tomh at BAMBI.CCS.FAU.EDU (Tom Holroyd) writes:
>Nice post, but I have some comments...
>>From: erwin at trwacs.fp.trw.com (Harry Erwin)
>>Stimulus invariance is associated with semantic invariance, but _not_ with
>>syntactic invariance (at the level of the olfactory nucleus). The _same_
>>pattern in the nucleus _means_ different things from month to month. I've
>>speculated that the cortex downloads a "semantic" model to the olfactory
>>nucleus, the nucleus then monitors the stimulus using the model, reporting
>>minor novelties (discrepancies between the model and the stimulus) and
>>major novelties (unmatchable stimulus patterns) to the cortex (in a
>I have a problem with the term "downloads". Does this mean that the
>cortex flips a switch in the olfactory nucleus and then dumps a semantic
>model into it? What is the switch? What causes it? What does the model
>look like? How is it transmitted to the nucleus? How is the model
>created in the cortex? How is it stored in the nucleus? How long does
>the transfer take and what's the baud rate? :-)
>Kidding aside, the concept of a "model" as an entity that can be transported
>around the brain leads to all sorts of problems. Homunculii, question
>begging, etc. The brain does not have a von Neumann architecture.
I agree with your criticisms, but there are two pieces of evidence that
have to be accounted for:
1. Short-term but not long-term invariance, and
2. The ability of the cortex to "prime" the olfactory nucleus to detect
specific olfactory patterns.
I suspect the cortex does "flip a switch in the olfactory nucleus and
then dump a semantic model into it." I even have some suspicions as to the
mechanism, since there is something somewhat similar going on in the
cerebellum. The switch might be a generalized activation pattern that
primes the nucleus for the arrival of a pattern. The pattern would be
stored by a competitive learning process, since patterns already present
would hold their own, and the new pattern would be picked up by free
components of the pattern-recognition network. I suspect there's a key
that the newly loaded network components use to recognize each other. I
even suspect the key is a "carrier wave" pattern that has been seen
in use elsewhere in the same role. Creation of the model in the cortex may
be similar to the creation of motor sequences in the cortex. The transfer
occurs in parallel and probably takes on the order of 100 msec. (If it
sounds like I've been thinking about how to build this system, it's purely
>The discussion of the chaotic dynamics of the olfactory system was wonderful!
>For example, you say:
>>Since they operate in a state space with hyperbolic fixed points, it's easy
>>to perturb them slightly and send them whereever you want. ...
>>you can perturb them as they approach a hyperbolic fixed point and send
>>them anywhere on their manifold. The energy needed to control a chaotic
>>process is many orders of magnitude less than the energy needed to control
>>a process that has a fixed point or stable limit cycle.
>This is a great picture, and shows that the system's behavior can be altered
>by small, subtle changes in the interactions of the nucleus with its
>So we have a picture of a complex dynamical system being driven by
>patterned inputs and interactions with other cortical regions. Then you
>go and download a model into it, as if it were a RAM chip. :-) It seems much
>clearer to me to continue with the dynamical reasoning: slow parametric
>changes in the olfactory nucleus alter the attractor layout - these are
>internal to the nucleus itself (like changes in synaptic strength).
>The changes can be both 'autonomous' structural changes in the nucleus
>itself, or as a result of coordinated pattern changes between the nucleus
>and the cortical regions it communicates with.
You complain about my positing rapid downloads, but that's necessary to
deal with rapidly evolving situations and to take account of learning.
Such processes also occur in the cortex--the mechanism of visual search
(search for a known pattern, not recognition of novelty) depends on a
download to the appropriate area of the inferior temporal cortex of the
image to be searched for. We know that pyramidial cells are highly
sophisticated processors. Synaptic switching in the dendritic arbor is a
significant, high-rate process; the pyramidial cells are known to do
complex computation, multiplexing the results in the spike train they
generate (where the information content is overlaid on a carrier wave that
is often used to index the data being transmitted); and individual
pyramidial cells are known to be able to remember hundreds of image
elements after a single presentation of each element. So we're dealing
with sophisticated processing elements, and positing that they can be
rapidly downloaded with information is not disturbing to me.
>Remember that the olfactory system has to cope with many problems during
>development and over the course of evolution. It is not known a priori
>which scents will be meaningful. It is not known a priori