Motor cortex activity across movement speeds is predicted by network-level strategies for generating muscle activity

Abstract

Learned movements can be skillfully performed at different paces. What neural strategies produce this flexibility? Can they be predicted and understood by network modeling? We trained monkeys to perform a cycling task at different speeds, and trained artificial recurrent networks to generate the empirical muscle-activity patterns. Network solutions reflected the principle that smooth well-behaved dynamics require low trajectory tangling, and yielded quantitative and qualitative predictions. To evaluate predictions, we recorded motor cortex population activity during the same task. Responses supported the hypothesis that the dominant neural signals reflect not muscle activity, but network-level strategies for generating muscle activity. Single-neuron responses were better accounted for by network activity than by muscle activity. Similarly, neural population trajectories shared their organization not with muscle trajectories, but with network solutions. Thus, cortical activity could be understood based on the need to generate muscle activity via dynamics that allow smooth, robust control over movement speed.

Flexible multitask computation in recurrent networks utilizes shared dynamical motifs

“…Smoothness of dynamics was a prevalent feature in our networks, where similar computations were implemented in nearby regions of state space, on a similar dynamical landscape (Fig.4a-c). This feature of smoothness is consistent with biological noise-robust networks 13,16,47 .…”

Thalamic control of cortical dynamics in a model of flexible motor sequencing

“…Therefore, after full optimization of the thalamocortical loop, the activities in the network are of moderate magnitudes (Figures 3E and 3F), with norms (Figure S3E) that indicate that the directions of larger activity patterns tend to not fully align with the readout weights (which themselves have unit norm). Likewise, moderately large neural activity patterns in movement-irrelevant dimensions have been observed in motor cortical dynamics (Russo et al, 2018;Saxena et al, 2021). Finally, we find that noise-robust solutions lead to smaller magnitudes of the thalamocortical loop weights compared to unoptimized networks (Figures 2C and 2F).…”

Ecologically pre-trained RNNs explain suboptimal animal decisions

“…RNNs have become a widely used tool to investigate the neural mechanisms underlying the behavior of animals in laboratory tasks (Mante et al 2013; Sussillo 2014; Sussillo et al 2015; Carnevale et al 2015; Barak 2017; J. X. Wang et al 2018; Remington et al 2018; Mastrogiuseppe and Ostojic 2018; Yang and Wang 2021; Feulner and Clopath 2021; Saxena et al 2021). The usual approach is to train the RNN directly on the task of interest and then investigate the underlying circuit mechanisms.…”