Fig. 3: Cortical bulbar feedback represents stimulus identity, contingency, and behavioral outcome.

a Histogram of individual bouton response correlation values across trials as a function of behavioral contingency (Hits, H vs. false alarms, FA vs. correct rejections, CR) in Odor (Left) and Sound (Right) trials. Bouton responses were sampled between cue onset and end of the delay period (before licking, “Methods” section). Inset: Bouton response stability across conditions (H/H, H/CR, H/FA, CR/FA) reported using as reference the 90th percentile of the Hit/Hit bouton response correlation distribution (bootstrap analysis, “Methods” section). b Individual bouton response stability analysis for trials where mice subsequently licked the reward spout (hits and false alarms). Note the differences in trial-to-trial response correlation distributions when comparing Odor H/H vs. Odor H/FA vs. Odor FA/Sound FA trials. a and b: N = 20 sessions, 3 mice. Two-sided One-way ANOVA and multiple comparisons of means: * = p < 0.05 compared to ‘odor H/H’ bouton stability. c Principal component analysis (PCA) for one example session: feedback bouton ensemble response trajectories plotted in a space defined by the first three principal components (74.5 and 73.7% variance explained respectively for odor and sound trials); population response trajectories rapidly diverge as a function of trial contingency for both Odor (Left) and Sound (Right) trials. Miss (M) trials were excluded from the analysis due to their low frequency (<3%). Different task periods in each trajectory are represented by distinct traces (baseline: thin continuous; cue: thick continuous; delay: thick interrupted; report: thick dotted line). d Multi-layer perceptron classifiers were trained to decode stimulus identity (odor or sound), behavioral contingency (H, FA, CR), trial instruction (Go or No-Go), and behavior (lick or no-lick) in the delay version of the task. Top: Average classifier performance across all sessions normalized relative to baseline performance. When shuffling trial labels on the training data, the average classifier performance was 0. Bottom: Distribution of the number of licks per second across all sessions. See Source Data and Supplementary Table 1 for exact data points and p-values.