I have been studying variations on Freeman and Yao's KII/KIII models of
the olfactory system ("Freeman Networks") as a possible basis for pattern
recognition processes to be used in air and ground vehicle control
systems. This morning, I finally got a model working that generalizes
the KIII model by replacing the simple AON and PPC models used by Freeman
and Yao with content addressable memory arrays like their OB model. This
model is designed so that all components, including the deep pyramidal
cells at the back end, correspond in a one-to-one way with the sensory
input channels (somewhat similar to the KIII model--I know this is a
massive simplification of reality.). This model has some interesting
dynamics.
The nominal dynamic relationships between the C (deep pyramidal) cells
are chaotic. However, if the Olfactory Bulb (OB) has been trained to
recognize a specific pattern of activations, and that pattern occurs in
the input, the C cells corresponding to the pattern become synchronized,
although the other cells remain in chaotic relationships. I have not yet
been able to transmit the input activation levels through the network to
the C cells so that the input pattern appears at the output (modulated
by the chaotic "carrier waves"), but I expect that by reducing the
couplings between cells in the content addressible memories, I will be
able to avoid the averaging process that I get now.
Implications:
1. It appears that semantics are carried by spatial synchronizations
between cell excitation patterns. This means that we recognize objects
(and semantic concepts) when multiple cells that are normally
uncorrelated become correlated. Individual pulse trains carry no
semantics. It is when multiple pulse trains have the same "carrier wave"
that we have semantics.
2. Qualia, on the other hand, I suggest are carried by the activation
intensities (pulse rate) of the individual cells. Note that the sensory
data from the olfactory receptors are forwarded on to the cortex. I've
decided this is required for the cortex to be able to request pattern
matching later.
3. The Olfactory Bulb (OB), Anterior Olfactory Nucleus (AON),
PrePyriform Complex (PPC) system appears to be an early processing
point that serves to recognize semantic objects (based on patterns of
sensory stimuli) for immediate action. I conjecture that this system
basically passes the sensory data through to the cortex as invariant
signals, superimposing its own in-band signalling (by synchronizing the
pattern of specific components of the sensory data) when the sensory
pattern is recognizable. Lack of invariance is specifically associated
with this in-band signalling.
4. This model immediately suggests mechanisms for data processing in the
cortex, and tends to confirm Barry Richmond's previous results.
Some questions emerge:
1. For Leslie Kay--is this consistent with what Freeman has been
seeing?
2. For Marvin Minsky--how does this interact with your "Society
of Mind" model?
I'm currently trying to pull this together for a poster session
paper at the 2nd Experimental Chaos Conference. Once I have
permission for public release, I'll post a notice so interested
people can request a copy.
Harry Erwin
Preferred address: erwin at trwacs.fp.trw.com
Also reachable at: cs103_03 at gmuvax.gmu.edu (where I'm teaching)