting higher resolution data on the parts of the image which
are more important, and lower resolution for the parts of less importance.
I think the same thing must happen in these learning networks (and the same
thing _does_ already happen in my simple pulse sorting network).
> It does however enjoy playing
> with dishes and soapy water. If you were to hide food under clean
> dishes and not under dirty dishes I suspect the Ape would learn to
> discriminate between the two and maybe learn to wipe off the dirt
> if it could associate the result, a clean dish, with a reward.
I've only heard about this second hand from you so I don't know what chimps
can actually be trained to do or not.
> current state (dirty dish)
> choose operator (wipe dish)
> desired state (clean dish)
Well, yes, that's one way to talk about it - but not what that yields
itself to a very useful implementation. That's a standard GOFAI approach
that fails to make it possible to be conditioned. To build a network that
can be conditioned, requires a very different approach. Though, whatever
approach you select, the above sort of thing has to be happening abstractly
in some way.
> > I don't think there's anything special about the perception
> > of clean dishes that isn't included in the general problem
> > of being able to perceive anything, from a face, to an ice
> > cream cone. It's all the same problem.
>
> This is where GOFAI can give some insight, for in trying to duplicate
> certain human skills, such as recognizing a face or inferring a goal
> (clean dishes) in the actions of another you get some idea of what
> the learning system has to achieve and thus a measure of its
> difficulty
> by how hard it is to duplicate. Calculus is easy, seeing is hard.
Yes, I think in effect that "seeing" is almost everything. However, what
most people working on perception problems never thinks of, is the reward
side of the problem.
As I've said many times before, I think our actions are best understood as
our perceptions. That is, we perceive when to raise our hand, and we
perceive when to lower it - etc. Perception and action are not two
separate problems here - they are one and the same problem. It's the only
problem there is to be solved to solve AI. It's the problem of how a
generic perception->action system learns how to map sensory data to
effector data though conditioning.
You can't solve either the perception or the action problem without solving
both because there is no solution to one, that doesn't also require the
other be solved at the same time. You can't train actions, without
perception, and you can't understand perception, without understanding it's
only purpose is to drive action.
If you try to ignore the action side of the perception problem, then there
is much you can't explain about how our perception is tuned to what is
valuable to us - which is also a way of saying how our attention is focused
on what is valuable to us. All question of value in this generic learning
side of the problem must be explained by conditioning by a reward signal.
Evolution hard-coded our primary rewards, but we have to learn the value of
secondary rewards though experience.
Something interesting I just thought of about the idea of warping pixels to
represent a higher resolution for the parts of the image that are more
important. It could be looked at as warping the pixels so that each pixel
covered an area of equal importance. That is, if face information was in
general twice as important as background data the the face pixels could
cover half the area, and end up with a the same value of information as the
background pixels what covered twice the data. So looking at the problem
like that, we could say the way the network needed to be adjusted, was so
that all pixels represented the same average value over time.
In my networks, I think of each signal generated in the network as the same
as a "pixel" because it represents some aspect of the environment. But
with this idea of making the signals have equal value, any signal that
showed itself as having higher value, would need to be adjusted so it had
less value, or covered less of the data. Which would mean in my pulse
sorting networks, to adjust the network to route fewer pulses to that
signal - which is oddly opposite of how I have been thinking about this
problem in the past because if looks like we are saying the network should
be adjusted to do less of what has worked in the past instead of more of
it. Interesting. Need to think more about that.
casey
>
> The article says HM was able to draw a detailed map of the home
> he moved to 5 years after the surgery. Considering how little
> he could learn, that was surprising.
Brain damage is never that precise as HM's parahippocampal gyrus
was still intact.
> That doesn't seem to be very accurate for Wearing. Being able
> to anticipate events in a movie is learning. But it's sure
> nothing like you describe as "doesn't remove the ability to
> lay down memories as to how to do something". HM on the other
> hand seems to have more learning ability left.
The actual combination of areas effected will vary between
individuals. I think you are being rather pedantic on this.
The point is that remembering episodes in your life is a
different kind of subject matter to learning how to do
something like tie a shoelace.
Memory of one kind of subject matter can be impaired without
it affecting memory of another kind of subject matter.
The suggestion is that the hippocampus is a module involved
in the storage of one kind of subject matter, spatial maps
and locations, and other modules are involved in the older
types of subject matter to be learned.
On the other hand emotional memories such as fear are handled
by the amygdala. In a simple case, such as a shock, the signal
is sent via the thalamus straight to the amygdala. This enables
the fast reaction and dose of adrenaline you experience when
you mistake a stick for a snake or someone presses a finger
into your back. However animals with a cortex also performs
a higher level analysis on the stimuli and the cortex sends its
results to the amygdala but that takes longer.
> Even with my belief that the cortex is fundamentally a
> homogeneous learning network, what you end up learning
> does end up being distributed to different parts of the
> network.
One way to find out what the cortex is used for is to remove
it from simpler animals and see what effect it has on the
things they can learn.
I don't know exactly what functions the circuits of the
neocortex performs but we do know it enables finer analytical
analysis of sensory data and finer synthesis of motor actions
for other brain areas.
> Every part ends up performing different functions. If you
> just go around and cut out random parts of the network, you
> are going to lose random parts of your behavior, and lose
> various associated abilities to re-learn even the type of
> behavior that was lost because you no longer have the right
> topology network to learn those behaviors.
You might ask how one piece of cortex receiving data from the
eyes decides to process color, another piece decides to
process texture, another decides to process motion and so on.
> My point however, is just because you call some behaviors
> "memory" and other behaviors "motor skills" and make up
> different labels for the different types of "memory' and
> different types of "motor skills" doesn't in any way
> disprove the theory that the cortex operates as one large
> generic operant conditioned learning system.
I wasn't talking about the neocortex hardware module, I was
talking about the hippocampus module. Being able to tell me
that you learned a skill yesterday is different to demonstrating
a skill you learned yesterday. The second one doesn't require
the ability to remember learning the skill.
> It was not evolution that figured out for us that dollar bills
> and coke bottles should be considered valuable in our perception
> system.
>
>
> Humans LEARN what is valuable by simple and well understood
> secondary reinforcement. It's well understood and fully
> implemented in our reinforcement learning algorithms already
> for the small scale problems. The only thing we are missing,
> is a workable implementation for high dimension problems.
And isn't the above example one of a high dimensional visual
input problem of recognizing a dollar bill and coke bottles,
which our pattern recognition system does so well, a system
that has been in action for other visual patterns in our
foraging past?
The simple example I gave previously was a program designed
to recognize the nine strokes 0 1 2 3 4 5 6 7 8 9. It turns
out it can also learn to recognize other strokes NEVER SEEN
BEFORE such as L T V J. The same applies to humans learning
to recognize dollar notes and coke bottles using a system
designed to recognize predators, trees, rivers and so on.
And the planning skills of a forager can be expated to the
planning skills for another kind of man made environment.
I have suggested that the innate visual preprocessing, which
would be required by our foraging ancestors, can indeed be
exaptated to tie knots, ride a bike, do calculus and so on.
Whereas calculus is hard because we do not have an innate
calculus module, learning to recognize 3D objects is easy
because we do have an innate object recognition system.
> In high dimension problems (all problems of interaction with
> the real world) we 1) don't know the true state from the
> sensory data 2) couldn't store value estimates for each state
> even if we did known it, and 3) will never get to visit most
> state->action pairs and won't get to visit any of them even
> twice.
Backgammon is such a high dimensional problem. The reason it
can be solved with the use of an ANN simply means that an ANN
measure of similarity is true for the high dimensional space
of backgammon states. In the case of chess a small difference
in the input state may require a completely different move.
In the case of Go it may require a visual analysis which is
easy for us but we haven't figured out how to evolve or build
such an analytical visual preprocessor for Go.
I see these low level module's functions as perhaps
impossible for any generic reinforcement learning module
to learn within the time and resource limits of any
practical system and that is the reason I don't believe
it has all been replaced by a generic learning module
although I have no issue with the fact humans learn
ALL their high level academic and social behaviors.
> The perception system has to divide up the world into "things"
> (which is the internal abstract representation of the state
> of the environment).
Yes we have to decode the environment into things. We actually
decode the environment into a set of basic things and this is
perhaps done for us by an innate preprocessing system, providing
something simpler than the high dimensional input, for us to
learn about. Our higher level learning system doesn't have to
deal with the onslaught of data coming through our senses
instead it deals with a simplified representation of the "world
out there" in the form of objects, actions, causality, substances,
surfaces and so on.
JC
Curt Welch
> On Dec 29, 1:12=A0pm, c...@kcwc.com (Curt Welch) wrote:
> >
>
> > The article says HM was able to draw a detailed map of the home
> > he moved to 5 years after the surgery. Considering how little
> > he could learn, that was surprising.
>
> Brain damage is never that precise as HM's parahippocampal gyrus
> was still intact.
>
> > That doesn't seem to be very accurate for Wearing. Being able
> > to anticipate events in a movie is learning. But it's sure
> > nothing like you describe as "doesn't remove the ability to
> > lay down memories as to how to do something". HM on the other
> > hand seems to have more learning ability left.
>
> The actual combination of areas effected will vary between
> individuals. I think you are being rather pedantic on this.
>
> The point is that remembering episodes in your life is a
> different kind of subject matter to learning how to do
> something like tie a shoelace.
Sure, but you seldom talk like that. You normally talk is if the memory
system itself was a different _type_ of system, and that as you might
suspect, is what raises objections from me. There is no induction in the
data to support the idea memories of past events is a different type of
brain function than memory of past actions. There is only the rather
obvious indication that different parts of the brain perform different
functions.
> Memory of one kind of subject matter can be impaired without
> it affecting memory of another kind of subject matter.
Yes, of course, if you wipe out the grandmother cell you impair the ability
to have memories of grandmother without impairing the ability to have
memories of apple pie.
> The suggestion is that the hippocampus is a module involved
> in the storage of one kind of subject matter, spatial maps
> and locations, and other modules are involved in the older
> types of subject matter to be learned.
Yes, that's true. But again, there's no indication that the _type_ is
different.
Am I remembering it correctly that the hippocampus is reaully just the
inner edge of the cortex?
> On the other hand emotional memories such as fear are handled
> by the amygdala.
Yeah, that's clearly something different because it's not part of the
cortex.
> In a simple case, such as a shock, the signal
> is sent via the thalamus straight to the amygdala. This enables
> the fast reaction and dose of adrenaline you experience when
> you mistake a stick for a snake or someone presses a finger
> into your back. However animals with a cortex also performs
> a higher level analysis on the stimuli and the cortex sends its
> results to the amygdala but that takes longer.
Yeah, no clue what's going on there. I don't believe it's all that
relevant to AI. It's highly relevant to fully understanding humans, but
not AI.
> > Even with my belief that the cortex is fundamentally a
> > homogeneous learning network, what you end up learning
> > does end up being distributed to different parts of the
> > network.
