Unexpected hypotheses arising from brain circuit simulation

[Richard Granger]

  Jim writes:
> the idea that cortical (olfactory) processing involved iterative
> response refinement and specificity was actually a central thesis of the
> monograph published by Lynch in 1986.
>
>Reference:  G. Lynch, Synapses, Circuits, and the beginnings of memory. MIT
>Press. 1986

If that's so, then its author doesn't know it.  The monograph neither
contains nor presages the hypothesis that appears in our 1990 Science
paper.  (Perhaps this is being confused with operations of excitatory
associational feedback fibers, which Lynch in 1986 hypothesized might cycle
repeatedly in response to a single input, rapidly building a
"representation" of an odor.  Such an operation is of course utterly
unrelated to the finding being discussed: there's no mention of bulb, no
mention of theta cycles, no mention of inhibitory feedback, no hierarchy.
The author states that the hypothesis that appears in the Science paper was
not even conceived of at the time of the 1986 monograph.)

The 1986 monograph was an early and fruitful step in the field of modeling
of real biological systems.  It is replete with attempts at identifying
computational concomitants of a range of biological phenomena, and with
compiling and integrating data related to research on LTP, the olfactory
system, and the hippocampus.  In particular, a central phenomenon is the
(4-8 Hz) theta rhythm, which:
  i) entrains cells throughout the olfactory system, hippocampus and much
of neocortex exclusively during learning and exploration (not during sleep,
or in a home cage, or passive restraint, etc.), (Macrides, 1975; Macrides
et al., 1982; Komisaruk, 1970; Otto et al., 1991; Kauer 1991), and
  ii) has been shown to be the optimal stimulation pattern for induction of
LTP (Larson et al., 1986; Diamond et al., 1987) due to an endogenously
occurring time-dependent gaba-b inactivation of gaba terminals (Mott &
Lewis, 1991).

For historical interest, after the monograph, what one finds is a
succession of papers by us, all on studies of this system (two papers in
1988 and four in '89), all attempting to identify various aspects of
function from the structure and operation of the system.  Our studies
focused in particular on the theta rhythm (such as the marvelous fact that
in small mammals the rhythm literally drives overt behavior (sniffing,
moving whiskers, etc., all at 5 Hz) during exploration).  Those papers make
the historical case glaringly:  they contain some interesting early
findings on the computational advantages of the synchrony itself provided
by theta, and on local lateral inhibition, and on refinement of recognition
with episodes of training and incremental steps of LTP, but nothing at all
on hierarchical clustering (the topic of the 1990 Science paper).   In
particular, the 1988 papers made the point that the initial sniff clustered
inputs (as was expected from work by other researchers), but nothing on
later sniffs subclustering.

Then we took on the puzzling observation that the cortical model was
producing different outputs over time, as it sampled a single input.  We
showed that not only do the initial cortical responses tend to empirically
cluster inputs, as mentioned, but also that later cortical responses,
paradoxically, tend to differentiate those same inputs.  We observed that
the system's inhibitory feedback (Price, 1973), and the long-lasting nature
of the resulting bulb granule cell IPSPs (Nicoll, 1969; Mori, 1987) tended
to selectively remove the same portions of the inputs that were giving rise
to the initial responses, and that therefore the system might not simply be
first clustering and then differentiating, but actually clustering,
sub-clustering, sub-sub-clustering, etc., over iterative cycles of the
theta rhythm.  This constituted a new hypothesis; not successive episodes
of learning; not steps of LTP; not cycling associational pathways; but
successive iterative feedforward excitation and feedback inhibition, giving
rise to a sequence of distinct different outputs on successive theta
cycles, traversing a hierarchical tree from general (clusters) to specific
(subclusters).  We wrote this up in 1989 and submitted it to Science, where
it was accepted and appeared in 1990.  It turned out that a relationship
could be shown between this circuit behavior and a nonstandard method of
hierarchical clustering; and it further turned out that this method was
unusually efficient with respect to time and space complexity.

Thus a puzzling operation that arose from interactions among a large number
of features in a biological simulation, turned out to yield an unusual
computational function, and did so in an unusually efficient manner.  It
was this finding that was surprising (to us, and to the reviewers and
editors of Science).

>It was clearly the objective of the subsequent model by Granger and Lynch
>to see if this specific idea could be incorporated into a "cortical like"
>structure.

Clearly not, given the above.  It was clearly our objective to explore the
model for its behaviors and functions, and to attempt to identify and
characterize the emergent properties of its many, many parts.


Our hypothesis is that a function of the olfactory system is the iterative
decomposition of inputs over successive rhythmic cycles of operation:
specifically, that sequential outputs of the cortex (every 200 msec) give
different information, corresponding to successive hierarchical clusters
and subclusters.  This is a novel hypothesis (or was, in 1990), and has
been much cited and studied for its behavioral, neurobiological,
psychological, and computational consequences.  It was derived directly
from, and incorporates, the detailed characteristics of the biological
system itself: theta, LTP induction and expression rules, sparse
connectivity, feedforward excitation, LOT and associational pathways,
feedback inhibition, mitral, tufted and granule cells, local lateral
inhibition, differential time courses of EPSPs and IPSPs in different cell
types, axonal arborization radius of inhibitory interneurons, etc.  We'd be
glad simply to take the credit for having come up with it by inspection, if
that were true.  The fact that we actually came up with it only after
extensive construction and observation of anatomically and physiologically
realistic models is a notable methodological point.  That this new idea has
given rise to a fruitful series of subsequent behavioral, physiological and
theoretical studies (in our labs and others) is a notable consequence.
That the hypothesis will undoubtedly turn out to be wrong in many ways is
the point of ongoing scientific inquiry.  We should continue to doubt, and
to study, and to counter-hypothesize, and to experiment.

-Rick Granger



[Footnote:

>In fact, as I remember, the Granger model assumed that the only LTP was in
>the synaptic connections made by the Lateral olfactory tract (LOT) not in
>the association fiber system.  Kanter and Haberly (1990) actually showed
>that the association fiber system is the major source of LTP in olfactory
>cortex.

No, actually, our models have LTP both in the LOT and the association fiber
system.  We have even studied the individual effects of these two pathways,
and their possible differential contributions to the function of the
overall system.]