Re: Invariant Recognition, Grandmother Cell, and Memory Hierarchy
- From: "dan michaels" <feedbackdroids@xxxxxxxxx>
- Date: 30 Jul 2005 14:30:25 -0700
>
>Hawkins further writes about the importance of prediction and the need
>to organize memory into feedback-driven hierarchies.
>
>
>I agree with Hawkins that invariance is an important aspect of
>intelligence. No moderately complex intelligent organism could
>function without this ability. I also agree with him that memory is
>organized into sequences that are used for making predictions.
>However, I strongly disagree that hierarchy is achieved via direct
>memory feedback.
Hi Louis. I don't think that's what HAwkins is saying. From
reading Hawkin's book and my own knowledge of the visual system,
I have a very different idea of how it works from what you indicate
here. Namely, feedback really has nothing to do with the hierarchy.
Rather, the hierarchy is actually a result of the "feedforward"
pathways, not the feedback. The hierarchy starts in the retina
[photoreceptors -> bipolar cells -> ganglion cells], and proceeds
to the LGN, then to cortical regions [first V1, and then V2, V4,
MT, on and on], and from there to the association areas in the
frontal lobes, etc. This hierarchy was first recognized by the
neuroanatomists, starting 100+ years ago, and really is
independent of feedback. Dave van Essen identifies at least 14
levels in the hierarchy of the visual system.
In addition, regards invariance, Hawkins is simply following the
lead of 50 years of neurophysiology studies, which in fact do
indicate that more and more abstraction is seen in the cell
responses as you go up the chain. Photoreceptors are simply pixel
[light dot] transducers, but bipolar cells are already the first
unit in the chain of information abstraction. They interact with
horizontal cells to extract information about "simultaneous-
contrast" in the visual field. Retinal ganaglion cells are
another level of this. Then up in V1, you have simple, complex
and binocular disparity cells. V4 codes for color. MT codes
for motion. Then, you have IT [IIRC] as the area that codes
for faces. On and on. IOW, each successive level you go into
the visual system, you have more abstraction and invariance
in the cell responses. At least 14 times.
This filtering and abstraction is probably mainly a feedforward
phenomenon. Furthermore, it is all dictated by genetics, and
probably has nothing whatsoever to do with memory, and little to
do with feedabck.
So, going up the chain you have more and more abstraction in
the responses. Feedback, as Hawkins describes it, I think,
mainly has to do with the "prediction" part of his model, not
so much formation of hierarchy and abstraction. As you go up
the chain, you have more and more abstraction in the
responses as a result of the genetically-dictated hierarchy.
Then, each level can send feedback signals back down the chain
to the previous level - ie, more abstract information is being
sent down to where the information is less abstract. This produces
comparison, and reinforcement, and tuning effects. And I also
think it has a *LOT* to do with temporal persistence. Feedback
is a dynamic temporal process, not just spatial.
Then, memory comes into play more and more towards the higher
levels of the hierarchy. After the signals have gone through
many levels of filtering and abstraction, they hit the memory
banks, and produce an output [I won't go into how associative
memory works], and this output provides the ultimate feedback
from the highest levels, which goes back down the hierarchical
chain as "prediction" information.
So, this is it. The hierarchy is anatomical, and is feedforward,
and is dictated by genetics and *nothing* else. It's a result
of evolution. The existence of feedback is also anatomical and
dictated by genetics. It's role is to carry predictions and not
to create abstractions/invariances. Memory comes into play
mainly at the higher most levels, and carries the experiential
information of the organism. Eg, in IT, you have face cells
and grandmother cells. IT is maybe 8 or 9 or 10 levels up in
the hierarchy of abstractions. It's high enuf that the memory
areas directly impinge upon it.
So, one theoretical way to get a cell that codes for your
own grandmother's particular face is to have a generic face
cell [whose connections are really dictated by genetics],
and to have output from associational memory directly
feeding back onto it. The feedback working in conjunction
with the basic operation of face-coding gives you a cell
that responds to gramma. Feedback is a differencing, comparison,
and resonance phenomenon.
All in all, if this isn't exactly what Hawkins has in mind,
it is how I see that the system might work. H calls it a
"memory-prediction" system, but I call it a "memory-prediction-
abstraction" system. There are really 3 different things going
on - at the very least, and each to a large extent has its
own part of the anatomy.
.
- Follow-Ups:
- Re: Invariant Recognition, Grandmother Cell, and Memory Hierarchy
- From: Gary Frank
- Re: Invariant Recognition, Grandmother Cell, and Memory Hierarchy
- From: Traveler
- Re: Invariant Recognition, Grandmother Cell, and Memory Hierarchy
- References:
- Prev by Date: Re: On blue squares
- Next by Date: Re: On blue squares
- Previous by thread: Re: Invariant Recognition, Grandmother Cell, and Memory Hierarchy
- Next by thread: Re: Invariant Recognition, Grandmother Cell, and Memory Hierarchy
- Index(es):
Relevant Pages
|
