Tone-in-noise detection behavior.

a. Schematic of go/no-go tone-in-noise detection task. Licking responses to target tones were rewarded, while responses to narrowband noise distractors were penalized with a timeout. Target tone frequency was fixed during a single behavior session and masked by narrowband (0.3 octave) noise centered at the same frequency with variable signal-to-noise (SNR). The “Catch” distractor was identical to the masking noise but with no tone. b. Behavioral performance of individual animals as a function of SNR (d-prime = Z[target response rate] - Z[catch response rate], n = 4 animals). Black line and error bars indicate the mean and standard error of the mean across animals. c. Left: Stimulus set for an example experiment where the target tone frequency was 2828 Hz. Right: both task relevant (catch vs. target) and task irrelevant (target vs. target, distractor vs. distractor) sound discriminations were studied.

State-dependent modulation of singe neuron target vs. catch discrimination.

a. Example peristimulus time histogram (PSTH) responses from a single recording site in A1. Heatmap color in each row indicates PSTH amplitude of one neuron. Dashed lines indicate sound onset / offset. Spikes were binned (20 ms), z-scored, and smoothed (σ = 30 ms Gaussian kernel). Example target responses are to the pure tone (Inf dB) target. Difference is computed as the response to target minus catch response. b-c. Mean z-scored response evoked by catch vs. Inf dB stimulus for each A1 neuron across passive (b) and active (c) trials. d. Histogram plots state-dependent change in target vs. catch stimulus discriminability for each A1 neuron. Neural d-prime is defined |Z[target] - Z[catch]|, and Δd-prime is the difference of active minus passive d-prime. e. Histogram of Δd-prime for dPEG neurons, plotted as in D.

Selective enhancement of task-relevant category representation in secondary auditory cortex.

a. Left: Representative A1 population activity during passive listening projected into a 2-dimensional space optimized for discriminating target versus catch responses. Each dot indicates the population response on a single trial, color indicates different noise (catch) or tone-in-noise (target) stimuli, and ellipses describe the standard deviation of responses across trials. The degree of ellipse overlap provides a visualization of the neural discriminability (d-prime) between the corresponding stimuli. Right: A1 population activity during active behavior. b. Mean population d-prime between sounds from each category (target vs. catch, target vs. target, and distractor vs. distractor, Figure 1C) for each A1 recording site (n = 18 sessions, n = 3 animals). c. Δd-prime is the difference between active and passive d-prime, normalized by their sum (D vs. D / T vs. T p = 0.048, Wilcoxon signed rank test). d. Single-trial population responses for a single site in non-primary auditory cortex (dPEG), plotted as in A. e. Passive vs. Active category discriminability for dPEG recording sites, plotted as in B (n = 13 sessions, n = 4 animals). f. Changes in discriminability per category in dPEG. Δd-prime for target vs. catch pairs (T vs. C) was significantly greater than for the other categories (D vs. D: p = 0.003; T vs. T: p = 0.005, Wilcoxon signed rank test).

Changes in neural decoding are correlated with behavior performance in dPEG, but not A1.

a. Scatter plot compares neural Δd-prime (active minus passive) for all tone-in-noise target vs. catch noise combinations against the corresponding behavioral d-prime for that target vs. catch discrimination. Line shows the best linear fit, and shading represents bootstrapped 95% confidence interval for slope. Left, data from A1 (n = 60 unique target vs. catch combinations, n = 3 animals, 18 recording sessions). Right, data from dPEG (n = 44 unique target vs. catch combinations, n = 4 animals, 13 recording sessions). b. Pearson correlation between neural d-prime and behavioral d-prime in each brain region. Error bars indicate bootstrapped 95% confidence intervals (A1: p = 0.082; dPEG: p = 0.002, bootstrap test).

Task-related changes in shared population covariability do not impact coding of task-relevant features.

a. Schematic of population response over many trials to a catch stimulus (grey) and target stimulus (red), projected into a low-dimensional space. Dashed line indicates the sensory discrimination axis and grey line indicates the axis of shared variability across trials during passive listening. Black lines indicate possible rotations in the axis of shared variability either toward or away from the discrimination axis during the task-engaged state. A larger angle (8) between the shared variability and the discrimination axes leads to increased discrimination accuracy. b. Alignment (cosine similarity) between the discrimination and shared variability axes during passive and active conditions. Error bars represent standard error of the mean. The axes become more aligned during task engagement in A1 (p < 0.001, Wilcoxon signed-rank test) and do not change in dPEG. c. Mean selective enhancement of neural target vs. catch discriminability across recording sites for simulated and actual data. Selective enhancement is the difference in Δd-prime for target vs. catch and target vs. target (inset). Simulations sequentially introduced task-dependent changes in mean sound evoked response gain, single neuron variance, and population covariance matching changes in the actual neural data. d. Model performance, defined as the correlation coefficient between simulated and actual selective enhancement. Performance of each model was evaluated against the performance of the shared variance model to check for stepwise improvements in predictions. Stars indicate significance at alpha = 0.05 level, bootstrap test. Colors indicate brain regions: dPEG / black, A1 / grey.