Freeman Paper

Harry Erwin erwin at trwacs.fp.trw.com
Wed Nov 25 14:05:11 EST 1992


Reference: Y. Yao and W. J. Freeman, 1990. "Model of Biological Pattern
Recognition with Spatially Chaotic Dynamics." Neural Networks, 3:153-170.

I've been going through this paper in detail. It's very interesting. They
provide a lot of insight into how a sensory process works.

The mammalian olfactory system consists of three major subsystems: the
olfactory bulb (OB), with the sensory receptors synapsing onto the mitral
cells and periglomerular neurons of the OB nucleus in the glomeruli, and
with granule cells interacting with the mitral cells to form neural
oscillators; the prepyriform cortex (PC), which monitors breathing,
sends reset signals to the OB, and reports via deep pyramidal cells to the
external capsule; and the anterior olfactory nucleus (AON), which monitors
the patterns presented by the OB and handles pattern-related interactions
with the OB. In isolation, each subsystem is periodic, and it is the
interaction of the subsystems that creates the chaos. The connections are
low-speed, and serve as low-pass filters. The OB has the highest natural
frequency, the AON is intermediate, and the PC is lowest.

The function of the OB is to convert the signals from the sensory neurons
into a stable pattern that the PC can recognize. It is organized as
content addressible memory. It reports a basal state if there is no input,
a chaotic state if there is an unrecognizable input, and a near-periodic
state if the input matches a pattern it knows. 

The PC monitors the breathing cycle and sends "mute" signals to the
granule cells of the OB to pull the system into a neutral "ready" state.
It can inhibit the AON and reports the "object" defined by the OB pattern
to the cerebral cortex.

The AON conducts training for the OB. This it does by sending synthetic
sensory data to the periglomerular neurons and resetting the OB with
"mute" signals similar to those provided by the PC. This training process
(what I previously called "downloading") is very interesting.

First, as Desimone demonstrated in September, neurons report not just
their conclusions, but also the evidence for those conclusions. This turns
out to be important. The AON can pull the OB to a neutral state by
afferent synapses onto the granule cells. Simultaneously, it can feed
synthetic sensory data into the mitral cells (which the sensory neurons
synapse onto as well). Where does it get the correlation between the
sensory data and what the OB is reporting to it? The OB reports not just
patterns, but the evidence for those patterns. Hence the AON can learn
how to simulate any sensory input.

The PC feeds the cortex further downstream with pattern data. It is also
able to detect the neutral state and the novel object states. Since both
it and the AON receive the sensory data associated with the novel object,
the PC can command the AON to feed it back, simultanously  pull the OB
to a neutral state, and so train the OB to recognize the pattern. The PC
correlates the pattern with the object it was trying to train the OB on,
using the retained sensory data. (Remember, neurons are _good_ at learning
patterns. Visual system neurons can retain hundreds of individual features
on one presentation each. This seems to involve synaptic switching, rather
than synaptic growth.) These patterns don't need to be learned permanently;
it's good enough to retain them for a period of time and then discard them.
Hence, the lack of invariance. The AON can instead use synthetic sensory
data to find out what the OB currently reports for a given pattern of data.

Neat, huh?! The OB does not do the model update. Instead, it passes
through the data needed downstream to do those updates. It appears that
semantic data networks are the way our brains naturally organize
information. Logic is not "natural." Instead what is natural is
overlapping cognitive domains.

This is a little like what I do when I redirect processed sensory input
through a sensory processing area. This is also how higher cortical
functions might "envision" manipulating a scene or object--rotating it,
etc. It also implies that the mechanism for converting sound into speech
is dumb and involves the maintenance of a "library" of words and
corresponding sound sequences in a CAM similar to the OB.

Cheers,


-- 
Harry Erwin
Internet: erwin at trwacs.fp.trw.com



More information about the Neur-sci mailing list