Re: What did that thread indicate?

Traveler <traveler@xxxxxxxxxx> wrote:
> On 26 Sep 2005 01:07:04 GMT, curt@xxxxxxxx (Curt Welch) wrote:

> Read also about the cocktail party problem. It is
> impossible to follow a conversation in a noisy environment without an
> anticipatory mechanism.
>
> And please, Kurt, don't reply that your network already does that. It
> does not.

:)

The coctail party problem is an intersting one.

How we focus our attention visually is not so hard. We move our eyes. How
we focus our attention on work is not so hard, we go sit down at the desk
instead of sitting down in the recliner and watching TV. So all the "focus
attention" problems which are solved by exetrnal beahvior, is easy to
explain with a network who's sole purpose is to learn to create the correct
behavior like mine.

Also, the problem of choosing what to think about is also easily explained
by a network like mine when you set up "thinking" as just another learned
behavior. So when we "foucs" by thinking about only one subject, you can
explain that simply by a nework which is constantly selecting the correct
behavior.

But, the cocktail party problem is a simple example that seems to show
something more complex happening. It's a clear example where a single data
stream flowing into the system is filtered, and the system has dynamic
control over the operation of that filter. So it's a function that can not
be explained by the external manipulation of the enviornemnt like you can
by moving your eyes.

So to do this, the system must have the ablity to break a data stream down
into different elements, and then selectivly direct, or block, the flow of
those elements to other areas for processing.

My network certainly has the ablity to decode a complex signal into an
unlimited number of sub elements. And it has the ablity to control the
switching of those elements to other parts of the network. So at that
level, the network clearly has the basic power it needs to do this.

But, is the decoding, and switching, which my network can do, the correct
type of decoding and switching, to allow a complex coctail party audio
signal to be broken down to individual converstaions? I have no clue. I
think there's a good chance it can, but I'm no where near that level of
experimentation to know the answer to that.

Louis, one thing you problay don't undrstand about my design is that the
feed-forward network is not the end answer. It's just the start. Larger
feedback loops are needed. I just don't yet know how to best use them.
It's research I haven't even started on. But, it's key to our power to
have internal thoughts. And our ability to focus on a single converstaion
in a cocktail party (and switch from one focuse to another) is closely
related to the same feedback loops we use for having our own private
conversations.

It's the feedback loops which creates your "anticipation" system because
the data being feed back creates some percentate of the "environment" the
beahvior system is reacting to. The data following in from a feedback loop
is setting up a context which the pattern matching system is responding to.

SO, just by adding feedback loops to a network like mine, I know it's got
the basic system needed to the "cocktail party" attention focus task. But
I don't know if it can actually do it.

Another more general way to look at the attention focus problem is if you
have 100 sensory inputs, and 1 "arm" motor output, the system must at all
times make a decision about what sensory data is currently bing used to
control the arm. Should it for example use the arm to grab the food on the
table, or should it use it to scratch my nose, or should I use to to
continue typing this message?

If you have a single level S-R system you have the problem of conflict
resolution where multiple S-R rules might be telling the system to do
different things in parallel. One rule might be telling you to raise the
arm, where as another might be telling you to lower it. So how does this
conflict get resolved?

You do it by using a selection system. It must, at some level, make a
choice between which rule will get to control the arm, instead of letting
all S-R rules try to control it at the same time. My network makes these
decisions at the lowest level. When a pulse gets routed one way or the
other, the node that made the routing decision is actually picking, which
down-stream node, is going to be allowed to make the next decision. So
instead of all rules being used in parallel, the rules are actually used to
pick other rules, so by the very nature of the operation of the network,
there are never any conflicts to be resolved at the lowest level.

What can however happen, is that it can quickly keep changing it's mind.
It could for example, decide to lower the arm, then 1 ms later, switch to
raising it, then 1ms later, switch to lowering it again. However, if it's
not "good" to be doing this sequence of fast up down arm motions, the
system will get punished for doing that, and it will learn not to do that.
It will learn to produce a long sequence of up motions or a long sequence
of down motions in order to get something done like grabing the food on the
table. And the prime system my type of network will use to learn to
produce a sequence of outputs (unguided by the environemnt) is motor
feedback loops. It will, for example, build a S-R rule that says, to the
effect, "if my last arm motion was up, then do up again". It can build
rules like that if the "last arm motion" is fed back as input to the
network to base the next output on. This is why I say the "motor cortex"
needs global feedback. It's needed for learning of fixed patterns of
behavior.

But, the big probelm of "focus" happens in the general case when you have
lots of senory inputs "fighting" for the control of a single motor output -
like the "arm" example.

My network is designed from the bottom up to solve that "attention" problem
- the problem of which sensory data is sent to which output. And the
decsion is made by each node antecipating what might come next based on
what has happaned in the recent past.

So, to say that my network has no attention system or ability to focus
attention, is only showing how much you are unable to understand any
solution which happanes to be different from your own.

But, whether my network has any hope of correctly duplicating human powers
of following a single converstaion in a cocktail party, (or whether your's
has) is something beyond both of our abilities because neiether of us has
such a system in operation at the moment.

BTW, if you check my from line, you will see my name is not "Kurt". It's a
mistake many people seem to make.

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