>
> One way to find out what the cortex is used for is to remove
> it from simpler animals and see what effect it has on the
> things they can learn.
Have they done that? When sections of the cortex are damaged in humans by
stroke or injury, the typical result is the lack of major and obvious
mental functions and behavior.
> I don't know exactly what functions the circuits of the
> neocortex performs but we do know it enables finer analytical
> analysis of sensory data and finer synthesis of motor actions
> for other brain areas.
Well, that of course is 100% consistent with my theories. When I try to
build AI my entire networks act as a perception system that decode the raw
sensory data into a alrge set of higher resolution features.
A problem with messing with animals is that it's clear the learning network
(cortex) was added by evolution later. Before it came along, behavior had
to all be hard-wired instead of learned. So evolution was forced to find a
way to make the generic learning systems work in parallel with the older
hard-coded behavior systems. Humans still have a few instinctive
hard-coded behaviors in us, though they are mostly overridden by the
learning system. A few, like our heart beat, and probably some aspects of
breathing, and digestion, are not under the control of the learning system
at all. But in lower animals, they seem to have less learning powers, and
more hard-wired behaviors. If the cortex is the new adaptive learning
module, and you disable it, what you end up with, is the lower level hard
wired instinctive circuits still at work.
So what such an experiment is likely to show, is more about how crafty
evolution was when it figured out how to use the two systems in parallel,
then you are to find out what the cortex on it's own can do.
To create strong AI, I don't think we need to include _any_ of those lower
level hard coded behaviors. We only need to include the generic adaptive
learning controller. In practical applications of AI, we might very will
use the same tricks evolution did and hard code some stuff, and then add an
adaptive controller on top of the hard coded behaviors just because it
saves money, or power, or size. But not becuase it's needed to make the
machine intelligent.
> > Every part ends up performing different functions. If you
> > just go around and cut out random parts of the network, you
> > are going to lose random parts of your behavior, and lose
> > various associated abilities to re-learn even the type of
> > behavior that was lost because you no longer have the right
> > topology network to learn those behaviors.
>
> You might ask how one piece of cortex receiving data from the
> eyes decides to process color, another piece decides to
> process texture, another decides to process motion and so on.
Well, there are two questions there. One is how would the parts of a
generic learning system distribute information though the network, and the
other, is why did the visual cortex distribute it the way it did.
My networks already do just the sort of thing you are talking about. They
automatically slice up the data into different features, and process the
different types of features in different parts of the networks.
They don't explain how the brain ended up with it's configuration of
features across the visual cortex, but they explain how features in general
can be distributed across the network automatically. How the brain ended
up with it's configuration is both a function of whatever sort of generic
adaptive wiring processes might be at work in the brain, along with generic
pre-wiring of the topology of those sections of the cortex.
But certainly when I try to figure out how to make network nodes that solve
the perception problem, looking at what type of features they produce, and
how those features would end up distributed though the network, is all
obvious stuff I'm always looking at as I think about different potential
designs.
> > My point however, is just because you call some behaviors
> > "memory" and other behaviors "motor skills" and make up
> > different labels for the different types of "memory' and
> > different types of "motor skills" doesn't in any way
> > disprove the theory that the cortex operates as one large
> > generic operant conditioned learning system.
>
> I wasn't talking about the neocortex hardware module, I was
> talking about the hippocampus module.
Maybe I've got my brain anatomy confused, but isn't the hippocampus just
the inner bottom edge of the cortex?
> Being able to tell me
> that you learned a skill yesterday is different to demonstrating
> a skill you learned yesterday. The second one doesn't require
> the ability to remember learning the skill.
Right. But again, there is no indication that there is a different _type_
of learning hardware at work for those two tasks - only that different
parts of the cortex are used for those two different tasks.
> > It was not evolution that figured out for us that dollar bills
> > and coke bottles should be considered valuable in our perception
> > system.
> >
> >
> > Humans LEARN what is valuable by simple and well understood
> > secondary reinforcement. It's well understood and fully
> > implemented in our reinforcement learning algorithms already
> > for the small scale problems. The only thing we are missing,
> > is a workable implementation for high dimension problems.
>
> And isn't the above example one of a high dimensional visual
> input problem of recognizing a dollar bill and coke bottles,
> which our pattern recognition system does so well, a system
> that has been in action for other visual patterns in our
> foraging past?
Yes. And your point is?
> The simple example I gave previously was a program designed
> to recognize the nine strokes 0 1 2 3 4 5 6 7 8 9. It turns
> out it can also learn to recognize other strokes NEVER SEEN
> BEFORE such as L T V J. The same applies to humans learning
> to recognize dollar notes and coke bottles using a system
> designed to recognize predators, trees, rivers and so on.
> And the planning skills of a forager can be expated to the
> planning skills for another kind of man made environment.
Sure, but if a tree recognize can be used to recognize coke bottles, then
it ISN'T A TREE RECOGNIZER! It's a GENERIC pattern learning machine.
You can't have it both way's John. It's not a HARD CODE TREE RECOGNIZER.
IT's a high quality GENERIC sensory pattern recognition hardware.
And yes, it was developed to deal with the wild, but it was still developed
to be a GENERIC pattern recognizer so that evolution didn't have to waste
any time hard-coding tree recognizers, and rock recognizers, etc.
> I have suggested that the innate visual preprocessing, which
> would be required by our foraging ancestors, can indeed be
> exaptated to tie knots, ride a bike, do calculus and so on.
Yes, but the fact that it can be done, proves it's not hard coded - it
proves it's a strong generic system that can deal with a very wide range of
sensory patterns never seen before by the hardware.
> Whereas calculus is hard because we do not have an innate
> calculus module, learning to recognize 3D objects is easy
> because we do have an innate object recognition system.
Calculus is hard because it's highly abstract. Too abstract for some
brains.
> > In high dimension problems (all problems of interaction with
> > the real world) we 1) don't know the true state from the
> > sensory data 2) couldn't store value estimates for each state
> > even if we did known it, and 3) will never get to visit most
> > state->action pairs and won't get to visit any of them even
> > twice.
>
> Backgammon is such a high dimensional problem. The reason it
> can be solved with the use of an ANN simply means that an ANN
> measure of similarity is true for the high dimensional space
> of backgammon states. In the case of chess a small difference
> in the input state may require a completely different move.
> In the case of Go it may require a visual analysis which is
> easy for us but we haven't figured out how to evolve or build
> such an analytical visual preprocessor for Go.
>
> I see these low level module's functions as perhaps
> impossible
Which is clearly why I'm never going to get you to work on it. It's too
hard for you to understand.
> for any generic reinforcement learning module
> to learn within the time and resource limits of any
> practical system and that is the reason I don't believe
> it has all been replaced by a generic learning module
> although I have no issue with the fact humans learn
> ALL their high level academic and social behaviors.
Well we went over this in the other thread. By looking at what humans can
and do learn, it's trivial to prove that the brain has solved the very
problems you claim can't be solved. The fact you can't understand that,
just means I'm wasting more of my time trying to get you to understand how
to solve it or to get you to help solve it.
> > The perception system has to divide up the world into "things"
> > (which is the internal abstract representation of the state
> > of the environment).
>
> Yes we have to decode the environment into things. We actually
> decode the environment into a set of basic things and this is
> perhaps done for us by an innate preprocessing system, providing
> something simpler than the high dimensional input, for us to
> learn about.
Well, the raw sensory inputs themselves are already a great simplification
of the high dimension state of the universe they are sensing. From there,
the brain doesn't really simplify anything - it actually creates more data
to work with like the 1 to 400 expansion of signals that happens in the
visual cortex if I'm remembering by numbers correctly. So this whole angle
of yours that the "problem which can't be solved" is solved by "evolution
giving us innate hard ware to simplify the data" is just absurd. The data
never does get simplified. It stays in a high dimension complex form all
the way though the brain and it comes out the other side as high dimension
complex data driving our actions.
The high dimension learning problem is not solved by simplifying it to a
low dimension problem. It's solved directly in it's high dimension form.
The only sort of simplification that happens is when we produce language
behavior to try and describe things. That's always a gross simplification
of the high dimension data that it attempts to describe.
> Our higher level learning system doesn't have to
> deal with the onslaught of data coming through our senses
> instead it deals with a simplified representation of the "world
> out there" in the form of objects, actions, causality, substances,
> surfaces and so on.
That's just not true at all. It's only true for the high level of
"langauge" and nothing else in the brain. When we attempt to use langauge
to describe what the brain is doing, or plan our actions, or describe an
event, or give labels to the "things" out in the world, we have very much
performed a gross simplification of the raw input data. But nothing like
that actually ever happens inside the brain as it's processing the data -
except of course, when the end result is the production of langauge. But
in that sense, all the output behavior ends up being a gross reduction and
simplification of the input data because we have far fewer outputs than we
have inputs to the brain.
The learning problem never gets to work on a "simple" version of the data
because the "simple" version you keep talking about just doesn't exist
anywhere in the brain. The high dimension learning problems the brain is
able to solve, can't be explained by simplifying it. It's solved by a
system that solves it as a high dimension problem - which I've already
roughly outlined many times how it can be done.
casey
> casey <jgkjca...@yahoo.com.au> wrote:
> [delete about brain]
> I don't believe it's all that relevant to AI.
Its relevance is in regards to a practical implementation.
>> You might ask how one piece of cortex receiving data from the
>> eyes decides to process color, another piece decides to
>> process texture, another decides to process motion and so on.
>
>
> Well, there are two questions there. One is how would the parts
> of a generic learning system distribute information though the
> network, and the other, is why did the visual cortex distribute
> it the way it did.
>
>
> My networks already do just the sort of thing you are talking
> about. They automatically slice up the data into different
> features, and process the different types of features in
> different parts of the networks.
You make claims about what your networks ARE doing but without
allowing anyone to test those claims they are meaningless. Are
you claiming that with a visual input one part automatically
takes on color, another motion, another texture and so on?
You gave me the impression it was all mixed together in your
nets, the latest of which I haven't seen.
> Sure, but if a tree recognize can be used to recognize coke
> bottles, then it ISN'T A TREE RECOGNIZER! It's a GENERIC
> pattern learning machine.
>
>
> You can't have it both way's John. It's not a HARD CODE TREE
> RECOGNIZER. IT's a high quality GENERIC sensory pattern
> recognition hardware.
>
>
> And yes, it was developed to deal with the wild, but it was
> still developed to be a GENERIC pattern recognizer so that
> evolution didn't have to waste any time hard-coding tree
> recognizers, and rock recognizers, etc.
Depends how you want to define "generic" as it doesn't have
the ability to recognize ANY pattern. Bottle tops and coke
bottles have the same features as natural objects.
http://www.owlnet.rice.edu/~psyc351/Images/MinskyPapertSpirals.jpg
Unlike bottle tops and trees you can NEVER learn to recognize
the two different types of spiral patterns just be looking at
them. You don't have the innate parallel machinery to do that.
