Re: Symbolic AI: Why Marvin Minsky and Curt Welch Are Out to Lunch

Traveler <traveler@xxxxxxxxxx> wrote:
On 22 Jul 2006 09:35:47 GMT, curt@xxxxxxxx (Curt Welch) wrote:

Traveler <traveler@xxxxxxxxxx> wrote:

Intelligence is discrete signal processing. It is based entirely on
the temporal relationships between unnamed (non-symbolic) signals. The
origin (sensors) and destination (effectors) of the signals are
irrelevant.

Then, what you are saying, is that your program can act "intelligent"
even if it's not connected to a stimulus source or an effector? That
your program "knows" what to do, even when it's not receiving any data,
and it's outputs are doing nothing?

Or, that your program will be just as intelligent, if the inputs are
from random number generators and the outputs are connected to nothing?

Why put words in my mouth? That is disingenuous.

Do you see the question marks at the end of the words? That means I'm
asking, "so is this what you are saying?". I put no words into your mouth.
They were rhetorical questions because I knew you wouldn't agree - but yet
it is what I see as the only conclusion your words could lead to. There is
nothing disingenuous about that.

Of course this is not true. Without being connected to an environment,
the signals have no meaning, and without any meaning, there's now way
for your algorithms to do anything useful.

This is not what this argument is about.

I think it is. That's why I wrote it.

Symbols don't have to be "named" to be useful.

They are not symbols if they are not named.

You should try looking up the definition of symbol in the dictionary.
Symbols aren't defined by their names, they are names defined by their
meaning - they are names that are defined by what they represent.

They have to have meaning.

Of course, they have meaning. That is the purpose of symbols, to
assign meaning.

Right. They are names that represent some meaning.

That means the symbol itself is a name, of the thing it is representing.
A spike coming from a light sensor is a name, given to the "light" which
that spike represents. By generating that spike, the sensor is in
effect, saying to the world, "I see some light". The spike is a name,
given to the light, by the sensor. The spike doesn't have a name, it is
a name. Symbols are names for the things they represent. The spikes
feed to your network always represent something, and the spikes coming
out of your network, always have meaning (such as "stop arm motion").

I am not arguing that signals have no meaning. Of course they do. This
is what the temporal relationships between signals is all about,
meaning.

Yes, we agree about that and I was pretty sure you felt that way. That was
why I am surprised with you keep contradicting yourself by saying spikes
aren't symbols. To say they aren't symbols is to imply they have no
meaning - that they represent nothing.

I am saying that the premise of symbolic AI is that the
meaning of a symbol (what it represents) is known to the symbol-based
intelligence a priori.

How can that be? They hand-code the meaning of symbols into the system so
that it can know the meaning. The machine doesn't know the meaning a
priori. It knows it only after it is taught the meaning by the action of
the guy hand-coding the meaning of "cat" into the knowledge database.

The only different in that approach is the fact that it's a Turing "adult"
machine instead of a Turing "baby" machine. Meaning instead of building a
learning algorithm that fills its own mind up with facts about the
environment, we, the creator, fill it's mind up for it. The intent is to
get the machine to end up knowing the same things either way.

There is no way for a symbol manipulation
system to use a symbol unless it knows its meaning beforehand.

Well, there's no way for a learning system to know the _correct_ way to use
a symbol until it learns what that may be. And it's the learning
algorithms in the system that extract that "correct" usage through some
system of learning. So when you first turn on your learning machine, it
also doesn't know the correct way to use symbols.

It's the same either way, they just use different learning algorithms - one
uses a human in the learning loop, one attempts to take the human out of
the learning loop.

This is
the reason that the deluded Cyc crowd has to hand feed zillions of
symbolic meanings into a database.

