De novo motor learning creates structure in neural activity that shapes adaptation

Animals can quickly adapt learned movements to external perturbations, and their existing motor repertoire likely influences their ease of adaptation. Long-term learning causes lasting changes in neural connectivity, which shapes the activity patterns that can be produced during adaptation. Here, we examined how a neural population’s existing activity patterns, acquired through de novo learning, affect subsequent adaptation by modeling motor cortical neural population dynamics with recurrent neural networks. We trained networks on different motor repertoires comprising varying numbers of movements, which they acquired following various learning experiences. Networks with multiple movements had more constrained and robust dynamics, which were associated with more defined neural ‘structure’—organization in the available population activity patterns. This structure facilitated adaptation, but only when the changes imposed by the perturbation were congruent with the organization of the inputs and the structure in neural activity acquired during de novo learning. These results highlight trade-offs in skill acquisition and demonstrate how different learning experiences can shape the geometrical properties of neural population activity and subsequent adaptation.


Trained on synthetic hand trajectories
Trained on actual hand trajectories Supplementary Figure S3: Trends in variance remain robust for synthetic movements and di↵erent measures of variance.a-h.Networks were trained to produce motor output (b) based on simulated (a) rather than actual "hand trajectories" based on monkey movements.Simulated trajectories did not have trial-to-trial variability for a given movement.
c-e.Same as Fig. 2c-d but for networks trained on simulated hand trajectories.f-g.Same as Panels c-e but for total variance summed across all features (across all units for unit activity, latent dimensions for latent activity, and output dimensions for position) rather than the variance for each individual feature.Variance is still calculated for each timestep across trials.Note that the patterns remained the same for synthetic trajectories as when trained on actual hand trajectories, showing that constrained dynamics are not a byproduct of variability in the monkey movements they were trained on.i-m.Networks were trained to produce motor output (i) based on actual "hand trajectories" based on monkey movements (j).These are the same simulations as in the main text.k-m.Same as Panels f-g but for networks trained on experimental trajectories.Note that the trends remain the same.While some trends are not as strong, they are likely due to masking from the variability inherent in experimental trajectories, since the trends remain strong for the networks trained on synthetic trajectories.e Supplementary Figure S5: Networks can generalize and adapt to perturbations that require movements within a learned range.a. Networks were trained on repertoires with movements that spanned di↵erent ranges (top).The range -10°to -50°, denoted by an red asterisk, was used for all simulations in the main text.b-c.Networks trained on two-movement repertoires in their respective ranges (i.e. a movement to -10°and a movement to -50°for the -10°to -50°range) were tested on target cues for movements equally spaced between 30°and -90°t o assess whether they could generalize to movements they were not trained on.The target cues were chosen such that networks were tested on movements they knew ('Trained'), movements that were within the range of known movements ('Within'), and movements that were outside the range ('Outside').d.Projections of the experimental (colored) and simulated (grey) latent dynamics onto the first three axes identified through canonical correlation analysis.e-f.RNNs were trained on the center-out reaching task with either one-hot encoded ('categorical') or angular inputs, and compared to monkeys trained on the same task (Monkey C, Monkey C2, and Monkey M).Data from Ref. 6. e.The 'deviation angle' (Figure 4k) was calculated and pooled for the latent activity of all targets before and after adaptation to a visuomotor rotation.Circles and error bars, median and 95% confidence intervals with bootstrapping.Grey, control with shu✏ed targets and time points.f.Congruency between angular input structure and neural space structure (see Methods) compared to the decay constants fit to learning curves based on the angular reach errors.Output for reassociation perturbation for angular input networks.Decay constants for angular and categorical input networks.Presented in Figure 5c.g.Output for reassociation perturbation for angular input networks.Decay constants for angular and categorical input networks.Presented in Figure 5f.
: Networks that can generate multiple movements produce more constrained neural dynamics.This figure presents additional data for Figure2b-d, and compares di↵erent variables across networks with di↵erent repertoires producing the same common movement.a. Unit activity for one example unit for networks trained on one (yellow) or two (orange) movements.Networks had the same random seed.Traces, di↵erent trials.b.Variance in unit activity, calculated per timestep across all trials for example unit in Panel a. c.Distributions for variance in unit activity, pooled across all units for one example seed.Dashed lines, median.d-f.Same as Panels a-c but for the second dimension of the latent dynamics in Panels d-e and for all dimensions in Panel f. g-i.Same as Panels a-c but for the motor output.In Panels g-h: Solid line, position along the x axis; dotted line, position along the y axis.: Networks reproduce experimental results from Ref. 54. a.This study examined the variance in reach angles to a given target following repeated movements sampled from a normal distribution around the target with di↵erent standard deviations.Larger variance of known movements was correlated with larger variance in reach angle.Figure modified from Fig. 2a from Ref. 54. b.Networks initially trained on di↵erent repertoires were tested on one shared movement.Networks trained on larger repertoires had a larger variance of known movements, and this was also correlated with larger variance in reach angle.
: Addition of noise shows underlying structure in multimovement networks.Noise was applied to networks trained on synthetic trajectories without trial-to-trial variability to guarantee that di↵erences would be due to di↵erent repertoires rather than di↵erent variability in the training data.a. Latent trajectories for an example singlemovement network.Purple line, trial-averaged trajectory without increased noise added.Pink gradient and grey lines, trajectories for example trials when increased noise (⌘ = 1) is added.Color gradient, time-course of the trial.b.Same as Panel a but for an example two-movement network.Noise is only added to the first movement.c. 90th-percentile tangling in the neural space 86 , which quantifies how deterministic future states are from current states.d-f.Same as Panels a-c but for the motor output and output space.g.Mean-squared error (MSE) of output when noise of increasing magnitude is added to the neural activity.Line and shaded area, median and 95% confidence interval.Left: MSE for both x and y directions.Middle: MSE only for the x direction.Right: MSE only for the y direction.h.Same as Panel g but for the variance in position.
b. Motor output for the target cues for each movement for a sample network.Colors, target cues; circles, targets for each movement.Colors for target cues that the networks have not been trained on have lower opacity.Grey backgrounds denote the range of known movements.c.Mean-squared error between the network output and target positions for previously 'Trained' movements, and movements 'Within' or 'Outside' the range of known movements.Note that MSE was lower for 'Within' movements, showing that the networks can generalize.d.Loss during adaptation training with counterclockwise VR perturbations of 10°.Traces and shaded areas, mean and 95% confidence interval across networks of di↵erent seeds.e.Decay constants for exponential curves fitted to the loss curves in Panel d.Circles and error bars, mean and 95% confidence interval with bootstrapping.
Supplementary FigureS6: Adaptation training with stochastic gradient descent (SGD) is better at overcoming catastrophic forgetting than FORCE.Following skill learning, networks were trained to adapt to counterclockwise VR perturbations of 10°on one movement, using either SGD or FORCE.Motor output (a) and MSE (b) for all movements in known repertoires following adaptation training.Note that networks trained with FORCE had worse 'catastrophic forgetting' (that is, forgetting of previously trained tasks when learning new tasks) of other movements that were not perturbed during adaptation training.Loss during adaptation training with SGD (c) and FORCE (d).Traces and shaded areas, mean and 95% confidence intervals across networks of di↵erent seeds.

