Semantics and Syntactic Representations

Harry Erwin erwin at trwacs.fp.trw.com
Wed Dec 1 11:17:28 EST 1993


This is in response to Leslie Kay's comments.

Thanks for the clarifications and corrections. It's the _simulation_ work
that seems to show pattern completion and matching processes. I'm aware
that there is significant divergence between my simulation models and the
biological studies (such as your work in progress), but I am
becoming more and more convinced from an engineering perspective that
reafference has to be understood as a process that (in part) inverts the
representation of the data in the deeper cortices to regain the
initial sensory signal. If the OB is a biological content addressible
memory (CAM) that is trained in real-time to track patterns of interest
and ignore patterns not of interest (through conditioning and habituation),
there has to be some mechanism for delivering the conditioning stimuli
when a pattern of interest is present. This is not sensitization via
interneurons, since only the pattern of interest is attended to. Either
the mitral cells in the OB are globally excited when the OB has matched
a pattern of interest, thus conditioning the inter-mitral cell connections
that are activated at the same time, or the specific mitral cells involved
in the pattern of interest are specifically excited, again conditioning the
connections of interest. The first mechanism has a number of drawbacks as
an engineering solution. First and foremost, it places a severe constraint
on the maximum response time of the loop from the OB to the deeper cortices
and back. Given the general 'trashiness' of olfactory data as presented to
the OB, this loop must complete before the next breath is taken to ensure
that the appropriate pattern is conditioned. This is particularly
important since the OB/PPC/AON loop modulates that pattern (by 'chaotic
resonance') in the simulation models of the KIII network.

The second mechanism is the preferred engineering solution, since it is
compatible with delayed reafference, to the point that the deeper cortices
can cue the OB/AON/PPC system to seek a known pattern of smells even when
that pattern has been 'forgotten' by the OB. Given the new evidence for
processes similar to the OB/AON/PPC system in other sensory modalities,
and the capabilities of the visual and auditory systems for known object
search, we have to address 'cueing' as a capability of reafference.

I'm currently investigating how reafferent stimulation of the inhibitory
'granule' cells in the AON can be used to scale the modulated signal
generated by the excitatory 'mitral' cells in the AON, since those
generate the reafferent signals to the OB in the KIII model. No results
yet, although in a non-oscillatory model of the AON, this mechanism
controls the amplitude of the signals back to the OB.

I understand your concern about the connection of this analysis with your
work in progress. My specific interest is the simulation of the KIII
network architecture, with the specific goal of gaining an understanding
of the engineering constraints on that network. The key issue is how it is
trained, which involves reafference, but from an engineering perspective,
not from a biological perspective. In the long run, the two streams of
investigation will have to be cross-validated. I am looking forward to
seeing your results when they are published.
-- 
Harry Erwin
Internet: herwin at cs.gmu.edu (not erwin at trwacs.fp.trw.com!)
Working on Freeman nets....



More information about the Neur-sci mailing list