Neural excursions from manifold structure explain patterns of learning during human sensorimotor adaptation

Abstract

Humans vary greatly in their motor learning abilities, yet little is known about the neural mechanisms that underlie this variability. Recent neuroimaging and electrophysiological studies demonstrate that large-scale neural dynamics inhabit a low-dimensional subspace or manifold, and that learning is constrained by this intrinsic manifold architecture. Here, we asked, using functional MRI, whether subject-level differences in neural excursion from manifold structure can explain differences in learning across participants. We had subjects perform a sensorimotor adaptation task in the MRI scanner on 2 consecutive days, allowing us to assess their learning performance across days, as well as continuously measure brain activity. We find that the overall neural excursion from manifold activity in both cognitive and sensorimotor brain networks is associated with differences in subjects’ patterns of learning and relearning across days. These findings suggest that off-manifold activity provides an index of the relative engagement of different neural systems during learning, and that subject differences in patterns of learning and relearning are related to reconfiguration processes occurring in cognitive and sensorimotor networks.

Separate and shared low-dimensional neural architectures for error-based and reinforcement motor learning

“…Specifically, we found that the cerebellum was overwhelmingly EL-aligned, consistent with its established role in implicit learning through sensory prediction errors (Wolpert et al, 2011; Taylor and Ivry, 2014). Our observation that large segments of the dorsal-attention network were EL aligned is likewise consistent with our previous findings that the engagement of the dorsal-attention network is associated with better performance during visuomotor adaptation learning (Areshenkoff et al, 2022). By contrast, we found that the putamen and accumbens were RL-aligned, consistent with their role in reward-based learning (Knowlton et al, 1996).…”

Distinct patterns of cortical manifold expansion and contraction underlie human sensorimotor adaptation

“…To reduce the influence of large individual differences in functional connectivity that can obscure any task-related changes (Fig. 1D; see also [55]), all connectivity matrices were centered according to a Riemmanian manifold approach (see Materials and Methods; [56][57][58]). To demonstrate the effect of this centering procedureand its importance for elucidating learning-related effects in the data-we projected participants' individual covariance matrices, both before and after centring, using uniform manifold approximation (UMAP; [59]).…”

“…We centered the connectivity matrices according to a previously described procedure that leverages the natural geometry of the space of the covariance matrices [56][57][58]. First, a grand mean covariance matrix, Sgm , was computed by taking the geometric mean covariance matrix across all participants and epochs.…”

Reconfigurations of cortical manifold structure during reward-based motor learning

“…6A highlights the learning curves for two example subjects: An individual who learned the hidden shape quite rapidly (a ‘fast learner’ in green) and a second individual who only gradually learned to trace the hidden shape (a ‘slow learner’ in red). To quantify this variation in subject performance, we opted for a pure data-driven approach and performed functional principal component analysis (fPCA; 75 ) on subjects’ learning curves, which allowed us to isolate the dominant patterns of subject variability see Methods for further details; see also 63 . Using this fPCA approach, we found that a single component — encoding overall learning — captured the majority (∼75%) of the variability in subjects’ learning curves (Fig.…”

“…For the benefits (and general necessity) of this centering approach, see Fig. 2, and for an additional overview, see 63 .…”