In order to "see" we have parallel machinery which makes
assumptions about how the real world is put together to the
extent that we will see things that aren't really there.
We will see virtual edges, 3D in a 2D image, false movements,
things that are the same size or color as being a different
size or color and no amount of learning will stop us seeing
them that way because our INNATE visual system is FIXED to
see them that way.
The kinds of illusions we experience flag possible assumptions
used by our innate visual system to decode the retinal images
into a representation of what is out there.
>> Whereas calculus is hard because we do not have an innate
>> calculus module, learning to recognize 3D objects is easy
>> because we do have an innate object recognition system.
>
>
> Calculus is hard because it's highly abstract. Too abstract
> for some brains.
And you think seeing isn't highly abstract? Isn't recognition
a process of extracting common qualities from specific examples?
Recognition of an object means abstracting out what is common in
a set of examples of that object. And if you think that is easy
why don't you have a go at trying to duplicate it.
The question you need to ask is why calculus is easy to program
and "seeing" is not easy to program. If seeing is easier than
calculus why can't we solve the seeing problem?
The ability to make those visual abstractions is innate whereas
the other kinds of abstractions are not. We cannot for example
"see" 4D objects because we lack the hardware to do this. This
is why it is hard visualize a 4D object.
Calculus is NOT HARD it is just not interesting to most people.
To quote Richard Dawkins, "Whenever I feel intimidated, [by math]
I always remember Silvanus Thompson's dictum in Calculus Made
Easy: 'What one fool can do another can'.
If you really want to understand calculus you only have to put
the time into it. If you really want to understand vision you
will have to put a lot more time into it - so far human vision
is not understood - it is THAT hard.
If we didn't have the parallel hardware with its build in assumptions
to learn to recognize natural objects and had to use the methods used
to understand calculus, or discriminate the two types of spirals, we
would find vision impossible to do in real time!
> The high dimension learning problem is not solved by simplifying
> it to a low dimension problem.
>
> It's solved directly in it's high dimension form.
I wonder what you really mean by that last statement?
> It's solved by a system that solves it as a high dimension problem
> - which I've already roughly outlined many times how it can be done.
It's physically impossible for the brain to match the high dimensional
problem and not required as shown by chess playing programs that
solves
winning a game of chess by using simplifying assumptions called
heuristics
without the need to solve the high dimensional problem.
What exists physically in the brain is not a high dimensional version
of the real world but rather a simplified version of that world.
JC
casey
Curt wrote:
> You can't have it both way's John. It's not a HARD CODE
> TREE RECOGNIZER. IT's a high quality GENERIC sensory
> pattern recognition hardware.
>
>
> And yes, it was developed to deal with the wild, but it was
> still developed to be a GENERIC pattern recognizer so that
> evolution didn't have to waste any time hard-coding tree
> recognizers, and rock recognizers, etc.
You have responded this way a number of times as if you
imagine I think every object has its own recognizer?
You claim humans have replaced the old learning system with
a new generic learning system and that is why we can learn
all our high level behaviors. I am suggesting that our
learning system is nothing more than an enhancement of the
old one and it has its limits and is innate.
Only being able to recognize natural objects is not one of
its limits! It turns out the SAME mechanisms allow it to
recognize coke bottles and bottle tops without the need
to add a new generic learning system on top.
What about our complex motor acts? The idea of using a car
doesn't occur to an animal. It is not a lack of motor learning
skills or visual processing skills on their part. Have you
seen dogs using skate boards or Apes steering a tractor?
Instead of looking for the sudden appearance of a magic new
generic learning system you might consider it in terms of
an evolutionary enhancement of the brain to make use of the
way our bodies have evolved.
JC
pataphor
> In my networks, I think of each signal generated in the network as the same
> as a "pixel" because it represents some aspect of the environment. But
> with this idea of making the signals have equal value, any signal that
> showed itself as having higher value, would need to be adjusted so it had
> less value, or covered less of the data. Which would mean in my pulse
> sorting networks, to adjust the network to route fewer pulses to that
> signal - which is oddly opposite of how I have been thinking about this
> problem in the past because if looks like we are saying the network should
> be adjusted to do less of what has worked in the past instead of more of
> it. Interesting. Need to think more about that.
Great. So you just have reinvented dreaming. It is the readjusting of
weights that is necessary to make the network stay effective. As such it
is the opposite of reinforcement learning. If you continue at this pace
someday you will discover consciousness, which is another layer above
these automatic systems. In this layer it is possible to completely
rotate weights to other positions, which is about as far away as one can
get from reinforcement learning without directly uploading and
reprogramming brains.
The way you keep talking about RL is like insisting that human biology
is essentially plant biology, because in the end all light coming from
the sun is converted by plants and all other biology (or at least most)
is dependent on that source of energy one way or another. But such
insistence is highly unproductive. At some point we have to switch from
talking about plants to talking about animals or humans. It is the same
with talking about materialistic versus nonphysical phenomena.
The universe may be evolving to some new kind of order, but in order to
be able to talk about these phenomena we have to switch terms as new
memetic constructs appear. It is not unlikely humans will end up in much
the same situation as rain forest Indians, as their ecosystem is
destroyed by large industrial consortia that they finance themselves
because they keep buying their products and do not oppose their
influence on the government and the election process. War zones are also
an ideal place for developing technology that will ultimate lead to our
own destruction.
What we need to do is to bypass our brain's control structures that
literally put us to sleep so we can't do lucid dreaming, or that induce
us to follow our instincts -- like greed and status -- which will
deliver us in the hands of these mega corporations. I suppose the mega
corporations will fall victim to Internet technologies in the same way
we individual humans are prey to them. See for example the recent
banking crisis.
The solution is always the same, only follow your feedback the first
time, but as soon as a pattern emerges that can be preyed upon, switch
to a new behavior. That way we will be able to always gain the positive
side of the feedback without ever experiencing the hammer.
At least until we die. Maybe death is the ultimate liberator of all the
energies we have accumulated in this guerrilla technique of saving our
energy from larger entities that form out of them and that try to
enslave us. But in order to do so we have to find a way towards the more
and more abstract, and lingering at old terminology and outdated ways of
thinking in order to accomplish some grand theory where everything fits
top-down and bottom-up is nothing more than inviting a blind and quick
and purposeless death, instead of a grand enlightenment and unification
with the cosmos by going nova.
P.
'Happy 2010'
Curt Welch
> On Dec 24, 3:24=A0pm, c...@kcwc.com (Curt Welch) wrote:
> > pataphor <patap...@gmail.com> wrote:
> > > Yevgen Barsukov wrote:
> >
> > > [...]
> >
> > > > We can look at two extremes:
> >
> > > > A) purely instinct driven animal, which is optimized
> > > > to propagate his genes
> >
> > > > OR
> >
> > > > B) purely cultural programing driven animal. Basically monkey
> > > > infested with cultural virus that cares not at all about
> > > > propagating of monkey genes, but just about propagating itself
> > > > onto other monkeys.
> >
> > > > Real humans are somewhere between A and B and there are probably
> > > > many variations with different weights. In overall during history
> > > > of human civilization there have been a gradual drift from A to B.
> >
> > > A very interesting view. And on the surface it seems highly credible.
> > > But what I am missing here is a sense of identity, or self. It is
> > > like this view is completely from the outside and not giving us any
> > > meaning or personal space. That's why it scares the hell out of me,
> > > however logical and credible it sounds. Are you sure you are not
> > > missing some third component you forgot to include in your model?
> >
> > > P.
> >
> > Well, for my 2 cents on this, I agree that we are the result of both a)
> > respo=
> nse
> > to nurture is even mostly under the control of the "will of the
> > cultural virus" is a huge failure to understand how the propagation of
> .
>
> Lets stop right there. You use the word "meme" in many of
> your statements, and that is not what I am talking about,
> at least not in the sense how Dawkings defined "meme". He defined it
> is some cultural content which is irrelevant for purpose of Life and
> propagates just because it has effective reproduction mechanisms and
> because the immune system of actual Life it parasites on did not yet
> figure out how to stamp
> it out.
Well, I'd have to reread the book, but I'm fairly sure he doesn't define it
that way. I'm fairly sure he defines it as _any_ element of human behavior
that can drift from human to human by learing - not just the useless
behaviors but all the good and bad ones as well. Whatever his definition
is, my usage is that it's any behavior that can drift from person to person
though learning. Though it might be easier understood (though less
accurate) as any belief that moves from person to person.
Reading http://en.wikipedia.org/wiki/Meme I see it doesn't support your
claim.
> Being a zoologist involved with mechanisms rather than reasons,
> Dawkings is not focusing fundamental thermodynamical nature of Life
> and so it is all a game of reproduction to him.
Right, He's only addressing the causality chain starting at the DNA
instead of starting at physics.
> In reality Life and
> all its forms are just such a necessity in certain chemical
> environments as a whirlpool in a stream when water flow exceeds
> certain rate.
Yes, I agree. And I am responding to these articles out of order so I've
already commented some on this in replies to your other articles.
> And it does not matter if this is Life which decision programming is
> stored in DNA or it is stored in neural states or it is stored in
> scrolls of papyrus - point is, Life will take whatever form is
> thermodynamical optimal at this moment and will be most capable of
> accelerate entropy increase.
Right, but the path is oddly constrained - just like the actual path of a
rock rolling down a hill is oddly constrained by it's shape, and the odd
things it ends up running into on the way down the hill. Even though
gravity is the over weling force in control of the rock and gravity is
pulling it straight down hill, the rock doesn't roll straight down hill.
> When I talk about cultural programming, I mean it with the same level
> of inherent belonging to Life as genetic programming. Humanity is just
> a Life form, some of which is not stored in DNA, but stored in books,
> CDs and Usenet archives.
Well, if you mean the real life form of human could be thought of as
everything we build, like the spider web should be thought of as part of
the spider, that's fine. But if you mean the information is in the book,
you have to be very careful not to make the error most people make. The
meaning of the signs in the book is not in the book. It's in the human
brain that wrote the book, and in the human brain that read the book, but
not in the book itself. Much of what people talk AS IF it were in the
book, is not at all in the book.
We say this book has in it a story about Odysseus and his voyage. But the
book contains no such story. It only contains ink marks on paper. Those
ink marks have the power to create the story in the brain of the reader,
when they look at the ink marks, but the story only exists in the brain of
the reader, it never existed in the book. I was just jumping on Tim in my
last post about making this same error by trying to claim that our human
memes exist in computers. (Not yet). When a computer can read the
Odyssey and understand it, then those memes will have moved from a human to
an AI. But they never existed in the book.
> What we see around us - houses, cars, large
> hadron collider and cruiseships are phenotypes of this Life-form.
Yeah I agree with that. They are the products of evolution as well. Or
from the Dawkins DNA perspective, they are all part of the transport and
protection system the DNA built around itself.
