Item

Abstract

Functional Magnetic Resonance Imaging (fMRI) research in sensorimotor learning focus on
two separate paradigms: (1) task-based (tfMRI), where brain changes are evaluated ac-
cording to activity elicited by performance of the task, or (2) task-free, i.e., resting-state
(rsfMRI), where changes are reflected in spontaneous, internally generated brain activity.
While the former paradigm allows careful control and manipulation of the task, the later
allows unrestrained motor learning tasks to take place beyond the limitations of the scanner
environment. Machine learning approaches attempting to model these two types of measure-
ments together to explain physiological effects of learning remained unexplored. Although
these paradigms yield results showing considerable overlap between their topographical pat-
terns, they are usually treated separately. Consequently, their relationship, and how or if
any behaviorally relevant neural information processing mediates it, remains unclear. To
resolve this ambiguity, new methodology was developed guided by questions of sensorimotor
learning in motor tasks having dynamics completely specified mathematically.
First, basic fMRI methodological considerations were made. Machine learning methods that
claimed to predict individual tfMRI task maps from rsfMRI activity were improved. In
reviewing previous methodology, most methods were found to underperform against trivial
baseline model performances based on massive group averaging. New methods were devel-
oped that remedies this problem to a great extent. Benchmark comparisons and model
evaluation metrics demonstrating empirical properties related to this predictive mapping
previously unconsidered were also further developed. With these newly formed empirical ob-
servations, a relationship between individual prediction scores and behavioral performance
measured during the task could be established.
Second, a complex motor learning task performed during an fMRI measurement was designed
to relate learning effects observed in both types of measurements from a single longitudinal
learning session. Participants measured while performing the task show they learn to exploit
a property that drives brain activity in certain regions towards a state requiring less active control and error correction. Reconfiguration of functional activity in task-evoked and task-
free activity from these behavioral learning effects were investigated, applying methodology
developed earlier in an attempt to relate them together. Predictions of individual task-
evoked responses from rsfMRI provide a relative measure of dependence, however, remain
limited for reasons understood from the methodological study. No rsfMRI reconfiguration
due to learning was detected, yet changes over the course of learning in task-evoked activity
appear significant. Increasing recruitment of the Default Mode Network (DMN) during the
task explain these changes. These results support that minimal reconfiguration of the cortex
suggestive of plasticity effects are needed to find task solutions in a passively stable space.