[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: subtraction and distribution function (was ACI/TI)

Dear Al, Chris, Richard and list

Just wanted to second what Chris wrote about interpreting brain activations. Very few people hold on to the view of single brain entity doing a single task. The buzz word for a while has been "distributed network". Thinking of distributed network allows us to use techniques like functional connectivity that measure strengths of connections between regions, across or within conditions and subjects.

At the risk of tooting our own horn, let me tell you about the large-scale multiple-level type of modeling that we do, which allows us to interpret the fMRI data and investigate mechanisms underlying the activations. This goes beyond data-fitting in that the assumptions we make are based on cellular level data and the predictions we make are matched or tested against fMRI data.

I chose just one type of the auditory continuity illusions alluded by Richard Warren (Btw, an excellent book and chapter, I seem to have highlighted almost all of it). The reason glides were chosen because I already had a model (matched with fMRI data) that processed glides. So all I had to do was input the ACI stimuli. And it worked. So now the question is why did it work? I did not create the model to accomplish this illusion nor were any parameters changed.

This what I think: Yes, the periphery is involved, but it is not the "seat" or primary source of this illusion. I would guess MGN or higher, and in fact would put my money (btw, I am looking for a job) on a distributed network of regions with the Superior temporal gyrus/sulcus at the apex (in a hierarchical sort of way). The divergent connections from
MGN->Primary Auditory Cortex -> Secondary Auditory Cortex -> STG/S result increasing "fan-out" which in turn results in increasing spectro-temporal windows of integration. (The sources for this claim are listed in Husain et al, 2004, Neuroimage). This allows the system to fill in over the noise-filled gaps in the stimuli. I tried different ways to "break" the illusion. I decreased the lateral inhibition in the model in primary auditory cortex (details at Husain et al, Journal of Cognitive Neuroscience, 2005). But this caused the cells to not select for direction of the glides. Then I decreased the fan-out (the divergence of the connections) and found that while the sweeps were processed fine, the illusion disappeard. And in my model, the crucial connection was from PAC to SAC and from SAC to STG/S.
(I have speculated on other types of illusion like phonemic restoration but because I have not modeled them, they will remain speculations)

I can of course be wrong. But this is a testable hypothesis and we hope to have more data within a few years to verify it. And I think that some version of this mechanism (increasing spectro-temporal windows of integration) is the key, whether it is cortical (as I say) or
somewhere else in the central auditory system.

Regards, --fatima

Fatima T. Husain, Ph.D.
Research Fellow
Brain Imaging and Modeling Section

Date: Tue, 13 Dec 2005 19:47:28 -0500 From: Al Bregman <bregman@xxxxxxxxxxxxxxxxxxxx> Subject: Re: subtraction and distributed function [was Re: The Auditory Continuity Illusion/Temporal Induction]

Dear Chris,

Thank you for your thoughtful reply.  It is clear that using imaging data
is a lot more complicated than people think.  You wrote:

"The question, as you suggest, should be "how" distributed
functions take place, and is unlikely to be answered by any one method,
but rather through the development of computational models on the basis
of descriptive data about the brain and behavior."

The approach that you suggest (fitting models to data) could be a
productive one, assuming that the models are based on assumptions about
actual types of processing rather than being of the "hip bone is connected
to the ankle bone" type (North American readers will know the song).

Best wishes,


Albert S. Bregman,
Emeritus Professor
Psychology Dept., McGill University
1205 Docteur Penfield Avenue
Montreal, Quebec
Canada  H3A 1B1

Voice & Fax: +1 (514) 484-2592

Aug 15 - Sept 15 annually:
        Voice & Fax  +1 (207) 729-0986

----- Original Message -----
From: "Chris Stecker" <cstecker@xxxxxxxxx>
To: <AUDITORY@xxxxxxxxxxxxxxx>
Sent: Monday, December 12, 2005 2:57 PM
Subject: subtraction and distributed function [was Re: The Auditory
Continuity Illusion/Temporal Induction]

Hello Al and list,

I'd just like to interject that the difficulty (folly?) of localizing
distributed functions in the brain is not lost on (all) fMRI
researchers. I've tried to focus attention on this issue when describing
to others the direction of my own work in auditory fMRI. There are
really two kinds of approaches through which we could imagine using the
subtraction technique. The first depends on assuming modularity of
function in the brain. If that assumption holds, then two carefully
designed conditions---one that includes the process of interest, and one
identical to the first except for the lack of that process---can be used
via subtraction to identify the module in question. (This is of course
not a new idea and has been used to look at chronometry, etc. for quite
some time now.) Many of us, however, doubt the applicability of strict
modularity, and instead recognize that---even with perfectly designed
conditions---many things are likely to change in related modules (if one
expects modularity to hold weakly) or throughout an extremely
distributed system that subserves the process of interest along with
many others (if one does not). The second approach is to use subtraction
as a tool to examine the effects of specific manipulations upon the
"activations" observed in the brain, and interpret those effects mainly
in a descriptive sense. For example, we might look at tone-evoked
activations at different sound levels. We might use subtraction to
isolate sound-related activity ("sound" - "silence") which we then
compare across levels, or we might directly compare activations produced
at different levels ("intense" - "soft"). Either way, we try to describe
the sensitivity to tone level in different regions of the brain (which
might in turn tell us a lot about the kinds of computations each region
is likely to be involved in) rather than to localize the "module that
processes sound level" or the "module that processes intense sound." [It
might also be worth pointing out that "subtraction" in this case can be
replaced by correlation or another statistical technique for assessing
sensitivity to manipulation of the independent variable across
potentially many levels; I'm not sure how to interpret such data via the
modularity assumption, but it makes good sense for the descriptive

Personally, I think the evidence for distributed function (throughout
the auditory cortex at least) is pretty good, and strongly prefer the
second, descriptive, approach to interpreting fMRI data. So I agree with
your post, especially regarding concerns about designing appropriate
conditions for subtraction and about the interpretation of subtraction
results via an implicit assumption of modularity. But I want list
members to realize that the method of subtraction can be employed
without that assumption, even though they are often encountered
together. The question, as you suggest, should be "how" distributed
functions take place, and is unlikely to be answered by any one method,
but rather through the development of computational models on the basis
of descriptive data about the brain and behavior.

-Chris Stecker