Figure 4. A condition-invariant signal during speech initiation.

(A) A large component of neural population activity during speech initiation is a condition-invariant (CI) neural state change. Firing rates from 200 ms before to 400 ms after the go cue (participant T5) and 100 ms to 700 ms after the go cue (T8) were decomposed into dPCA components like in Kaufman et al. (2016). Each bar shows the relative variance captured by each dPCA component, which consists of both CI variance (red) and condition-dependent (CD) variance (blue). These eight dPCs captured 45.1% (T5-words) and 8.4% (T8-words) of the overall neural variance, which includes non-task related variability (‘noise’). (B) Neural population activity during speech initiation was projected onto the first dPC dimension; this ‘CIS₁’ is the first component from panel A. Traces show the trial-averaged CIS₁ activity when speaking different words, denoted by the same colors as in Figure 3B.

Figure 4—figure supplement 1. Further details of neural population dynamics analyses and additional datasets.

(A) Cumulative trial-averaged firing rate variance explained during each dataset’s speech initiation epoch as a function of the number of PCA dimensions (top set of curves) and dPCA dimensions (bottom set of curves) used to reduce the dimensionality of the data. The dotted lines mark the eight dimensions used for dPCA in the condition-invariant signal analyses for panels C-E and in Figure 4. In addition to the T5-words and T8-words datasets (blue curves) shown in Figure 4, this figure also shows results from the T5-syllables and T8-syllables datasets (orange curves) as well as the T5-5words-A and T5-5words-B replication datasets (gold and purple curves). (B) Cumulative trial-averaged firing rate variance explained during all six datasets’ speech generation epochs. The dotted lines mark the six dimensions used for jPCA in the rotatory dynamics analyses for panels E-H and Figure 5. PCA and jPCA curves are one and the same because jPCA operates within the neural subspace found using PCA. (C) Firing rates during the speech initiation epoch of each dataset were decomposed into dPCA components like in Figure 4A. The additional inset matrices quantify the relationships between dPC components as in Kobak et al. (2016): each element of the upper triangle shows the dot product between the corresponding row’s and column’s demixed principal axes (i.e. the dPCA encoder dimensions), while the bottom triangle shows the correlations between the corresponding row’s and column’s demixed principal components (i.e. the neural data projected onto this row’s and column’s demixed principal decoder axes). More red or blue upper triangle elements denote that this pair of neural dimensions are less orthogonal, while more red or blue lower triangle elements denote that this pair of components (summarized population firing rates) have a similar time course. Stars denote pairs of dimensions that are significantly non-orthogonal at p<0.01 (Kendall correlation). (D) Neural population activity projected onto the first dPC dimension, shown as in Figure 4B for all six datasets. (E) The subspace angle between the CIS₁ dimension from panels C,D and the top jPC plane from panel F, for each dataset. Across all six datasets, the CIS₁ was not significantly non-orthogonal to any of the six jPC dimensions (p>0.01, Kendall correlation) except for the angle between CIS₁ and jPC₅ in the T5-words dataset (70.5°, p<0.01). (F) Neural population activity for each speaking condition was projected into the top jPCA plane as in Figure 5A, for all six datasets. (G) The same statistical significance testing from Figure 5B was applied to all datasets to measure how well a rotatory dynamical system fit ensemble neural activity. (H) Rotatory neural population dynamics were not observed when the same jPCA analysis was applied on each dataset’s neural activity from 100 to 350 ms after the audio prompt. Rotatory dynamics goodness of fits are presented as in the previous panel.

Figure 4—figure supplement 2. Neural population dynamics when viewed across a range of reduced dimensionalities.

(A) Condition-invariant neural population dynamics during speech initiation are summarized as in Figure 4 for the T5-words (top) and T8-words (bottom) datasets, but now varying the number of dPCA components used for dimensionality reduction. The eight dPCs results shown in Figure 4 and Figure 4—figure supplement 1 are outlined with a dashed gray box. (B) Rotatory neural population dynamics during speaking are summarized as in Figure 5 for the T5-words (top) and T8-words (bottom) datasets, but now varying the number of principal components used for the initial dimensionality reduction (after which the jPCA algorithm looks for rotatory planes). The six PCs results shown in Figure 5 and Figure 4—figure supplement 1 are outlined.