> > The genes have total control over what memes it will allow the body to
> > except. =A0The memes have very little control over the genes
> > (currently). There is no doubt that currently, the genes are in control
> > here - not the memes. =A0The memes are emergent behaviors that the
> > genes allow to happen=
> .
> > When the genes create emergent meme behaviors in our society, and those
> > memes fail to do a good job of keeping the genes alive, then those
> > genes (and the memes they allows to emerge) all die off together.
>
> Genes are in control of physiological functioning of the body
> (including its input/output system such as hearing, vision etc).
> What it can not control is what these systems are hearing and seeing.
That's right. It has some control over it, but not a lot.
> Here is one extreme example - A baby can be placed in a dark room
> without any information input (choice A) or into a room full of toys,
> parents, aunts and other siblings (choice B). What it will learn is
> completely determined by these choices. But - neither of these two
> choices in controlled by the genes in the slightest. What is it
> controlled by?
Right, humans don't perform sensory deprivation experiments on babies. And
why is that? IT'S BECAUSE THE GENES BUILT US TO NOT _WANT_ TO DO THAT. we
have no free will when it comes to what we want. The heart wants what it
wants they say. That's because our primary wants and desires are hard
coded into us by our DNA, and they hard-coded a statistical system that
estimates the value of sensory clues based on hard coded statistics. And
all our learned behaviors, are statistical derivations of those wants and
desires. We have no control over any of it. Our behavior is
pre-determined by our DNA. Our DNA built a system to statistical calculate
optimal behavior - and what we do is what that statistical calculation says
we do. We have no control over any of if. Free will is an illusion
created from ignorance.
What we talk about as "our control" and "free will" really isn't true
causal control of anything. It's rationalizations we tell ourselves
because it's optimal behavior according to the statistics that actually
control us.
> It is controlled by cultural programming of the parents and other
> programmed monkeys this baby is in charge of.
Yes, when the environment changes, the behavior changes because a change in
the environment creates a need to change behavior to remain optimized per
the behavior optimization system built into us.
> So is everything else what baby will learn, which in turn will define
> what it will teach to its own babies and so on and so forth. By
> controlling the environment from the time when brain is starting to
> learn, cultural program is the master of the content, while genes are
> just responsible for the execution of whatever content is presented by
> obediently delivering sensory inputs carefully selected and controlled
> by cultural programmers down the throat of open and receptive blank
> brain.
Well, I'm glad you think it works that way. It doesn't. The brain has a
fairly straight forward behavior optimizations system called operant
conditioning. But the value of a stimulus IS not controlled by the
environment. It's not the environment that makes the food "Good" and the
dog shit "bad". Its our DNA. And it's the DNA's assignment of value to
these environmental conditions that causes HOW and in WHICH DIRECTION our
behavior changes - not the environment.
If what you think was true, we could do a little experiment. In one
environment we raise a baby in an environment where everyone around him
ate, and used, and talked about, normal human food. In the second,
everything was exactly the same, except the food was rocks and sand.
If what you say is true, the baby would act exactly the same in both
environments. He would like the rocks just as much as he liked his
cookies.
Of course the baby won't be "fooled" by the rook food environment because
his wants and desires don't come from the environment. They come from his
DNA. And likewise, the behavior doesn't come from the environment directly
either. They come from a statical process that attempts to optimize the
behavior to maximize the systems needs in any given environment.
The only memes that move from one person to another, are the ones which are
expected to improve the needs of the person the meme enters. Memes are
allowed to enter only if the pass the gates of the optimization function -
the optimization process created by, and defined by, OUR DNA.
> Note that such receptiveness itself has been evolutionary reinforced
> because babies that would not learn
> got eaten by hyenas, fall of the entrance of the cave, got burned by
> the fire or otherwise died from being incompatible with culture-
> defined phenotype of humanity and so did not propagate their genes.
> Those the very genes that you think me cause opposition to
> cultural programming have been eliminated during last 500 000 years of
> culture-defined evolution of monkeys. Only the
> "culturally receptive" monkeys have been selected and became
> the backbone of our modern gene pool.
>
> Admittedly, there are many behaviours (instincts) that have peen pre-
> programmed genetically at the time that precedes formation of cultural
> programming system and are still in many cases duplicating or
> contradicting cultural programming. But honestly - make an experiment
> and analyze yourself for say last 2 hours: how many of actions you did
> were caused by instinct and how many by cultural programming?
You don't understand human learning.
Yes, most my behaviors can be called a direct result of cultural
programming. But why is that? Is it because humans take into them any
idea their culture offers them? Nope. It's because we have a hard wired
innate meme evaluation system working in us. When we see someone else do
something that's useful what do we do? We say, "damn! good idea! I'll do
that too!".
With a few billion humans, all with almost identical meme evaluation
hardware, constantly on the lookout for better behaviors, what do you
expect to happen? Someone on the other side of the world finds one by
accident, and the behavior becomes a meme in that persons brain and is used
over and over. Then other people see him using this behavior, they
evaluate it as good as well, and they copy it. Before long, the behavior
as spread across the world.
Why did that meme spread though the culture? Because it was a "selfish
meme" that found some magical way to spread? Shit no. It spread because
our DNA meme evaluator liked it - liked it more than whatever else it was
using before the new meme was accepted.
If there were no hardwired meme evaluators, what would control the spread
of the memes? Why would one have more power to spread then another? If a
human trips and falls face first into the mud, why isn't that "fall face
first into the mud" meme not spreading though the culture faster than the
"watch where you walk" meme?
It's because the power of a meme to spread, has nothing to do with the
meme, and everything to do with the value assigned to the meme, by the
innate meme evaluation hardware built into us by our DNA.
> ....
> I am sure you found that excluding purely physiological functions like
> breathing, eating and urinating most of your actions were caused by
> cultural programming and have no instinctive basis.
Nope. I argue that we are in fact blank slates, but not as blank as you
would suggest we are. Though the behavior of typing on the keys like this
to post a Usenet message in English is a behavior I learned directly from
my environment, the reason I learned it, and the reason I use it, is
because my hard-wired behavior evaluation system determined it was the
optimal behavior for me based on my environment and my past experience and
on the prime needs and values hard wired into me - all by my DNA.
> > Humans can't be forced to accept any meme. They only accept the memes
> > tha=
> t
> > are proven to useful to prevent harm to the body or harm to our
> > reproductive success - which is a requirement created by the genes.
> > =A0Th=
> ey
> > are the gate keepers of what memes are allowed to take hold in us and
> > always have the upper hand.
>
> I think this statement is not even true to actual memes that Damkins
> defined. The whole point of memes is that they posess mechanisms of
> propagation that are unusually sucessful (such as deviation in a
> lyrics of a hymn that is caused by typical accustics of a chappel).
> As for applicability to cultural programming, as I have shown above,
> baby's genetic make-up has no say whatsoever what programming it will
> accept, it just defines the
> structure of machinery that is being USED in this programming.
> And this machinery is pretty much the same, except of some
> pathological cases where it is broken.
[repeat everything I said above again] :)
> >
> > We don't accept just any meme we hear from society. =A0Our brain judges
> > t=
> he
> > value of the different memes it is exposed to, and accepts the ones it
> > evaluates as being the best - the most useful. =A0But the evaluation
> > syst=
> em
> > the brain uses, is specified by our genes. =A0Which is why the genes
> > have ultimate control over the memes - not the other way around in any
> > sense.
> >
> > Learning systems are flexible in that they have the power to learn a
> > wide range of very odd behaviors - like swinging a stick at a little
> > white bal=
> l
> > in order to maximize our odds of survival (Tiger Woods). =A0But they
> > are =
> not
> > so flexible that they will learn just anything they are exposed to.
> > =A0Th=
> ey
> > don't work like a camera just capturing and duplicating _anything_ it
> > see=
> s.
> > It's nothing like that.
>
> I quite agree. It is not simple "mimicking" that is why I don't like
> work meme. Cultural programming technology is just as complex and
> incomprehensible as genetic programming in DNA - it has been evolving
> ever since language was created or maybe even before, and so it has
> countless tricks and optimizations both to store and to imprint useful
> for humanity information which we are not nearly aware of. It is also
> rapidly changing.
It's not hard to understand any of it in the broad since once you gain a
valid understanding of how learning works and where our behavior comes
from.
> For example in the past some empirical optimizations found by deadly
> trial and error experiments (for example that you need to wash hands
> before eating) had to be stored in religious stories and fairy-tales,
> now many of such optimization have well understood reasons (existence
> of bacteria, known since very recent in historical therms discovery)
> and so can be imprinted through direct causal explanation that is
> based on education. Interestingly, in both cases end result is the
> same - monkey is going to wash its hands and not die from dysentery,
> even though methods to achieve this results appears opposite.
Well, let me use your example to talk about out learning works and why
those complex memes managed to survive.
When we learn behavior, it comes to us though a complex statistical
behavior evaluation system. Our brain only produces the behaviors that it
has calculated to be the best behavior possible at any instant in time. In
effect, it's compared all possible behaviors, ranked them all, and picked
the one with the highest rank based on a highly complex statistical ranking
system.
If we choose to mimic some behavior we see someone else do (like hand
washing) or not, is a function of what that complex system selects. It
will be based on thousands or millions of parameters - it's far too complex
for a human to understand, and if the reasoning for a behavior choice could
be described, it would basically require enough words to produce a dump of
the current wiring of the human brain. So no amount of rationalizing why
we make some choice can every be anything but the tip of the iceberg in
terms of the real reason a human made a behavior choice. However, we can
look for, and attempt to identify some major influences in any action
(which we do all the time).
Why would we choose to mimic hand washing if some religion told us to do
it? Why would following the advice of some religions view be something our
brain learned as a "good" thing to do?
Well, mostly it comes down to probabilities. How many similar things have
we mimicked, and how did that work out for us in the past? The answer is
that we learned very early that at times it's good to mimic, and at times
it's bad to mimic. Are brain develops (as it does for all behavior) a very
complex evaluation function to estimate the expected worth of mimicking any
behavior based on any, and all, sensory clues it has to work with. Some of
those clues come from a direct evaluation of a behavior based on our
understanding of the behavior itself. It the action makes blood leak out
of a human, our evaluation system warns us away from the behavior. If the
action leads to the other person getting money, we see it a better
behavior.
But other clues come from the source of the behavior - in who is doing it.
If we mimic our parents when we are young, life gets better for us (in
general). This is something most of us learn for obvious reasons. These
clues stick with us typically for life. Anyone that our system evaluates
as being similar to our parents, become high value targets to mimic. And
every time we mimic one of these people, it creates another test of our
mimic quality evaluator. The more "luck" we have mimicking a person, the
more of their behavior we are likely to mimic in teh future.
Why does a religious meme of hand washing spread? Because it exists in
people who have established themselves as worthy of being mimicked.
And why does some set of cultural beliefs survive? - because either 1) it's
harmless crap that got lucky, or 2) like real genes, the meme is actually
useful to the survival of the human and their DNA.
