Fig. 4: The structure in neural space of multi-movement networks resulting from de novo learning is responsible for their patterns in adaptation.

Networks were given either angular (a) or categorical (d) inputs to simulate different learning experiences. After training on different repertoires, they had to adapt to a 10° VR perturbation. b Latent activity, for example, network with angular inputs trained on four movements during preparation (500 ms before go cue) and execution (1000 ms after go cue). Each trace corresponds to the trial-averaged activity for each movement projected on the neural manifold computed before adaptation. Solid lines, activity before adaptation; dotted lines, activity after adaptation. c Left: Input structure, measured as the cosine dissimilarity between input vectors for pairs of movements for the network in Panel b. Right: Neural structure, measured as the normalized median Euclidean distances between latent trajectories during preparation and execution for different movements for the same network. e, f. Same as panels (b, c) but for a categorical input network. Inset in Panel e: zoomed view. g Congruency between the input and neural structure, quantified as the Pearson’s correlation between their dissimilarity matrices. Congruency for mismatched input-activity pairings shown as control. Circle and error bars, mean of congruency values for each seed (n = 10) and 95% confidence intervals (CIs) with bootstrapping. h Motor output following skill-learning for angular input networks. Bottom: Loss during adaptation training. Traces, shaded surfaces, smoothed mean, and 95% CIs across networks of different seeds (n = 10). i Same as panel h but for categorical input networks. j Decay constants for exponential curves fitted to the loss curves in panels (h, i). Circles, error bars, means, and 95% CIs with bootstrapping. k Relative weight changes during adaptation. Circles and error bars, means of the median changes across all weights for each seed, and 95% CIs with bootstrapping. l Schematic for a ‘deviation angle’ between the ‘adjacent movement vector’ (red solid line in panel b, e) and the ‘adaptation vector’ (red dotted line in panel b, e). m Deviation angles during adaptation. Circles and error bars, means of the median deviation angles across all time steps for each seed and 95% CIs with bootstrapping.