Right - because they chose to try and build an adult machine instead of a
baby machine that can learns all that on it's own. It's a perfectly valid
approach - the same approach used when we try to hand-code chess playing
skills into a machine in order to make a chess playing machine. The goal
of a project like Cyc is to build a generic symbol manipulating machine
that could talk to us. And to some extent, they have managed to create
just that. But it's far short of what humans are doing, because we are
learning machines at the core - and it doesn't just affect our ability to
go to school and learn, it's shapes our behavior as we talk. It defines the
purpose behind all our behavior. What I write at the end of this message
was shaped by what I learned, as I wrote the beginning of the message.
That type of constant and continual behavior shaping through constant
learning, is what is missing from Cyc.

This is not required in a
non-symbolic learning system based on temporal relationships. Meanings
must be learned and the only way to learn them is to discover how
signals are correlated temporally. Learning is a thus a probabilistic
process.

Well, I agree with the idea you are after - you and I are basically working
on the same problem using similar techniques to try and solve it. But all
you have done is created a baby machine instated of an adult machine. You
have created a machine that attempts to fill it's own knowledge database up
instead of one which requires us to fill the knowledge database up for it.
You are still building a symbolic AI system - you just added learning, and
you chose to work with far more generic symbols (pulses) instead of the
higher level abstraction of "words" (or symbols near that level of
abstraction.

The symbols have meaning before your machine learns them. So to argue the
symbols don't have meaning is wrong. And your machine doesn't really need
to understand the meaning of the symbols as much as it needs to understand
how to respond to them - what behaviors to create in response to the
symbols. So what your, and my machine is really doing, is assigning
meaning to our output signals more so than defining meaning in the input
symbols.

I fell the need to also point out to you that Turing in his famous 1950
paper laid out two approaches to AI. The first was to create an adult
machine, where we would have to hand-code knowledge into the machine, and
the second, was to create a baby machine, where it would extract knowledge
from the environment on its own.

Symbolic AI projects like Cyc are mostly exploring his first suggestion.
You are simply exploring his second suggestion.

You write these messages talking about how dumb these guys are, and then
you go about doing exactly what Turing suggested be done 50 years ago and
act as if you have some great new insight into AI that people like Turing
never understood. Your attempt to elevate yourself above them is only
doing them, and yourself, a great injustice.

Louis Savain

I do believe at this point, that any AI project other than a learning
machine, has now hope of creating what we think of as intelligent behavior.
It will at best create "smart" machines that perform tasks like playing
chess, or talk to us in English about the contents of a knowledge database.
To actually act human and intelligent, it has to have self purpose which
comes from being a learning machine - the purpose that drives the learning
(directs what it learns), is what creates all our uniquely human,
intelligent, high level purpose. And what we are always trying to learn,
is always, what behavior we should be producing.

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

Relevant Pages

  • Re: Symbolic AI: Why Marvin Minsky and Curt Welch Are Out to Lunch
    ... the temporal relationships between unnamed signals. ... signals have no meaning, and without any meaning, there's now way for your ... spike represents. ... By generating that spike, the sensor is in effect, ...
    (comp.ai.philosophy)
  • Re: universal concept list
    ... > elements whenever you forget a word shape or word meaning. ... > how to respond to those situations isn't absurd. ... The topic under discussion is learning the word in the first ... You have the strangest ideas about language. ...
    (sci.lang)
  • Re: universal concept list
    ... >> after some period of time with what was originally learnt. ... > you don't remember the meaning for in the target ... > target language but not its meaning would make learning ... If I see in a German ...
    (sci.lang)
  • Re: universal concept list
    ... >> elements whenever you forget a word shape or word meaning. ... the process of learning the word. ... > learning a new word in a foreign language. ... learn 100 Finnish words that had similar English shapes. ...
    (sci.lang)
  • Re: FAQ * r.g.l.faq * 01/17/2006
    ... > meaning of a word, you have been blessed with another learning ... nyuk nyuk nyuk ... >> Webster's dictionaries spells it exactly the same way I did. ... >> learning opportunity. ...
    (rec.gambling.lottery)
Loading