Memes that are harmful to survival get evolved out of existence. The more
harmful, the faster they are removed. Harmless behaviors can live for a
very long time simply by random chance.
Memes that survive the test of time, prove their value. The older the
meme, the more likely that the meme is actually valuable to survival.
Our meme evaluation system picks up on this fact indirectly. The people
that are keepers of old wisdom, are typically good to mimic. Not because
they _say_ they known the truth, but by all these indirect clues our brain
picks up on, such by who else mimics their behaviors, and how close their
behaviors are to the behaviors we currently believer are the "good"
behaviors. The more they act like us, the more likely their behaviors
which are different from us, are good behaviors for us.
The location we see a behavior, and the time, also give us clues about
whether the behavior might be good to mimic.
The net result is that our environment is full of useful clues that gives
us a huge wealth of information about the odds of whether mimicking a
behavior will be good for us or not. Our brain uses all those clues and
all it's past experience in it's calculations, but we have high level
understanding of almost none of the clues the brain will use.
So whether that meme of hand washing takes hold in us, is a complex
function of a life time of past experience accepting, or rejecting,
different memes based on a huge volume of environmental clues available to
make that calculation with.
By using such a wealth of data to make value estimates, the brain is able
to make some very wise choices about the worth of mimicking a given meme
(even without the help of a high level logical and rational analysis of the
choice).
Because the brain uses complex probability functions to accept and reject
behaviors, the function can make some big errors - even if on average, the
choices are highly accurate and highly useful to our survival. Some memes
will survive, simply because they are able to fool the gatekeeper. If the
meme is actually harmless (or mostly harmless), fooling the gatekeeper is
harmless. The error won't get corrected very quickly (if at all). Urban
myths and superstitions for example are mostly harmless memes that fool the
gatekeeper. But the memes that are actually bad, tend to quickly re-train
the gatekeeper. Fool me once, same on you, fool me twice, you must be
really good at tricking my meme gatekeeper!
The bottom line however is that good memes spread for good reasons,
harmless memes spread but our harmless, and bad memes get weeded out. But
the definition of "good" and "bad" is hardwired into us from our DNA. And
our DNA, is hard wiring a definition of good and bad, which is as good for
survival as our DNA as it can make it.
> =A0All changes to our behavior happen for very
> > specific reasons - because the change is shown by the internal behavior
> > worth estimation system to be better than the previous behavior.
> >
> > Looking at the odd things humans do, you might not understand how on
> > eart=
> h
> > some behaviors could exist. =A0But they all tie back to the experience
> > th=
> at
> > person has had in their life relative to their internal drives defined
> > by their genes.
>
> In many cases experiences of a single person are not statistically
> significant to verify that certain behavior is useful or not for
> humanity. However, experimentation of humanity as a whole can span
> over hundreds of years of maintaining certain behavior mandated by
> cultural programming and find with absolute significance that say
> washing hand or not marrying cousins is harmful or useful.
Ah, same points I was touching on above!
However, the fact that time tested memes have more value is something our
brain is able to statically identify though our own experience. It's why a
1000 year old religion has so much more power in spreading its memes to the
new generation than a 6 month old religion.
This is the same trick that Google uses to identify web pages. They use
statistics to create a page value ranking, which uses value by association
as a key indicator of estimated true value. They use links to create a
closeness measure, and then use a statistical process to make the web of
connections converge on a ranking that is based on the idea that pages of
value get more links from other pages of value. Good stuff clusters with
good, and bad clusters with bad.
The brain with it's low level statistical system doesn't know to know why
hand washing is is good. It only has to figure out it's coming from the
same place other good stuff came from. The more associations (links) the
meme of "hand washing" has with other "good stuff", the more value the meme
has.
Our personal first hand experience is very limited, but it's more than
enough to form a very good web of values by association. If lots of good
"stuff" happens around our parents (they feed us all the time) we form a
web of trust based on that experience when we are young. The guy that
keeps taking our food from us creates an associated web of low values by
association, etc.
All our learned herbivores are evaluated based on this complex learned web
of stored experiences and assocations. Trust the force Luke!
> Individual human does not discover by himself a damn thing, and even
> when he does in most cases it is illusion specially created by
> cultural programming hidden methodics to keep things interesting and
> to encourage the particular individual
> monkey who was lucky to be the final step in a long chain of discovery
> process to disseminate the resulting information. This even applies
> for purely empirical and serendipitous discoveries, because the
> definition of the problem and significance of the observation
> themselves is created by cultural programming.
You are just way off base.
[repast everything I wrote above a third time].
Believe me, if I hunt you down, and slap you in the fact for being so
stupid, you will learn something from it. And it's because you DNA hard
wired you not like being slapped in the face. But your statistical
avocation system will make you cringe or duck every time you see someone
looking like the idiot Curt Welch walking towards you with his hand raised
because by association, you hard wired learning system senses real danger
it needs to escape from.
> Real discovery processes span over many generations and
> include huge loads of information created by hundreds or thousands
> which admittedly can fit into a single monkey brain, but is mostly
> loaded there ready to consume. IF you want to challenge this
> assertion, give me an example of a discovery you made, and I will show
> you how cultural programming has totally caused and defined it.
Well, here's a few things I discovered today.
I discovered that when I find the lights left on in the Forge, but yet the
door is locked, and the parking lot is empty, Phil was probably there
working but left to for a short time to do something.
I discovered the when trying to do forge work with antibiotic cream in my
right eye hammering the vision blurry, it's hard to hammer objects straight
- they keep coming out bent!
I discovered that after taking a multiple week break forge work, I couldn't
even remember the correct order for making a simple drive hook with a
twist, and that if you put the twist in before adding the bend for the
drive spike, I didn't locate the twist in the right place.
I discovered it is possible, and not all that difficult, but not very easy,
to forge a 1 inch ball on the end of a 3/4 square tapered shaft. I
discovered 5 or 6 different hammer and anvil techniques for doing that as
well today, as well as discovering how easy it was to get a fold in the
steel leaving a scar in the finished ball.
I discovered a technique for making a basket spiral feature out of 1/4"
square stock by double folding and forge welding the ends before twisting.
I also discovered that when the twist is made like that the sharp corners
of the bars end up pointing outward instead of the flat sides - which is
the effect I was actually trying to duplicate.
I discovered than when driving in the dark in the rain, with ointment in my
right eye, I really need to focus more on the right side of the car with my
left, because I ran into a low hanging limb knocked down by the weather
that I didn't even notice until about the same instant I hit it.
I discovered when my neighbor Don calls to warm me about the cars parked in
the street, he may not understand we are the ones having the party.
I discovered a new way to talk about reinforcement learning by using a
metaphor of memes being restricted by a fence erected by our DNA.
I discovered a new way to talk about reinforcement learning by using the
metaphor of memes being evaluated by a gatekeeper built by our DNA.
I discovered a I can link the way the brain works to the way Google works.
That's 10 things I discovered on my own today. I can keep going if you
like.
You might not be very good at exploring and inventing and discovering, but
I've spent my whole life making many important discoveries every day of my
life. Some of them spread to others as cultural memes, most are only
important to me, many are just re-discoveries others made before me.
All the behavior we each learn comes from a constant process of exploring
and discovering which is happening in each of us whether you understand
it's happening or not.
> > Learned behavior are just as instinctive as fixed behaviors. =A0They
> > are =
> just
> > more complex. =A0That is, the instinctive behavior of learning is
> > operant conditioning.
> >
> > With a simple instinct, we might sense heat, and move away from it.
> > =A0Th=
> e
> > hardware is pre-build to always perform that action to that type of
> > stimulus. =A0With operant conditioning, we see some fire, move our hand
> > towards it, get burned, and then the internal behavior system adjusts
> > the map so we are less likely to produce that same behavior in the
> > future. Though the "move away from heat" is pre-wired in one case, the
> > "adjust ou=
> r
> > behavior to prevent sensing too much heat" is innate in the other.
> > =A0It'=
> s an
> > innate behavior created by the genes in both cases. =A0The first is
> > just stimulus response, the second is stimulus, response, evaluation,
> > stimulus=
> ,
> > different response. =A0The sequence is innate and unchangeable in both
> > an=
> d
> > fully specified by the genes - who always have the upper hand (for
> > now).
> >
> > As long as humans continue to be grown as they are now, under the
> > control of our genes, the genes will continue to have the upper hand
> > here. =A0But=
> the
> > more we use our intelligence, to change our bodies, and our genes, the
> > mo=
> re
> > complex the entire process of evolution becomes and the less we will be
> > able to say the genes are in control.
> >
> > However, the upper hand of evolution will never lose control. =A0That
> > is, whatever form we evolve to, it will only exist, if it proves to be
> > good a=
> t
> > survival. =A0At the moment, our genes are still the prime controller of
> > o=
> ur
> > evolution, and the memes really control almost nothing - they were
> > create=
> d
> > by and allowed to exist, by our genes.
>
> Evolution of genes is already secondary for humanity, it is evolution
> of ideas that is primary. Think about a thousand of wild aborigines
> with perfectly fine genes, and handful of cultured man with AK-47
> having the power of ideas evolution. Who will have survival advantage?
Well, if the environment is the outback, the AK47s probably won't save the
guys. :)
Sure, our learning brain gave us a huge step up in our survival skills
because valuable survival wiring of our brain could move from person to
person without having to go though the genes. DNA based life is already a
powerful reinforcement learning machine that is the source of the
intelligent design of all DNA based life. It found a way to harness that
same type of power when it discovered how to build reinforcement learning
technology into our control system (the brain) so that it could evolve
quickly within the lifetime of a single organism.
Calling one of these learning systems primary and the other secondary
doesn't make much sense in terms of evolution. They are two different
forces of evolution at work - one which works at lightning speeds compared
to the other.
The first system is evolving the design of the _entire_ body, while the
second only evolves the wiring of the brain. The brain's learning system
is 99% slave to the DNA learning system because of the heavy one-sided
control of DNA on the design of the brain, and most especially, the design
of the reward centers that defines the primary values by which all
behaviors are judged by the brain's learning system.
Also, the entire human race is really the DNA learning system. A single
race works by maintaining their "code" not in a single human, but in the
gene pool of all humans which are part of the same effective breeding pool.
The ratios of a given gene in that pool really defines the current
"strength" of that gene in the learning machine. So our genetic design is
as much a result of the netire population working for millions of years, as
is our meme design. That, is the current design of a human brain after
exposure to our culture.
So the entire population of the human race is part of both a working DNA
learning process, and a meme learning process. The advantage of memes is
that they are created, and killed, and shared, far faster than the DNA
discoveries which take a life time to test.
