LESSONS FOR METHODOLOGY IN COGNITIVE NEUROSCIENCE
FROM A COMBINED NEUROCOMPUTATIONAL MODELING-fMRI PROJECT
John Bickle,
Department of Philosophy and Neuroscience Graduate Program,
University of Cincinnati, P.O. Box 210374,
Cincinnati, OH 45221-0374.
Email: bicklejw@email.uc.edu
Two prominent methodologies in recent cognitive neuroscience
are neurocomputational modeling (often coupled with computer simulation)
and functional neural imaging. Unfortunately, these methodologies are typically
employed in isolation from each other. Yet they can be combined to overcome
some methodological shortcomings of each. I demonstrate this point empirically
by describing some recent results from a joint neurocomputational—functional
magnetic resonance imaging (fMRI) project (Bickle et al. 2001). This
project begins with published models and computer simulations of the cellular
mechanisms of frontal circuits that compute sequences of saccades from an
initial fixation point (Bickle et al. 2000; Bernstein et al
. 2000). Those models were built up from details revealed by anatomical and
single-cell physiological studies of primate frontal eye fields (FEFs), various
frontal “working memory” regions, and anterior cingulate cortex (ACC). Activity
by a computer simulation of the full model reflects realistic primate saccade
sequencing.
However, in order to get realistic performance from our
simulation, we had to choose some architectural and computational parameters
that cannot at present be justified neurobiologically. This is typical of
most neurocomputational work constrained by “low level” cell-physiological
data. In keeping with our explicit neurocomputational focus, we treat
these assumptions as predictions about neural processing to be tested
by the appropriate laboratory methods. A number of these predictions from
the saccade command model turn out to testable using fMRI.
To do so, we developed a sequential saccade task (block
design) of increasing complexity. Subjects executed four cycles of blocks
of 2-, 3-, and 4-step saccades and a control motor task. All saccade targets
on each run appeared and extinguished during the latency of the first saccade.
Subjects were asked to saccade to the targets in the order of presentation.
During the entire 8.5 minutes (510 seconds) of the four cycles, we took 340
fMRI data points (one full brain image in 24 slices every 1500 milliseconds).
Activation data were acquired at 3T using BOLD (Blood Oxygen Level Dependent)-sensitized
T2*-weighted gradient-echo EPI. (Technical details will be presented and
explained during my talk.) A 3-D Modified Driven Equilibrium Fourier Transform
whole brain scan was performed in an axial plane to provide high-resolution
anatomical images for co-registration of activation maps. Image post-processing
was performed using Cincinnati Childrens Hospital Image Processing Software
(CCHIPS©) developed in the IDL® software environment.
Cross-correlation between the BOLD signal intensity time course and the reference
function was performed for each subject on a pixel-by-pixel basis; pixels
above a statistically appropriate correlation threshold were overlaid on
the co-registered anatomical image. In this manner we computed activation
maps for each subject. These statistical parametric maps were then transformed
into Talairach space for composite analysis.
To identify frontal regions subserving sequential saccadic
activity, we first identified regions that were significantly activated during
saccade activity in any of the contrasts with the control motor task. Then
we identified voxels in the grouped data where the amplitude of the BOLD-sensitized
fMRI signal was correlated significantly (p < 0.01) with the number of
saccade steps per trial (i.e., with increasing “saccade burden”). These strategies
clearly identified FEFs, ACC, and two regions of frontal lobe previously
associated with working memory functions in humans. This verified the areas
from which single-cell data was gathered in constructing our neurocomputational
model. We then plotted time courses for each of these regions in each subject,
graphing level of BOLD signal intensity against fMRI data point. These results
validated a computational assumption-cum-biological prediction of our model,
namely that FEF activity increases montonically as saccade burden increases.
However, this analysis also disconfirmed an important assumption of our model.
We had assumed, based on the single-cell data, that FEFs only have the capacity
to compute two-step saccade sequences, and had to rely on frontal working
memory areas to store dimensions of earlier saccades when saccade sequences
involved more than two steps. Yet we didn’t see significant activity in frontal
“working memory” areas until the four-step task. This result forces us “back
to the drawing board” to increase the biological plausibility of our neurocomputational
model in light of this new data.
Besides their intrinsic interest, results from our combined
neurocompuational modeling—fMRI study have two broader methodological/philosophical
implications. First, we now have a piece of actual scientific research that
treats purely computational and architectural assumptions of a neurocomputational
model as predictions and tests them using methodologies like fMRI. This project
disables the common worry that neurocompuational models aren’t scientifically
valid because they cannot be tested or falsified. But the advantages of a
combined neurocomputa-tional—neuroimaging methodology run in the other direction
as well. fMRI has yet to transcend a “neo-phrenology” status among some neuroscientists.
Partly this attitude is fed by the “fishing expedition” strategy employed
by a lot of fMRI research. Our combined methodology generates specific predictions
about regions of interest for fMRI and greater focus for the design of experimental
tasks and controls. And when combined with neurocomputational models based
on known cell-physiological processes, fMRI becomes directly applicable to
bridging the current gap between cognitive and cellular/molecular neuroscience.
Second, the neurocomputational operation our model proposes
as implemented in the frontal saccade command regions (vector subtraction)
generalizes to other frontal regions. Anatomically, primate FEFs consist
of “standard frontal cortex with distinct granular layer”: the same type
that constitutes most of frontal cortex (Parent 1996). Hence most frontal
regions have the cell types, distributions, and anatomical connections capable
of implementing the same computational operation as FEFs to compute ordered
sequences of neuronal activation patterns. Frontal regions subserve many
higher cognitive processes (e.g., language production, planning, problem
solving, complex motor commands); and virtually every higher cognitive (and
con-scious) process is “sequential” in a fashion similar to saccade sequences
(Bickle, Worley, and Bernstein 2000). By combining biologically plausible
neurocomputational modeling with functional neuroimaging to investigate the
cellular mechanisms of sequential saccade commands, we are at the same time
postulating potential, biologically plausible cellular mechanisms for many
types of sequential cognitive processes, implemented by cytoarchitecturally
similar frontal regions of the primate brain.
Bernstein, M., Stiehl, S., and Bickle, J. (2000). “The
effects of motivation on the stream of consciousness: Generalizing from a
neurocomputational model of cingulo-frontal circuits controlling saccadic
eye movements.” In R. Ellis and N. Newton (eds.), The Caldron of Con-sciousness:
Motivation, Affect, and Self-Organization. Amsterdam: John Benjamins,
135-161.
Bickle, J., Holland, S., Schmithorst, V., and Avison,
M. (2001). “Cellular mechanisms of saccade sequencing revealed using a combined
neurocompuational—fMRI approach.” Proceed-ings of the First World Congress
on Neuroinformatics (Technical University Vienna), 571-588.
Bickle. J., Worley, C., and Bernstein, M. (2000). “Vector
subtraction implemented neurally: A neurocomputational model of some cognitive
and conscious processes. Conscious-ness and Cognition 9: 117-144.
Parent, A. (1996). Carpenter’s Human Neuroanatomy , 9th Ed. Baltimore: Williams and Williams.