By distinguishing characteristics? (This would include things like
little dings on your car or the junk in its back seat that distinquish
it from a similar make and model of the same years and color. )
If so, that is handled by Granger’s system in the manner described in my
response to the question copied below.
Now when you are dealing with objects that have an identical appearance,
such as Diet Coke cans (the example I normally use when I think of this
problem), often the only thing you can distinguish them by is – again –
their distinguishing characteristics. But in this case the
distinguishing characteristics would be things like their location,
orientation, or perhaps relationship to other objects. It would also
include implications that can properly be drawn from or about such
characteristics for the type of thing involved.
For example, if you leave a Diet Coke can (can_1) downstairs in your
kitchen and go up to you bedroom and see an identical looking coke can
next to your bed, you would normally assume the can next to your bed was
not can_1, unless you had some explanation for how can_1 was moved next
to your bed. (For purposes of dealing with the hardest part of the
problem we will assume all coke cans have been opened and have the same
amount of coke with roughly the same level of carbonation.) If you go
back down stairs and see a Diet Coke can exactly where you left can_1,
you will assume it is can_1, itself, barring some reason to believe the
can might have been replaced with another, such as if you know someone
was in your kitchen during your absence.
All these types of inferences are based on generalities, often important
broad generalities like the persistence of objects, that take the
learning of even more basic or more primiative generalities (such as
those needed for object recognition, understanding the concept of
physical objects, the ability to see similarities and dissimilarities
between objects, and spatial and temporal models), all of which take
millions of trillions of machine opps and weeks or months of experience
to learn. So I hope you will forgive me and Granger if we don’t explain
them in detail. (Goertzel in "Hidden Pattern", I think it is, actually
gives an example of how an AGI could learn object persistence.)
However, the whole notion of AGI is built on the premise that such
things can be learned by a machine architecture having certain
generalized capabilities and having something like the physical world to
interact in and with. Those of us who are bullish on AGI think we
already have a pretty good ideas how to make system that can have the
required capabilities to learn such broad generalities, or at least get
us much closer to such a system, so we can get a much better
understanding of what more is needed, and then try to add it.
With such ideas of how to make an AGI, it become much easier to map the
various aspects of it into known, or hypothesized, operations in the
brain. The features described in Granger’s paper, when combined with
other previous ideas on how the brain could function as an AGI, would
seem to describe a system having roughly the general capability to learn
and properly inference from all of the basic generalizations of the type
I described above, such as the persistence of objects, and what types of
objects move on their own, and with what probabilities under what
circumstances. For example, Granger's article explains how to learn
patterns, generalizations of pattersn, patterns of generalizations of
patterns, and with something like a hippocampus it could learn episodes,
and then patterns from episodes, and generalizations from patterns from
episodes, and patterns of generalazations from episodes, etc.
Yes, the Granger article, itself, does not describe all of the features
necessary for the brain to act as a general AGI, but when interpreted in
the context of enlightened AGI models, such as Novamente, and the
current knowledge and leading hypotheses in brain science, it is easy to
imagine how what he describes could play a very important role in
solving even mental problems as “DEEP” (again with reverb) as that of
determining whether the Diet Coke can on the table is the one you have
been drinking from, or someone else’s.
Has there been a little hand waving in the above explanation? Yes, but
if you have a good understanding of AGI and its brain equivalent, you
will understand the amount of hand waving is actually rather limited.
Ed Porter
============= from prior post ====================
“RICHARD>> “How does it cope with the instance/generic distinction?”
I assume after the most general cluster, or the cluster
having the most activation from the current feature set,
spreads its activation through the matrix loop, then the
cluster most activated by the remaining features spreads
activation through the matrix loop. This sequence can
continue to presumably any desired level of detail supported
by the current set of observed, remembered, or imagined
features to be communicated in the brain. The added detail
from such a sequence of descriptions would distinguish an
instance from a generic description reprsented by just one
such description..
A misunnderstanding: the question is how it can represent multiple
copies of a concept that occur in a situation without getting confused
about which is which. If the appearance of one chair in a scene causes
the [chair] neuron (or neurons, if they are a cluster) to fire, then
what happens when you walk into a chair factory? What happens when you
try to understand a sentence in which there are several nouns: does the
[noun] node fire more than before, and if it does, how does this help
you parse the sentence?
This is a DEEP issue: you cannot just say that this will be handled by
other neural machinery on top of the basic (neural-cluster =
representation of generic thing) idea, because that "other machinery is
nontrivial, and potentially it will require the original (neural-cluster
= representation of generic thing) idea to be abandoned completely.
-----Original Message-----
From: Richard Loosemore [_mailto:[EMAIL PROTECTED]
Sent: Monday, October 22, 2007 2:55 PM
To: agi@v2.listbox.com
Subject: Re: Bogus Neuroscience [WAS Re: [agi] Human memory and number
of synapses]
Edward W. Porter wrote:
Dear Readers of the RE: Bogus Neuroscience Thread,
Because I am the one responsible for bringing to the attention of this
list the Granger article (“Engines of the brain: The computational
instruction set of human cognition”, by Richard Granger) that has caused
the recent kerfuffle, this morning I took the time to do a reasonably
careful re-read of it.
[snip]
In his Sun 10/21/2007 2:12 PM post Richard Loosemore cited failure to
answer the following questions as indications of the paper’s
worthlessness.
“RICHARD>> “How does it cope with the instance/generic distinction?”
I assume after the most general cluster, or the cluster
having the most activation from the current feature set,
spreads its activation through the matrix loop, then the
cluster most activated by the remaining features spreads
activation through the matrix loop. This sequence can
continue to presumably any desired level of detail supported
by the current set of observed, remembered, or imagined
features to be communicated in the brain. The added detail
from such a sequence of descriptions would distinguish an
instance from a generic description reprsented by just one
such description..