We are certainly entering a time where the behavior learning system is
going to have increasing influence on the genetic learning system. Genetic
engineering and all other related technologies are going to give our
intelligence increasing control over our genetics. But there's a real
danger with such a change in the control loop. Our brain's learning system
as a technology that only works well to maximize our odds of survival if
the innate evaluation system is finely tuned to match what's really
important to our survival. The DNA learning system we call the human
species (with it's sex system of reproduction) is constantly adjusting that
primary evaluation system to keep the learning system working. If you flip
that around, and allow the brain to modify the DNA learning system, then
the system could become unstable and stop functioning - and die. The more
the brain can modify the DNA, the closer we get to wire heading problems.
However, the real top level check the system can't escape from is the force
of natural selection, so if the systems behavior controller starts to make
DNA changes that are bad for survival, that sub-set of the species will die
off and take both that DNA and that learning system out of the equation.
The advantage of having the learning brain is that we can evolve behavior a
lot faster to adapt to a changing environment. Quicker adaptation (even
if it's just for our behavior instead of for our entire body), is a clear
advantage. But it also has the odd side effect of making the environment
change very quickly as well, because we use our power to quickly change as
well as quickly adapt. Global warming is one example where our power to
quickly change our environment is threatening to out-strip our powers of
quick adaptation. If the climate changes too much, too quickly, we will
just wipe ourselves out. It might be that a rise in intelligence creates
this spiral of increasing changes to the environment which eventually
become too large and too fast for the intelligence to keep up.
> This is even more obvious if you look at thermodynamical
> implications. One human is only accelerates about 100W worth of
> entropy increase (result of genetic evolution) but the car he drives
> accelerates 250 kW of entropy increase this car is a phenotype of
> humanity that is caused by cultural programming and has no
> representation in genetic programming at all.
Yeah, that's all true and valid. We have genetic information stored in two
places - the first in in our DNA, the second is in our brain wiring. OUR
DNA is only shared from parents to child. The brain wiring is shared from
any living human to any other living humans. But it's not in our computers
yet. Human knowledge doesn't exist in wikipedia, it only exists in our
brain wiring.
But the car is represented in the _wiring_ of the brain built by the DNA.
If humans die today, cars die with us - no new cars will be produced after
the last human dies. The car is linked back to the DNA the same way our
arms are. If the DNA dies, it all dies. The car is just a two step link
instead of single step.
> It is not to say that genetic evolution has finished, it still
> continues but in the environment created by cultural evolution
> and to satisfy thermodynamical optimum behavior of Life of the
> humanity as a super-organism defined by both cultural and genetic
> programming rather that of an individual monkey.
The creation of sexual reproduction killed the evolution of the individual
monkey and turned it into an evolution of a gene pool of a species instead
of the evolution of individual genes down a replication tree. It was a
huge technological improvement in sruival systems.
Adding to that system, the strong learning brain that was powerful enough
to learn to trade programming quikcly, kicked it up yet another notch.
But what's evolving is not "culture", it's "human brains". What you keep
talking about as human couture evolving separately from humans is just not
what's happening. What's evolving is a large pool of human brains. And as
a pool of brains which communicate, they have actually formed one large
intelligence which is itself learning by evolving as one large control
system for the whole species. It's the global brain - or at lest the human
global brain.
Trying to talk as if culture is something separate from the human brain by
using words like "information portability" and making the error of
believing the control systems which are evolving currently exist in our
computer databases is a common error people make. It's another topic
people are lost and confused about that I write endlessly about here. It's
the mind body problem - which is an illusion that has most the world is
totally befuddled about. It's the illusion that creates the mind body
problem, and gave rise to beliefs in a non physical human soul. And though
much of that has been removed from modern belief systems, it still hands in
there as endless confusion over consciousness and the mind body problem,
and embedded in our culture (our brains) in the form of language where we
are trained to talk about mental events being non physical as if that is
something that can even happen. That is, that ideas, and feelings being
somehow separate form the brain. When we talk about ideas, we are talking
about brain behavior, not something separate from the brain, and not
something created by the brain. We are talking about the brain itself.
And when we talk about human knowledge, or our "memes" it's just an
abstract way of talking about the human brain.
The evolution of human memes is just the physical evolution of the human
brain. And so far, it's not "left" the brain in ANY sense. The evolution
of the human brain hasn't moved to computers yet. Our DNA, our cell
structure (which interprets the DNA), and our brain, and it's wiring, and
all the things human build, like roads, and buildings and cars, are one
huge adaptive system full of millions of symbiotic relationships between
parts. But currently, at the middle of it all, in the top level driver
seat (or the lowest level foundation of what drives it all), is the DNA
based adaptive learning machine we call the human species. It's what
driving it all and keeping it all working.
> DNA is still a storage medium used by this super-organism, along
> with CDs and paper, and it still has extremely efficient and extremely
> robust survival programming that will be used and reused by humanity
> for thousands years to come. But it has lost its uniqueness as the
> _only_ way how evolving programming can be stored.
Yes, once strong learning was added to the brain, we gained another
sub-system that had it's own system of evolution by learning. But it's
still _us_ It's our brain and what it does. It's not ideas separate from
the body that can float off like a soul at death (or before death).
So the question Tim and I debate in circles, is how is the invention of
strong AI going to change the path of evolution for this large super
organism?
Tim talks as if the AI will just replace the human body and take over the
work of building roads and buildings and generally moving energy at ever
faster rates. Well, the AIs don't need our roads, or our buildings, or our
farms, or our tools, or the factories for making human-tools like cars.
Much of what was built as extra shell around the human body would have no
use by the AIs - they will just tear it all down, kill us all off, and get
on with their job of increasing entropy.
Human civilization is just the external parts of this large super organism
to support all the systems inside the body. If the AIs take over, they
won't just replace the humans, they will kill and re-build the entire
super-organism.
The super organism known as humanity is not suddenly going to fall apart
and reform in completely different structure because the super organism
invents one new learning algorithm called human-level AI.
We are going to tie it into our value system, so it keeps locked into our
values - which ultimately is the survival of this super organism and it's
core of selfish DNA. Ultimately, all AI is going to do, is make the super
organism more intelligent. We already function as one large intelligence,
and each new communication system we invent makes us more intelligent by
helping the brains share knowledge faster and better. With AI, the
computers stop acting as fancy communication systems and start to add more
core intelligence to the super organism without having to grow more brains.
Our memes will then spread out into the AIs - but becuase they are
configured to be slaved to our DNA created values, the whole system will
stay in rough sync, and memes will move freely from humans to AIs and back.
In the long run, this super organism might evolve it's design to replace
the core DNA/biology core that currently drives it all. But there's no
reason to believe the invention of AI is going to make it flip because AI
by itself is not a good super organism. Something must define it's values.
If you allow the values to be directly defined by the super organism, it
will wirehead itself and stop carrying about survival. If the whole thing
makes that mistake at the same time, the super organism dies.
The value system that drives the learning brain, is defined by the DNA and
tuned by millions of years of evolution. But the learning brain has now
built itself a huge set of derived values in it's meme storage system that
drives the behavior of the super organism - that drive all our social
systems.
AI by itself is not a super organism. We might in time, be able to design
a supper organism to replace us, but a computer controlled robot running a
fancy learning algorithm is no super organism. Our entire culture and
society, and government structure, and belief systems, are all here to
support this core foundation of DNA based humans. It would all have to be
replaced and redesigned from the ground up if you replaced the core biology
with robots.
Even with the help of billions of AIs, redesigning somethign that took
evolution millions of years to create, and a 100 thousand years of
brain/meme evolution of finding optimal social structures to support
humans, is not going to be replaced by an entire new structure with an AI
core any time soon.
Our biological AI (our brain) is slave to the DNA "brain" (intelligent
learning system) already. When we create AI, it will just be hooked onto
the bottom of that stack where we will use our brain, to define the drives
of our billions of different AIs, and they will work for us, like we (aka
the brains) are already working for the DNA. In time, all this
intelligence is going to start reaching under the hood and redesigning the
core biology. How this super organism will self evolve by messing with
it's primary drive, system (DNA) is hard to guess. But changing it is like
trying to change the tires on on a moving car. Not an easy task. It's like
trying to change the foundation of a skyscraper after you get 100 stores
above it build. In time, piece by piece, the biological foundation of this
super organism may be replaced, but it will be a many thousand, if not many
million, year process.
What will happen in the short term, is we will add these AIs just like we
add all our other machines to the outer shell, to help us do more of what
we are already doing - keep searching for, and finding better ways, to move
more energy faster.
--
Curt Welch http://CurtWelch.Com/
Tim Tyler
> Yevgen Barsukov <evge...@gmail.com> wrote:
> It's because the power of a meme to spread, has nothing to do with the
> meme, and everything to do with the value assigned to the meme, by the
> innate meme evaluation hardware built into us by our DNA.
Except in cases like this:
''Computer scientists at the University of California, San Diego and its San
Diego Supercomputer Center (SDSC), Eureka-based Silicon Defense, the
University of California, Berkeley, and the nonprofit International
Computer
Science Institute in Berkeley, found that the Sapphire worm doubled its
numbers every 8.5 seconds during the explosive first minute of its
attack.''
-
http://www.scienceblog.com/cms/sapphire/slammer_worm_shatters_previous_internet_speed_records
> Why does a religious meme of hand washing spread? Because it exists in
> people who have established themselves as worthy of being mimicked.
>
> And why does some set of cultural beliefs survive? - because either 1) it's
> harmless crap that got lucky, or 2) like real genes, the meme is actually
> useful to the survival of the human and their DNA.
>
> Memes that are harmful to survival get evolved out of existence. The more
> harmful, the faster they are removed. Harmless behaviors can live for a
> very long time simply by random chance.
>
> Memes that survive the test of time, prove their value. The older the
> meme, the more likely that the meme is actually valuable to survival.
To *its* survival - not to the survival of DNA. We have been over this
several times now. I am repeating myself, but:
''Time and again, my sociobiological colleagues have upbraided me as a
turncoat, because I will not agree with them that the ultimate criterion
for the success of a meme must be its contribution to Darwinian
�fitness�.
At bottom, they insist, a �good meme� spreads because brains are
receptive to it, and the receptiveness of brains is ultimately shaped
by (genetic) natural selection. [...]
It is, of course, true that �Memes are utterly dependent upon genes,
but genes can exist and change quite independently of memes�.
But this does not mean that the ultimate criterion for success in
meme selection is gene survival. It does not mean that success
goes to those memes that favour the genes of the individuals
bearing them. To be sure, this will sometimes be so. Obviously
a meme that causes individuals bearing it to kill themselves has
a grave disadvantage, but not necessarily a fatal one. Just as a
gene for suicide sometimes spreads itself by a roundabout
route (e.g. in social insect workers, or parental sacrifice), so
a suicidal meme can spread, as when a dramatic and well-
publicized martyrdom inspires others to die for a deeply
loved cause, and this in turn inspires others to die, and so on.''
- http://alife.co.uk/essays/misunderstood_memetics/
> The bottom line however is that good memes spread for good reasons,
> harmless memes spread but our harmless, and bad memes get weeded out. But
> the definition of "good" and "bad" is hardwired into us from our DNA. And
> our DNA, is hard wiring a definition of good and bad, which is as good for
> survival as our DNA as it can make it.
