Extended Data Fig. 7: Analysis of neural recordings.
a, For each mouse, mean ± sem (n = 10 folds) R2 for a Weiner cascade ridge regression model73 fit to muscle activity without (top) or with (bottom) a cubic nonlinearity using the top principal components (PCs) capturing at least 99% of variance in neuronal firing rates. Retaining 99% of variance resulted in between 52 and 203 PCs being included across sessions (median 92 PCs). Muscle activity averaged in 10 ms bins was fit using the neural PCs over the preceding 100 ms (10 bins). The middle 50% of each recording was used to reduce compute time. This segment was divided into 1-second epochs, and a random 10% of epochs were held out for performance measurement. The first 100 ms of these epochs were omitted from testing to avoid overlap with training data. The mean ± sem R2 values across sessions and muscles were 0.310 ± 0.051 (ridge) and 0.546 ± 0.051 (with cubic). b, Neural activity maps for 48 neurons from one mouse. To illustrate varying sparsity, neurons shown have values equally spaced along the full distribution of sparsity values for the given mouse. Black text is sparsity, red text is each map’s maximum firing rate.
