Small, correlated changes in synaptic connectivity may facilitate rapid motor learning

Abstract

Animals can rapidly adapt their movements to external perturbations. This adaptation is paralleled by changes in single neuron activity in the motor cortices. Behavioural and neural recording studies suggest that when animals learn to counteract a visuomotor perturbation, these changes originate from altered inputs to the motor cortices rather than from changes in local connectivity, as neural covariance is largely preserved during adaptation. Since measuring synaptic changes in vivo remains very challenging, we used a modular recurrent network model to compare the expected neural activity changes following learning through altered inputs (Hinput) and learning through local connectivity changes (Hlocal). Learning under Hinput produced small changes in neural activity and largely preserved the neural covariance, in good agreement with neural recordings in monkeys. Surprisingly given the presumed dependence of stable neural covariance on preserved circuit connectivity, Hlocal led to only slightly larger changes in neural activity and covariance compared to Hinput. This similarity is due to Hlocal only requiring small, correlated connectivity changes to counteract the perturbation, which provided the network with significant robustness against simulated synaptic noise. Simulations of tasks that impose increasingly larger behavioural changes revealed a growing difference between Hinput and Hlocal, which could be exploited when designing future experiments.

Cerebellar-driven cortical dynamics enable task acquisition, switching and consolidation

“…3d-f and Extended Data Fig. S7) are computed as in 48 We record the RNN time-dependent activities (post non-linearity) given 1000 input examples in multiple periods: task 1 baseline, task 2 and task 1 switching (Fig. 3a).…”

Feedback-based motor control can guide plasticity and drive rapid learning

“…Extending our work to modular, multi-region RNNs whose activity is compared to neural population recordings across the sensorimotor network (such as in Ref. 59,61,62) may shed light into the distributed implementation across brain circuits of the different computations underlying feedback control.…”

“…Here, we hypothesised that the neural circuitry for feedback control may be exploited to drive the “plasticity” that enables successful motor adaptation. To address this hypothesis, we use a recurrent neural network (RNN) model 54–62 trained not only to produce a certain output, but to control it. The key difference here is that this model should be able to flexibly correct its output in the case of unexpected external perturbations.…”

“…Here, we hypothesised that the neural circuitry for feedback control may be exploited to drive the "plasticity" that enables successful motor adaptation. To address this hypothesis, we use a recurrent neural network (RNN) model [54][55][56][57][58][59][60][61][62] trained not only to produce a certain output, but to control it.…”

Local field potentials reflect cortical population dynamics in a region-specific and frequency-dependent manner

“…The latent dynamics are constrained by the neural covariance structure, which is shaped in part by circuit connectivity ( Sadtler et al, 2014 ; Oby et al, 2019 ; Okun et al, 2015 ; Feulner and Clopath, 2021 ; Feulner et al, 2021 ). Currents through these same connections as well as other biophysical properties of the circuit are the main contributors to the generation of the LFPs ( Mitzdorf, 1985 ; Einevoll et al, 2013 ; Lindén et al, 2011 ; Buzsáki et al, 2012 ; Pesaran et al, 2018 ; Figure 1A ).…”