Extended Data Fig. 1: Experimental setup, learning metrics, and image quality assurance.

(a,b) Task performance improves over days of training. (a) The percentage of trials resulting in reward significantly increases over learning (p = 2e-31; Pearson’s correlation coefficient). Data points correspond to the mean fraction of successful trials ± SEM. **** p < 0.001. n = 53 mice. (b) The reaction time (time from cue onset to movement onset; black) as well as the time from cue to reward (red) and movement onset to reward significantly decrease over learning (p = 8e-35 for cue-to-reward; p = 7e-50 for cue-to-movement; p = 6e-4 for movement onset to reward, Pearson’s correlation coefficient). Data points correspond to mean time ± SEM. **** p < 0.001. n = 53 mice. (c) Schematic of imaging schedule. In each animal, three dendritic fields were imaged on each of the first three training days (Imaging Session 1, IS1), after which each field was imaged in 5-day intervals for two additional imaging sessions (IS2 and IS3). This yielded three imaging sessions for each field: early, middle, and late sessions. (d) Spine density does not significantly change over learning (1-way ANOVA, main effect of session: F = 0.94, d.f. = 2, p = 0.39). Individual dendrites are plotted as colored dots; black line corresponds to the median spine density ± bootstrapped 95% CI.n = 76 dendrites / 23 mice.(e) Spine event frequency is stable over learning for MRS (p = 0.46, rank-sum test), and decreases for nonMRSs (p = 1e-5, rank-sum test). Spine event frequencies are higher for MRSs in late sessions (early: p = 0.17, late: p = 7e-7). Bars represent median ± bootstrapped 95% CI. n = 898 early MRSs, 820 late MRSs, 869 early nonMRSs, 788 late nonMRS. Multiple comparisons were corrected using the FDR method. (f) Example lever position trace, showing movement periods in green shadings. (g) Example of frame-wise 2d correlation with the corresponding calculated reference image (that is, an iteratively aligned and average-projected version of the imaging field) associated with the behavioral window shown in (d). Prior to motion correction (red line), frame correlations with the reference image show with a decrease during movements. After motion correction, mean frame correlations are higher regardless of movements (blue line). (h) Example x-y pixel shifts used for motion correction of each frame in the behavioral window shown in (f,g), showing that extra correction during movements successfully compensated for movement artifacts. (i) Average projection images from the ~4 s behavioral window indicated in (g). Prior tomotion correction, images are blurry and individual structures are difficult to resolve because of misalignment across frames (top). After motion correction, images are sharp, and individual spines are visible (bottom). (j) Motion correction generates stable frames over movement. When aligned to movement onset, frame correlations with the reference image (as in (e)) decrease sharply during movements in the raw, pre-correction images (red line), but are stable during movements for the post-correction frames (blue line). Curves represent mean change in reference imaged correlation. Shaded portions correspond to SEM.