Figure 1.
The DCMs in this paper were used to analyse fMRI data from three brain regions: (i) left posterior temporal sulcus (region P), (ii) left anterior superior temporal sulcus (region A) and (iii) pars orbitalis of the inferior frontal gyrus (region F). The DCMs themselves comprised the following variables; experimental inputs for auditory stimulation and
for speech intelligibility, a neuronal activity vector
with three elements (one for each region P, A, and F), exogenous connections specified by the three-by-three connectivity matrix
(dotted arrows in figure), modulatory connections specified by three-by-three modulatory matrics
for inputs
(the solid line ending with a filled circle denotes the single non-zero entry for this particular model), and a 3-by-2 direct input connectivity matrix
with non-zero entries shown by solid arrows. The dynamics of this model are govenered by equation 1. All DCMs in this paper used all-to-all endogenous connectivity ie. there were endogenous connections between all three regions. Different models were set up by specifying which regions received direct (auditory) input (non-zero entries in
) and which connections could be modulated by the speech intelligibility (non-zero entries in the matrix
).
Figure 2.
RFX posterior densities for input families.
The histograms show versus
for the
input families. Input family ‘P’ has the highest posterior expected probability
. See Table 1 for other posterior expectations.
Table 1.
Inference over input families.
Figure 3.
RFX Posterior densities for modulatory families.
The histograms show versus
for the
modulatory families. Modulatory family ‘F’ has the highest posterior expected probability
. See Table 2 for other posterior expectations.
Table 2.
Inference over modulatory families.
Figure 4.
Model level inference for input families.
For FFX (top panel) the figure shows that models in the P family have by far the greatest posterior probability mass. For RFX (bottom panel) models in both A and P families have high posterior expected probability, although the probability mass for P dominates.
Figure 5.
Model level inference for modulatory families.
For FFX (top panel) the figure shows that models in the F and BAL families have most probability mass. The expected posteriors from the RFX inference show a similar pattern (bottom panel). The ordering of models in this figure is not the same as the ordering of P models in figure 4.
Figure 6.
The figure shows the input (filled square and solid arrow) and modulatory connectivity (solid arrows) stuctures for four models in Occam's window (assessed using FFX). Note that all models also have full endogenous connectivity (not shown). These four models are (a) model with
, rank = 1, (b) model
with
, rank = 2, (c) model
with
, rank = 15 and (d) model
with
, rank = 16. All models have auditory input entering region P.
Figure 7.
Average Modulatory Connections from FFX for input family P.
The figures show the posterior densities of average network parameters from fixed effects Bayesian model averaging for the modulatory connections. Only forward connections from P to A and from P to F are modulated by speech intelligibility.
Figure 8.
Average Modulatory Connections from RFX for input family P.
The figures show the posterior densities of average network parameters from random effects Bayesian model averaging for the modulatory connections. Only forward connections from P to A and from P to F are modulated by speech intelligibility.