...which is totally rubbish. Genes have had no time to develop adaptations
to deal with memes. Given enough millions of years a memetic immune
system might develop to help weed out bad memes - but that is going to
be a very slow process - and evidently it has only just begun.
In the mean time, the bad memes get a free ride with the good ones -
which helps explain why we see so many useless religions, chain lettters,
superstition and nonsense.
> The evolution of human memes is just the physical evolution of the human
> brain.
Not unless you define "memes" that way - and that, I maintain, is a
very bad practice. In fact, external transmission media host most memes.
> The evolution of the human brain hasn't moved to computers yet.
What will move to computers is not the human brain, but its
functional capabilities. Of course many of those have already
migrated across. Machines have far better memories than us.
They are better at mental arithmetic - and can communicate
with each other much better than we can.
--
__________
|im |yler http://timtyler.org/ t...@tt1lock.org Remove lock to reply.
rs...@nycap.rr.com
> Curt Welch wrote:
> > Yevgen Barsukov <evgen...@gmail.com> wrote:
> > It's because the power of a meme to spread, has nothing to do with the
> > meme, and everything to do with the value assigned to the meme, by the
> > innate meme evaluation hardware built into us by our DNA.
>
> Except in cases like this:
>
> ''Computer scientists at the University of California, San Diego and its San
> Diego Supercomputer Center (SDSC), Eureka-based Silicon Defense, the
> University of California, Berkeley, and the nonprofit International
> Computer
> Science Institute in Berkeley, found that the Sapphire worm doubled its
> numbers every 8.5 seconds during the explosive first minute of its
> attack.''
One has to ask, “Does the concept meme have any staying power?”
The wordsmith Dawkins introduced it in 1976, and it is still with us.
The wordsmith typically envisions a cluster of thoughts and things
that he thinks has the striking aspect of newness. He dredges up a
word for it, and fills up a column, or a chapter. In this case, he
likened his concept to evolution. Now evolution is based on the
chemistry of the genome, but the meme has no such organic anchor. It
is adrift in a sea of words.
My feeling is that it will tinkle around in a few newsgroups for a
half-century and then drift away.
There is nothing there.
Ray
Tim Tyler
> One has to ask, �Does the concept meme have any staying power?�
>
> The wordsmith Dawkins introduced it in 1976, and it is still with us.
It is growing just fine at the moment - see:
http://google.com/trends?hl=en&q=meme
> Now evolution is based on the chemistry of the genome, but the
> meme has no such organic anchor. It is adrift in a sea of words.
Memetics is a kind of evolution too. Memes - like genes - consist of
information,
which can be instantiated in practically any physical medium. Genes
do not stop being genes simply because their information is temporarily in
the database of a genetic engineer. It is the same with memes - they are
not tied to any particular information storage medium.
> My feeling is that it will tinkle around in a few newsgroups for a
> half-century and then drift away.
Its competitors are doing much worse. Check out Boyd and
Richerson's "cultural variants" for example. Basically, it has
been a complete walkover for the memetic terminology.
zzbu...@netscape.net
> so? I have a basic engineering and arts foundation? Tell you what! I
> was chattig with some guy at nearly the last of the last fashionable
> 'set' in the hive and we said 'Dell!' but basically I cant imagine any
> hot industy engineering technology dept not taking on and paying, ne,
> nurturing the best minds to presuppose a product...? ....?...?...?...?
> anyways.....sci,lan
Well, as far as computers are concenred the only thing any
engineering departments for is supercomputers. So the
post gravryard people long ago switched to atomic clock
wristwatches,
holographics, laser disk i/o, Digital Books, Distributed
Processing Software,
and Post GM nomics.
N
wrote:
yeh...things havnt needed to alter much for a few thousand years, so
it seems, all excepting some pretty horrendous changes in t he use of
medicine to control people...consequentially I think the 'social
statistics' used in any research on development and inventive or
creative, ie. those to represent.'intelligence' can't be anything else
but skewed out of all proportion...
Civilization has been going on for many thousands of years before
type, symbols and database control or research too! Anyhows...you
guys are most used to a certain kind of symbolic reasoning and I
thought about it, structure, but a hell of a lot of the population of
the world still understand other images to represent ideas in other
languages.
N
people purely
picked out due to knowledge of their genes...this is added to the text
book
interpretations already out of date in print...and then there are over
anxious
mentors, people who design data bases and salespersons who desperately
want to see their text book sales ID conclusions be found 'correct' (a
man is
innocent until disproved)also bearing in mind that most Dr's of
medicine
frequently seek the 'abnormal' in order to generate som kind of work
load, I would
be very serious and critical of any developments in tech used to
idetify behaviours
alone on the supposition that 'these are research papers' say?
I understand the system to suggest that women who are prostituted have
the
very best chance of gene frequency, and prior to that, male rapists.
I don't believe that.
Curt Welch
> On Dec 30, 6:07=A0pm, c...@kcwc.com (Curt Welch) wrote:
> > casey <jgkjca...@yahoo.com.au> wrote:
> > [delete about brain]
>
> > I don't believe it's all that relevant to AI.
>
> Its relevance is in regards to a practical implementation.
>
> >> You might ask how one piece of cortex receiving data from the
> >> eyes decides to process color, another piece decides to
> >> process texture, another decides to process motion and so on.
> >
> >
> > Well, there are two questions there. One is how would the parts
> > of a generic learning system distribute information though the
> > network, and the other, is why did the visual cortex distribute
> > it the way it did.
> >
> >
> > My networks already do just the sort of thing you are talking
> > about. They automatically slice up the data into different
> > features, and process the different types of features in
> > different parts of the networks.
>
> You make claims about what your networks ARE doing but without
> allowing anyone to test those claims they are meaningless.
Not only have I allowed people to test it, YOU have written the code for
yourself and run it and have seen it do these things. The fact that you
don't grasp what it was doing, is simply beyond my control. But it's
certainly not because I'm making claims that can't be supported. All I've
done there for years is top try and share my ideas with people.
> Are you claiming that with a visual input one part automatically
> takes on color, another motion, another texture and so on?
Nope.
I'm claiming:
"They automatically slice up the data into different
features, and process the different types of features in
different parts of the networks."
I'm calming that different parts of the network automatically deal with
different parts of the information. I didn't claim that the way my network
sliced up the information and distributed it matched the way the cortex
does it. It most definitely doesn't match the the way the cortex does it.
> You gave me the impression it was all mixed together in your
> nets, the latest of which I haven't seen.
Nope, not all mixed together. The way my networks work is to slice up the
information and assign a different node to process each different slice of
the information. It automatically distributes the information evenly
across the entire network with no single node in the same network
processing the same information. Every node is configured to receive and
analyze a different slice of the information. That part works just fine.
What it doesn't do, is slice it up correctly. :) Trying to figure out just
what is correct has been what has stopped me for the past few years. But I
know enough to know that the way it does it, is not correct. Or, at least,
not optimal.
> > Sure, but if a tree recognize can be used to recognize coke
> > bottles, then it ISN'T A TREE RECOGNIZER! It's a GENERIC
> > pattern learning machine.
> >
> >
> > You can't have it both way's John. It's not a HARD CODE TREE
> > RECOGNIZER. IT's a high quality GENERIC sensory pattern
> > recognition hardware.
> >
> >
> > And yes, it was developed to deal with the wild, but it was
> > still developed to be a GENERIC pattern recognizer so that
> > evolution didn't have to waste any time hard-coding tree
> > recognizers, and rock recognizers, etc.
>
> Depends how you want to define "generic" as it doesn't have
> the ability to recognize ANY pattern.
That's right. It can't recognize all patterns. It's limited. That's just
a basic reality. The complexity of the patten any pattern recognizer is
limited by the complexity of the machine doing the recognition. Since we
can't build infinite sized machine, we can't recognize an infinite number
of patterns.
But if the limits of what it can be configured to recignize are themselves
generic, such a max length, or max resolution, or min or max frequency in
the pattern, or max complexity (max information in the pattern) and it's
able to recognize any classification of patterns within those genetic
limits, then I think of it as a generic pattern recognizer.
For example, in the binary number domain, if we build a recognizer that can
define a pattern class that includes any subset of 32 bit binary patterns,
then I call it generic, even though it's limited to 32 bit number patterns.
But if you build hardware that first does something like sums the number of
bits in a bit string, and then does "generic" pattern matching on the sum,
then that pattern matcher is not generic in the domain of binary bit
streams because there are many bit streams classifications we can define
that the machine won't be able to classify.
In general, the larger the set of patterns any system can be configured to
recognize, the more generic I would claim it to be.
And likewise, if the system included any highly specific domain defined
simplifications, I would claim it was not generic.
> Bottle tops and coke
> bottles have the same features as natural objects.
Yes, man bad graphics do very much mimic features from our environment.
Our visual system is very much optimized and configured to deal with the
features which are typical in our environment.
But how did it get configured like that? By nature or nurture? You argue
nature, but yet you can't produce any data to support the position that
isn't just as strong as the data supporting the opposite position.
> http://www.owlnet.ricea.edu/~psyc351/Images/MinskyPapertSpirals.jpg
>
> Unlike bottle tops and trees you can NEVER learn to recognize
> the two different types of spiral patterns just be looking at
> them. You don't have the innate parallel machinery to do that.
Yes, but is it missing some low level innate feature detector, or is the
feature just beyond the range of what our generic detectors can learn.
Do you grasp what the feature detectors have to recognize in order to
classify those two types of patterns? It has to create an even odd line
count detector for 15 lines.
Can you detect whether these patterns of lines have an even or odd number
of lines?
|
| |
| | |
| | | | |
I can. So to suggest we don't have the hardware is an invalid argument.
We obviously do.
Can you detect if this is even or odd:
| | | | | | | | | | | | | | |
I can't.
Do you honestly think it's because I don't have innate hardware for even
odd detection, or simply because the pattern is beyond the complexity and
size limits for my genetic detectors to recognize?
> In order to "see" we have parallel machinery which makes
> assumptions about how the real world is put together to the
> extent that we will see things that aren't really there.
What it does, is adapt its configuration based on the constrains that are
actually in the data. It's IMPOSSIBLE to explain how genetics is able to
perfectly match the strcuture of the visual network with the phsyical
structure of the eye without including learning systems that use the real
constrations in the data to auto-configure itself.
But once you have hardware that can auto-configure itself to adapt to the
constrains in the data crated by the eye and the lay out of sensors in the
eye, then that same hardware can auto-configure itself to adjust for those
constraints can adjust for constraints created by all the stuff "out there"
past the eye.
It makes no sense to believe that evolution would create, and use,
auto-adjusting networks to allow it to build an eye, and for some stupid
reason not use the same auto-adjusting technology to deal with trees and
rocks.