f
Target cue to +200 ms +200 ms to +1000 ms Target cue to +200 ms +200 ms to +1000 ms Supplementary FigureS7: Explicit preparatory epochs are not necessary for di↵erences in adaptation and structure.Networks were trained to produce movements without an explicit preparatory epoch.That is, networks started movement at the target cue rather than waiting for a go cue.Without an explicit preparatory epoch, networks still had preparatory-like activity in the first 200 ms after the target cue and execution-like activity for the next 800 ms after the target cue.a-f.Same measures as Figure4a-f, but with the distances in neural space separated into the first 200 ms after the target cue and 200 to 1000 ms after the target cue.g-k.Same measures as Figure 4g-l.Supplementary Figure S8: Simulation results reflect experimental monkey data.a. Hand trajectories produced by a monkey performing an eight-target center-out reaching task.Data from Ref. 6 (pooled from all sessions for Monkey M, see Methods).b.Simulated motor output when RNNs were trained on the same center-out reaching task with angular inputs.c Canonical correlation values between experimental and simulated latent dynamics.
: Adaptation with fixed input or recurrent weights.Loss during adaptation training with 10°counterclockwise VR perturbations with fixed input weights for networks with angular (a) and categorical (b) inputs (for 100 training trials).Traces and shaded areas, smoothed mean and 95% confidence intervals across networks of di↵erent seeds.c, d.Same as Panel a, b but for reassociation perturbations (for 300 training trials).Note that the results from Figure5a,d held when the input weights were frozen such that the networks could not simply rely on changing the input weights to counteract reassociation.e, f.Same as Panels a, b but with fixed recurrent weights (for 1000 training trials).Note that the patterns remain the same as in the main text, where networks were trained with plastic input and recurrent weights, although at longer timescales.: Adaptation to visuomotor rotation perturbation is within-manifold.Manifold overlap (see Methods) for network activity between skill learning and adaptation to a VR perturbation of 10°for networks with angular (a) and categorical (b) inputs.Line colors denote di↵erent random seeds.