Figure 4

Activity is explained by a small number of features that separate vergence and rotation variables

(A) Explained variance (EV; cross-validated) for the ROIs of an example fish using RRR.

(B) The cross-validated explained variance of the best RRR model (EV_RRR) was generally higher than the explained variance of the best MLR model (EV_MLR) (left). Right shows performance on training data, i.e., without cross-validation. Same fish as in (A).

(C) Population-explained variance (popEV, cross-validated; see STAR Methods) as a function of the number of features (rank) included in the model for each individual fish (gray traces; example fish in A and B is shown in red) and across fish (average of 8 fish, black trace; error bars correspond to SEM).

(D) The first two features for the example fish. Feature traces are scaled by their overall importance (V^∗S; see STAR Methods). For each feature, variables are color-coded and sorted into vergence, rotation, and tail variables. Vertical dotted line indicates a time shift of zero. Traces of five cross-validation runs are shown and closely overlap. Note that feature 1 groups activity related to vergence and swimming, whereas feature 2 groups activity related to rotation. Features 3–6 of this fish have only a minor contribution to the popEV (see Figure S5).

(E) Contribution of the latent regressors associated with features 1 (vergence) and 2 (rotation) to ROI activity for the same example fish (43,622 ROIs).

See also Figures S4 and S5.