The only way to explain how the brain has the generic learning powers it
has, is to assume it's got hardware that can identify, and adapt to, the
constrains in the sensory data created by the objects that exist in the
universe. It has to have that technology. And once it has that
technology, why not use it for all sensory systems. And lo and behold we
see it is using the same module, to process the data from all sensory
systems - the cortex.
> We will see virtual edges, 3D in a 2D image, false movements,
> things that are the same size or color as being a different
> size or color and no amount of learning will stop us seeing
> them that way because our INNATE visual system is FIXED to
> see them that way.
Except we know that's bull shit because the system doesn't become "fixed"
if it's not receiving real data. I don't know why you choose to ignore all
the facts that that are a direct contradiction of your position. If what
you are saying is true, then the brain wouldn't need real data to configure
itself. But it does need real data, and if the eye is defective, the
network configures itself in a defective way. If what you are suggesting
were true, the visual system would build itself correctly with or wither
the data flow.
> The kinds of illusions we experience flag possible assumptions
> used by our innate visual system to decode the retinal images
> into a representation of what is out there.
The illusions we experience show how the network become configured as a
RESULT OF it's adaptive learning system.
> >> Whereas calculus is hard because we do not have an innate
> >> calculus module, learning to recognize 3D objects is easy
> >> because we do have an innate object recognition system.
> >
> >
> > Calculus is hard because it's highly abstract. Too abstract
> > for some brains.
>
> And you think seeing isn't highly abstract? Isn't recognition
> a process of extracting common qualities from specific examples?
Sure is.
> Recognition of an object means abstracting out what is common in
> a set of examples of that object. And if you think that is easy
> why don't you have a go at trying to duplicate it.
I am. And when the answer is found, you will finally gasp how trivial it
actually is. The only think hard here, is finding the simple answer.
> The question you need to ask is why calculus is easy to program
> and "seeing" is not easy to program. If seeing is easier than
> calculus why can't we solve the seeing problem?
We can and will solve it and it will be easy when the answer is found.
ONly the finding is hard.
> The ability to make those visual abstractions is innate whereas
> the other kinds of abstractions are not.
Three is no argument you can make John that I can't counter with an equally
strong argument for the ther position.
When you are shown calculus for the first time, when you are 16, calculus
is hard, and seeing 3D objects is easy. Why do you think that might be?
Might it because by that point in your life, you have had 16 YEARS OF 24x7
EXPOSURE TO 3D IMAGE DATA, or 101760 hours of practice, And 0 HOURS OF
EXPOSURE TO CALCULUS?
Don't you suspect that after 101760 hours of exposure to calculus, it too
might become easy? It does.
> We cannot for example
> "see" 4D objects because we lack the hardware to do this. This
> is why it is hard visualize a 4D object.
It's because we never get 101760 hours of practice "seeing" 4D objects.
> Calculus is NOT HARD it is just not interesting to most people.
And far more important, most people never get exposed to it for even a
single hour in their life.
> To quote Richard Dawkins, "Whenever I feel intimidated, [by math]
> I always remember Silvanus Thompson's dictum in Calculus Made
> Easy: 'What one fool can do another can'.
>
> If you really want to understand calculus you only have to put
> the time into it.
So, your argument is, since we don't have the innate hardware to deal with
it, putting time into it adds it to our genetics and gives us the innate
hardware to deal with it?
> If you really want to understand vision you
> will have to put a lot more time into it - so far human vision
> is not understood - it is THAT hard.
And some huge percentage of what makes it hard is because the idiots think
like you do - they think they are looking at modules designed by evolution
and are trying to decode the design instead of understanding the entire
cortex is one large adaptive learning module and to understand it, you must
stop trying to understand the product of the learning algorithm, and
instead, understand the learning system that build the hardware.
The mistake they are making would be like looking at the neural network and
the weights that TD-Gammon created by learning and trying to understand how
to hand-code a network to duplicate the same game playing skills of
TD-Gammon. It's beyond what any human can understand. But how it learns
on it's own, is easy to understand. You just have to know that what you
should be looking for when you look at the brain, is a learning system, not
a vision system. The more time you waste trying to understand how it works
as a vision system, the more lost you will become. All we can ever do
there, is just barely scratch the service of what it's actually doing.
> If we didn't have the parallel hardware with its build in assumptions
> to learn to recognize natural objects and had to use the methods used
> to understand calculus, or discriminate the two types of spirals, we
> would find vision impossible to do in real time!
Yes, the processing is not trivial, but that has nothing to do with whether
it was innately wired by DNA or wired by a learning process.
> > The high dimension learning problem is not solved by simplifying
> > it to a low dimension problem.
> >
> > It's solved directly in it's high dimension form.
>
> I wonder what you really mean by that last statement?
I mean that there are no set of signals you can identify inside the system
that are a low-dimension internal version of the problem.
> > It's solved by a system that solves it as a high dimension problem
> > - which I've already roughly outlined many times how it can be done.
>
> It's physically impossible for the brain to match the high dimensional
> problem and not required as shown by chess playing programs that
> solves
> winning a game of chess by using simplifying assumptions called
> heuristics
> without the need to solve the high dimensional problem.
>
> What exists physically in the brain is not a high dimensional version
> of the real world but rather a simplified version of that world.
You don't seem to have any concept of what "high dimension" means.
Yes, the brain is a highly simplified model of the real world. No doubt
about that. But when the simplified internal representation is a billion
parallel real time pulse signals, the problem is still not "low dimension".
ONE real time pulse signal from a real world sensor is still HIGH
DIMENSIONAL.
You don't seem to have shown any desire to understand the true complexity
of the _learning_ problem the brain solves.
Curt Welch
> Just to clarify a point you persistently misread.
>
> Curt wrote:
> > You can't have it both way's John. It's not a HARD CODE
> > TREE RECOGNIZER. IT's a high quality GENERIC sensory
> > pattern recognition hardware.
> >
> >
> > And yes, it was developed to deal with the wild, but it was
> > still developed to be a GENERIC pattern recognizer so that
> > evolution didn't have to waste any time hard-coding tree
> > recognizers, and rock recognizers, etc.
>
> You have responded this way a number of times as if you
> imagine I think every object has its own recognizer?
No, that's really my take on what's happening showing though in that type
of response.
> You claim humans have replaced the old learning system with
> a new generic learning system and that is why we can learn
> all our high level behaviors. I am suggesting that our
> learning system is nothing more than an enhancement of the
> old one and it has its limits and is innate.
Of course, learning systems are innate, and learning systems always have
limits.
> Only being able to recognize natural objects is not one of
> its limits! It turns out the SAME mechanisms allow it to
> recognize coke bottles and bottle tops without the need
> to add a new generic learning system on top.
The fact that it can deal with items in the environment that evolution has
never had to deal with int he past is the very definition of generic. It's
the proof that it is generic.
The only question to understand here is what are the limits of that generic
system.
If we fully understood both its limits, and it's implementations, the fine
points of what we constantly debate could be better discussed. But we
don't, so we debate speculation.
If you could show me that the limits of the system were in fact highly
turned to match what exists in our environment, I would far better buy what
you are suggesting.
For example, when we look at the sun's light spectrum and compare it to the
sensitivity of the eye, we see the two are fairly well matched. Obviously,
the eye was evolved to match the EM characteristics of life on the surface
of the earth. The eye was innately tuned to match the environment. It's
not an adaptive learning system that adjusts it's EM sensitivity after
birth to match what it finds in the environment. Drop a human in an
environment where the EM light is in a different range, and the human will
just be blind to it becuase the eye is not an adaptive system in that way.
But the signal processing that happens in the brain is highly adaptive,
even though it's limited. But what are the limits? If we can define what
those limits are, and then measure the actual limits of the information we
find in our environment, and snow the two are highly matched, like the eye
sensitivity to the EM light in the environment, it would be far easier to
understand and debate the innate nature of human adpative powers.
However, my real position about all this goes back to the other side of the
argument. Whatever the limits are, the brain is more generic (less
limited) and more powerful, than any machine we have yet built - or even
have a clue to know how to build. So to solve AI, we have to figure out
how to build MORE GENERIC learning systems than anything we have yet built.
The learning system in TD-Gammon doesn't work for chess. But the SINGLE
GENERIC learning system in the brain works for Backgammon, and chess, and
go, and computer programming, and driving to the store to buy milk, and
doing AI research. One generic powerful adaptive learning system can adapt
itself to perform all these advanced tasks and millions others.
The learning algorithm in TD-Gammon is highly generic already. Very little
about about the game of Backgammon itself was hard-coded into the learning
system in TD-Gammon. Most of what's hard coded is the general dynamics
that exist in all board games. But yet, with all its generic learning
powers, and it's ability to equal human performance in the domain of
Backgammon, it can't touch human performance in any other domain like
chess. But yet the brain can. Which means the brain (no matter ho
limited) is still many orders of magnatude MORE GENERIC than the already
highly generic learning algorithm in TD-Gammon.
It's these facts that make me talk about what the brain does as generic
learning, and it's these facts, that make it clear to me that the missing
technology, the missing understanding, that we have to master to solve AI,
is stronger GENERIC learning.
It's not important to understand what evolution might have done in the
brain to optimize it's powers to solve specific survival problems of cave
men. Except, in what it clearly did - which was to develop a highly
GENERIC adaptive learning power.
> What about our complex motor acts? The idea of using a car
> doesn't occur to an animal. It is not a lack of motor learning
> skills or visual processing skills on their part. Have you
> seen dogs using skate boards or Apes steering a tractor?
>
> Instead of looking for the sudden appearance of a magic new
> generic learning system you might consider it in terms of
> an evolutionary enhancement of the brain to make use of the
> way our bodies have evolved.
Of course it's an evolutionary enhancement to make use of the way the body
evolved. What does it gain to talk or think like that. We know exactly
what the enhancement was - it's a generic learning system to control a real
time dynamic physical system.
We can FULLY UNDERSTAND that problem, simply by trying to add an adaptive
learning controller to a robot. How do we add adaptive learning powers to
a robot?
I fully understand the problem and no amount of looking at evolution of
humans will give me a better understanding of the problem.
The solution to the problem of giving robots adaptive learning powers will
not be found, or improved, by looking at what cave men had to deal with in
their environment. We have all the problems we need right here in front of
us today.
How do I make an adaptive learning controller (that is more generic, and
more powerful than the learning in TD-Gammon) to put into a robot to make
it learn on its own, how to collect empty coke cans in an office? To make
it learn, on it's own, how to deal with all the complex things such a robot
might run into in an office environment? That's is just one of a million
examples of the things humans can learn with their generic adaptive
learning controllers that our robots can't.
How much innate (non learning) hardware exists in humans is not relevant or
interesting. What's relevant, is the innate GENERIC problem solving
hardware that we don't yet know how to duplicate.
casey
I have responded to this in a new thread innate vs. learned.
JC
Curt Welch
You have some very weird ideas. :)