A misunnderstanding: the question is how it can represent multiple
copies of a concept that occur in a situation without getting confused
about which is which. If the appearance of one chair in a scene causes
the [chair] neuron (or neurons, if they are a cluster) to fire, then
what happens when you walk into a chair factory? What happens when you
try to understand a sentence in which there are several nouns: does the
[noun] node fire more than before, and if it does, how does this help
you parse the sentence?
This is a DEEP issue: you cannot just say that this will be handled by
other neural machinery on top of the basic (neural-cluster =
representation of generic thing) idea, because that "other machinery is
nontrivial, and potentially it will require the original (neural-cluster
= representation of generic thing) idea to be abandoned completely.
“RICHARD>> “How does it allow top-down processes to operate in the
recognition process?”
I don’t think there was anything said about this, but the
need for, and presence in the brain of, both top-down and
bottom-up processes is so well know as to have properly been
assumed.
Granted, but in a system in which the final state is determined by
expectations as well as by incoming input, the dynamics of the system
are potentially completely different, and all of Granger's assertions
about the roles played by various neural structures may have to be
completely abandoned in order to make allowance for that new dynamic.
“RICHARD>> “How are relationships between instances encoded?” ”
I assume the readers will understand how it handles temporal
relationships (if you add the time dilation and compression
mentioned above). Spatial relationships would come from the
topology of V1 (but sensed spatial relationships can also be
build via a kohonen net SOM with temporal difference of
activiation time as the SOM’s similarity metric).
Similarly, other higher order relationships can be built
from patterns in the space of hierarchical gen/comp pats
networks derived from inputs in these two basic dimensions
of space and time plus in the dimensions defined by other
sensory, emotional, and motor inputs. [I consider motor
outputs as a type of input].
Again, no: relationships are extremely dynamic: any two concepts can
be linked by a relationship at any moment, so the specific question is,
if "things" are represented as clusters of neurons, how does the system
set up a temporary connection between those clusters, given that there
is not, in general, a direct link between any two neurons in the brain?
You cannot simply "strengthen" the link between your "artichoke"
neuron and your "basilisk" neuron in order to form the relationship
caused by my mention of both of them in the same sentence, because, in
general, there may not be any axons going from one to the other.
“RICHARD>> “How are relationships abstracted?”
By shared features. He addresses how clusters tend to form
automatically. These clusters are abstractions.
These are only clusters of "things". He has to address this issue
separately for "relationships" which are connections or links between
things. The question is about "types" of links, and about how there are
potentially an infinite number of different types of such links: how
are those different types represented and built and used? Again, a
simple neural connection is not good enough, because there would only be
one possible type of relationship in your thoughts.
“RICHARD>> “How does position-independent recognition occur?”
He deals with this. His nodes are nodes in a hierarchical
memory that provides degrees of position and shape
invariance, or the type mentioned by Hawkins and the Serre
paper I have cited so many times. Granger’s figures 6 and 7
indicates exactly this type of invariance.
I have not looked in detail at this, but how does his position
invariance scale up? For example, if I learn the new concept of "floo
powder", do I now have to build an entire set of neural machinery for
the all the possible positions on my retina where I might see "floo
powder"? If the answer is yes, the mechanism is bankrupt, as I am sure
you realise: we do not have that much neural machinery to dedicate to it.
“RICHARD>> “What about the main issue that usually devastates any
behaviorist-type proposal: patterns to be associated with other
patterns are first extracted from the input by some (invisible,
unacknowledged) preprocessor, but when the nature of this preprocessor
is examined carefully, it turns out that its job is far, far more
intelligent than the supposed association engine to which it delivers
its goods?
What he feeds to his system are things like the output of
Gabor filters. I don’t think a Gabor filter is something
that is “far, far, more intelligent than the supposed
association engine to which it delivers its goods.”
He has to show that the system is capable, by itself, of picking up
objects like the "letter A" in a scene without the programmer of the
simulation giving it some hint. The fact that he uses Gabor filters
does not bear on the issue, as far as I can see.
This issue is more subtle than the others. Too much for me to go into
in great detail, due to time constraints. Suffice it to say that you do
not really address the issue I had in mind.
This is just an example of how a serious attempt to understand what is
good in Granger’s paper, and to expand on those good features, overcomes
a significant number of the objections raised by those whose major
motivation seems to be to dismiss it.
I think I have shown that none of my objections were overcome, alas.
Wikipedia, that font of undisputed truth, defines Cognitive science as
“Cognitive science is most simply defined as the scientific
study either of mind or of intelligence (e.g. Luger 1994).
It is an interdisciplinary study drawing from relevant
fields including psychology, philosophy, neuroscience,
linguistics, anthropology, computer science, biology, and
physics”
Based on this definition I would say the cognitive science aspect of
Granger’s paper, although speculative and far from fully fleshed out, is
actually quite good.
Cognitive science is more than just saying a few things that seem to
come from a selction of these fields.
I would welcome further discussion of these issues, but it might be
better for me to point to some references in which they are discussed
properly, rather than for me to try to do the whole job here.
Richard Loosemore
-----
This list is sponsored by AGIRI: _http://www.agiri.org/email_
To unsubscribe or change your options, please go to:
_http://v2.listbox.com/member/?&_ <http://v2.listbox.com/member/?&>
------------------------------------------------------------------------
This list is sponsored by AGIRI: http://www.agiri.org/email
To unsubscribe or change your options, please go to:
http://v2.listbox.com/member/?&
<http://v2.listbox.com/member/?&>