Invariant Recognition, Grandmother Cell, and Memory Hierarchy
- From: Traveler <traveler@xxxxxxxxxx>
- Date: Fri, 29 Jul 2005 14:32:29 -0400
In his book, "On Intelligence", Jeff Hawkins mentions the brain's
ability to recognize moving objects from various angles and under
changing lighting conditions. He writes:
[...] the problem of understanding how your cortex forms
invariant representations remains one of the biggest
mysteries in all of science. [...] So much so that no one,
not even using the most powerful computers in the world,
has been able to solve it. And it isn't for a lack of trying.
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.
The problem with hierarchical feedback is that it is too rigid. Let's
take the grandmother cell example. What feedback scheme could possibly
account for this? Consider that a grandmother cell fires continually
as long as grandma stays in the picture, regardless of distance,
orientation, lighting, movements, etc... Lately, as a result of my
continuing research in memory organization, my understanding of memory
organization (both STM and LTM) has risen to a point where I can
confidently claim that the solution to the invariance problem is
finally here. I believe that invariance is the result of two memory
mechanisms.
First, memory is organized into a huge number of individual sequences,
each consisting of seven temporally correlated nodes. This means that
the timing of a sequence is not static but dynamic. Hence, a given
sequence can be used to encode a sizeable range of temporal signatures
associated with the same phenomenon. An example of this is the ability
to recognize a learned melody regardless of tempo.
Second, an entire group of non-conflicting or related sequences
representing a complex and dynamic sensory object can be activated
(awakened) by a single parent neuron. Once activated, the group will
stay awake for as long as the object remains within the sensory space.
The parent control neuron is what is known in the AI community as a
"grandmother cell".
In this scheme, there is no strict, unyielding hierarchy wired into
the connectionist structure of memory during perceptual learning. More
importantly, hierarchy is not encoded with the help of feedback loops,
as Mr. Hawkins and others suppose. Contrary to conventional wisdom, a
grandmother cell does not emerge from a group of perceptual neurons
using a hierarchical feedback tree. Rather, it uses principles of
motor coordination to assemble the group. Memory sequences are useless
unless they are tied to motor behavior. Behavior groups continually
wake up and go to sleep (this is part of the attention mechanism)
depending on sensory circumstances and motor conflict resolution. The
result is that memory is divided into a huge number of small and
self-coherent behavioral clusters.
In conclusion, I predict that memory hierarchy is not a consequence of
perceptual learning but that of motor learning (an aspect of
intelligence that Hawkins almost completely ignores). I realize that
my hypothesis is rather counterintuitive but I am willing to go out on
a limb and stick to my claim. It is for these reasons that I believe
that Numenta, Inc. (Hawkins' recently formed AI company) will
ultimately fail. Unless, of course, they change their approach.
Links to Hawkins's sites:
http://www.onintelligence.org/
http://www.numenta.com/
Louis Savain
Why Software Is Bad and What We Can Do to Fix It:
http://www.rebelscience.org/Cosas/Reliability.htm
.
- Follow-Ups:
- Re: Invariant Recognition, Grandmother Cell, and Memory Hierarchy
- From: HMS Beagle
- Re: Invariant Recognition, Grandmother Cell, and Memory Hierarchy
- From: dan michaels
- Re: Invariant Recognition, Grandmother Cell, and Memory Hierarchy
- From: Gary Frank
- Re: Invariant Recognition, Grandmother Cell, and Memory Hierarchy
- Prev by Date: Re: new AI movie today
- Next by Date: Re: On blue squares
- Previous by thread: Experience versus Programming
- Next by thread: Re: Invariant Recognition, Grandmother Cell, and Memory Hierarchy
- Index(es):
Relevant Pages
|
