One last try... God help me...
Let's assume a 640 x 480 visual field with black and white pixels for
simplicity.
There are two approaches here, let's take what I call the raw approach.
One frame is grabbed. Each pixel of the frame is asserted into the database.
Each pixel is asserted to the model using the proposition PERCEPT(Visual, '
PIXEL x y c '),for example
PERCEPT(Visual, ' PIXEL 240 320 W ')
==> percept-76855
where "percept-76855" identifies that a specific monad which represents a white
pixel at x,y location 240 x 320 has been activated.
Once the entire frame is asserted to the model. A concurrence activation
process looks to chunk groups of percepts that have the same activation time
but are not grouped together. A UNISON relationship is synthesized for each
such chunk indicating that these percept monads occurred at the same time. For
example,
UNISON( < percept-76855, percept-672232, percept-89004, percept-74325,
percept-20876, percept-45694, percept-33368 > )
==> unison-35437
More unison relationships are formed on this tier. Then unison relationships
are formed on the next tier. For example,
UNISON( < unison-35437, unison-89465, series-89033, unison-23455,
series-45688, series-98809, unison-78900 > )
==> unison-1245388
This goes on until we have a final tier which represents the entire frame, lets
call it unison-9426772
These monads and schemes (or monemes as I call them) are the elements that
comprise your image at this point.
We begin again with the next frame. But now, instead of having to construct
each percept and intervening scheme,many of the schemes and reifying monads
will be reused because they already exist in the model.
The second approach is to assert visual features rather than pixels but the
principal is the same.For example, we could use OpenCV to ascribe various
features to the incomincg frame and only assert those features. Hence,
PERCEPT(Visual, ' EDGE ... ')
==> percept-93453
PERCEPT(Visual, ' CORNER ... ')
==> percept-93454
PERCEPT(Visual, ' CIRCLE ... ')
==> percept-93455
So that these can be grouped at a higher level of visual perception.
So the elements you refer to are the percepts and intervening schemes (the
monemes) that comprise the total experience, within the current model.
Does that answer your question, if not why not? These are the elements.
Then there are transformations that occur upon these elements.
~PM.
From: [email protected]
To: [email protected]
Subject: Re: [agi] Re: Superficiality Produces Misunderstanding - Not Good
Enough
Date: Tue, 23 Oct 2012 21:58:51 +0100
PM,
You have to identify the **elements** or “monads” pace you that are common
to those chairs – or just begin to identify *one* element common to them. [No,
a
complete analysis is not expected]. You haven’t done this. And these figures
are
not hard to analyse.
Supplying me with a long architecture or analysis of your system does not
address that problem in any way at all. There is nothing in what you have
written that addresses: what are the elements common to the different examples
of a concept/ chair, or a visual object, (or scene, or text anything else). You
just **presuppose** that you can analyse them without the slightest attempt at
a
demonstration/instantiation..
The problem of AGI is that of multiformity/novel transformations - we
are always faced with different objects/scenes where there may be no common
configurations of common elements – where each example can be considered as a
creative, novel transformation of the last – (see those chairs again)
- and yet the brain has done what no AI system or technology has
achieved, and found a way to classify them together. You’re not addressing
that.
Neither AFAIK is anyone else.
From:
Piaget Modeler
Sent: Tuesday, October 23, 2012 6:47 PM
To: AGI
Subject: RE: [agi] Re: Superficiality Produces Misunderstanding -
Not Good Enough
I have a paper that details how this works, send me an email if
you'd like to get a copy.
In PAM-P2 percepts (whether visual, auditory, proprioceptive, etc.) are
asserted to the
current model. These assertions activate "monads". Monads in
turn activate schemes.
Each scheme has a reifying monad which is activated by the scheme according
to the
scheme's merge type (PASSthrough, AND, OR, NAND, NOR, NOT). The merge
type is
defined by the relationship that the scheme represents. Basic
relationships are
UNISON, SERIES, OPTION, CASE, TYPE, etc.
Activation flows along several dimensions: perception, expectation,
intention, thus
a monad can be activated in multiple ways. Percepts activate
monads along the
perception dimension. The entire image presented will ultimately
activate a single
reifier as it passes through several tiers of pattern abstraction. In
addition the
sound of the word "Chair" will be activated in sequence with the image and
ultimately the reifying monad for the image and the monad for the word will
be
bound together in a series scheme.
This will occur again for the remaining training instances. The
series schemes
will also be assessed by processes which will identify commonalities and
predictions.
As the training examples are repeated, predictions ensue and the monads are
activated along the expectation dimension. When predictions are satisfied
there is
there is a shift which occurs from from sequence to concurrency and also
from
sequence to optionality, which happens as part of automaticity.
Activation along the expectation dimension triggers simulation, whereby
expected
monads can be "visualized" or activated in the forward model. This
activation is
propagated throughout the forward model from reified monads down to
perceptual
monads.
Another type of simulation is also possible, but we'll save that for
another day.
Check out the site http://piagetmodeler.tumblr.com for some diagrams
of how
this works.
This is a good start.
Cheers.
~PM.
From: [email protected]
To: [email protected]
Subject: Re: [agi] Re:
Superficiality Produces Misunderstanding - Not Good Enough
Date: Tue, 23 Oct
2012 17:29:23 +0100
PM & Aaron,
You do realise that whatever semantic net system you use must apply to not
just one chair, but chair after chair – image after image?
Bearing that in mind, explain the elements of your semantic net which you
will use to analyse these fairly simple figures as **chairs**::
http://image.shutterstock.com/display_pic_with_logo/95781/95781,1218564477,2/stock-vector-modern-chair-vector-16059484.jpg
Let’s label these chairs 1-25 (going L to R from the top down, row
after row)
Start with just 1. and 2. top left and explain how your net will recognize
2 as another example of 1.
How IOW do you define a “chair” in terms of simple abstract forms?
Then we can apply your system, successively, to 3. 4. etc.
This is the problem that has defeated all AGI-ers and all psychologists and
philosophers so far.
But Aaron (and PM?) has a semantic net solution to it - if you
can solve jungle scenes, this should be a piece of cake.
I am saying, Aaron, you do not understand this problem – the problem
of visual object recognition/conceptualisation//applicability of semantic
nets.
You are saying you do – and it’s me who is confused. Show me.
-------------------------------------------
AGI
Archives: https://www.listbox.com/member/archive/303/=now
RSS Feed: https://www.listbox.com/member/archive/rss/303/21088071-c97d2393
Modify Your Subscription:
https://www.listbox.com/member/?member_id=21088071&id_secret=21088071-2484a968
Powered by Listbox: http://www.listbox.com
