Canonical cortical architecture supports the emergence of noise-invariant auditory representations

Neurons in the auditory system must represent behaviorally relevant sounds in the presence of background noise (BN) to support noise-invariant perception and behavior. Although primary auditory cortex (ACtx) has been implicated in constructing noise-invariant representations, it remains unclear which excitatory subpopulations within ACtx carry out this transformation from noise-dependent to noise-invariant coding. To address this, we presented pure tones with and without continuous BN to head-fixed mice and used two-photon calcium imaging to record sound-evoked activity from three major excitatory subpopulations in ACtx: layer (L)2/3 in- tratelencephalic (IT) neurons, L5 IT neurons, and L5 extratelencephalic (ET) neurons. L2/3 IT neurons exhibited strong noise dependence at the level of single-neuron responses, pairwise interactions, and population representations. In contrast, deep-layer pathways showed greater noise invariance, with L5 IT neurons preserving stable representations most consistently and L5 ET neurons exhibiting more limited invariance at the population level. These findings reveal a functional division of labor in ACtx, in which superficial neurons remain noise-dependent and deep-layer broadcast pathways, particularly L5 IT, preferentially carry noise-invariant represen- tations, suggesting that excitatory subpopulations contribute differentially to the construction and propagation of noise-invariant codes. Introduction1 The auditory system continuously faces the task of accurately representing sound features, even2 if they are embedded in background noise (BN). Robust, or ideally invariant, coding of acoustic3 attributes such as frequency and intensity enables animals to detect threats and allows humans to4 understand speech in complex acoustic scenes (e.g., the cocktail-party setting). The construction5 of such a noise-invariant neural code, one that represents acoustic characteristics regardless of BN,6 poses a central challenge for auditory processing. As BN increases and the signal-to-noise ratio7 (SNR) worsens, the acoustic waveform at the ear is distorted, cochlear responses are reshaped,8 and these changes propagate through the ascending auditory pathway [1–9]. Despite extensive9 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted December 22, 2025. ; https://doi.org/10.64898/2025.12.19.695511doi: bioRxiv preprint work on BN effects across numerous auditory brain regions [2, 10–14], the precise locus and circuit10 mechanisms by which noise-invariant representations first emerge remain unresolved.11 Sounds reaching the ear are decomposed by the cochlea into frequency components that are12 relayed along the ascending auditory pathway [15, 16]. When sounds are embedded in BN, their13 spectrotemporal structure becomes distorted, producing measurable changes in neural representa-14 tions in the auditory periphery [10, 11], midbrain [12], and cortex [2, 13, 14]. BN can modify both15 single-neuron and population responses, shifting firing rates and reshaping population activity pat-16 terns [10–12], degrading spatial cues [10], suppressing sound-evoked activity [2, 17], and modulat-17 ing baseline firing [1]. Across the pathway, representations generally become more noise-invariant18 as signals ascend toward auditory cortex (ACtx)[1, 9]; responses in primary ACtx are closer to19 noise-invariant than in the periphery yet remain measurably affected by BN [4, 18], whereas non-20 primary fields often exhibit stronger invariance [4, 9, 19]. These observations identify ACtx as a21 critical locus where invariance to BN is refined, motivating the question of how cortical circuits22 support this transformation.23 ACtx is a hub for sound processing and the brain-wide broadcast of auditory information,24 roles that are particularly important in complex acoustic environments where sounds are embed-25 ded in BN [7]. Inactivation of ACtx significantly impairs pure tone detection in BN, and disrupting26 somatostatin- or parvalbumin-positive interneurons produces similar behavioral deficits [6, 20], un-27 derscoring a role for ACtx in supporting noise-invariant perception. Building on these findings,28 recent work suggests that ACtx transforms noise-dependent inputs into noise-invariant representa-29 tions through adjustments of excitatory-inhibitory balance [2, 5, 6] and cholinergic neuromodulation30 [3]. Nevertheless, the computations that implement noise invariance remain unclear, in part be-31 cause prior studies have not considered the functional diversity of excitatory subpopulations within32 the cortical microcircuit.33 Within ACtx, excitatory neurons are stratified across layers and projection classes [21–26].34 Two major subpopulations, intratelencephalic (IT) and extratelencephalic (ET) neurons, differ in35 their anatomical, genetic, and functional properties [27–29]. In ACtx, layer (L)2/3 IT neurons form36 dense intracortical networks and are positioned to perform local computations on incoming sensory37 input [24]. In contrast, L5 IT and ET neurons constitute broadcast pathways: L5 IT neurons38 project widely within the telencephalon [26, 28, 30–33], whereas L5 ET neurons target subcortical39 structures such as the inferior colliculus and thalamus [25, 26, 29]. These subpopulations make40 distinct contributions to cortical function, including learning perceptually relevant information41 [22], amplifying sensory signals after injury [32], and regulating thalamic transmission [25]. Given42 these divergent roles and projection patterns, we tested whether L2/3 neurons are more susceptible43 to BN, and whether L5 IT and L5 ET neurons preferentially transmit noise-invariant signals to44 downstream targets.45 We examined noise invariance at three analytical levels (single-neuron, pairwise, and popu-46 lation) in three major ACtx excitatory subpopulations: L2/3, L5 IT, and L5 ET. Using in vivo47 two-photon calcium imaging in mice, we recorded responses to pure tones presented with and with-48 out BN. L2/3 neurons exhibited pronounced noise-dependence at all three levels, while both both49 L5 IT and L5 ET populations exhibited robust, noise-invariant coding. Thus, while superficial50 neurons alter their representations in the presence of noise, deep-layer output pathways maintain51 stable representations, indicating that noise invariance is preferentially expressed within the cortical52 output stages of ACtx.53 2 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted December 22, 2025. ; https://doi.org/10.64898/2025.12.19.695511doi: bioRxiv preprint Figure 1: Two-photon recordings from awake, head-fixed mice. (A)Schematic illustrating excitatory subpopulations in auditory cortex (ACtx). (B) Two-photon imaging setup for recordings in awake, head-fixed mice. (C) Schematic of acoustic stimulation in the absence (No-BN) and presence (BN) of continuous background noise. (D) Example fields of view (FoVs) from each recorded subpopulation. Imaging depths were 206 µm (L2/3), 415 µm (L5 IT), and 494 µm (L5 ET). (E) Example sound-evoked calcium responses at the best frequency for one representative neuron from each FoV shown in D. Results54 To assess noise invariance in sound representations across excitatory subpopulations of ACtx, we55 performed in vivo two-photon calcium imaging to record activity from three targeted groups: L2/356 (n = 961 neurons; N = 7 mice), L5 IT ( n = 2576; N = 6) and L5 ET ( n = 566; N = 4) (Fig 1A).57 GCaMP8s expression was selectively driven in each subpopulation using genetic and viral strategies58 (see Methods). Head-fixed mice were imaged through a cranial window over right ACtx, while pure59 tones were delivered monaurally to the left ear (Fig 1B). Tones spanned 4–45 kHz in half-octave60 steps and 20–70 dB SPL in 10 dB SPL steps (8 frequencies × 6 intensities = 48 unique conditions),61 each repeated 20 times. Each mouse completed two sessions: one without BN (No-BN) and one62 with continuous BN at 50 db SPL (Fig 1C–D). Fluorescence calcium traces were extracted and63 deconvolved using standard approaches [34, 35]. All deconvolved responses were z-scored relative64 to their pre-stimulus baseline (Fig 1E).65 BN attenuates single-neuron responses in L2/366 We first asked whether single-neuron responses in each excitatory subpopulation are noise-invariant.67 To do so, we analyzed two complementary aspects of single-neuron activity: (i) mean stimulus-locked68 responses, summarized by frequency and intensity tuning curves, and (ii) trial-to-trial response dis-69 tributions, which capture variability not reflected in the mean. Tuning analyses assessed additive70 and multiplicative changes between No-BN and BN conditions, whereas distributional analyses71 quantified stimulus information and its modulation by BN.72 For each sound-responsive neuron (see Methods), we computed frequency response areas (FRAs)73 by averaging sound-evoked activity across all trials for each frequency–intensity combination (Fig 2A).74 From these FRAs, we derived frequency and intensity tuning curves for the No-BN and BN condi-75 tions by marginzalizing over the complementary stimulus dimension. To compare frequency tuning76 across noise conditions, we aligned each neuron’s tuning curve to its best frequency (BF) and av-77 3 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted December 22, 2025. ; https://doi.org/10.64898/2025.12.19.695511doi: bioRxiv preprint eraged the resultant BF-centered curves for each subpopulation. BF-centered frequency tuning78 curves showed a significant reduction in L2/3, but no change in L5 IT or ET (Fig 2B top row,79 two-way ANOVA, main effect for BN, L2/3: p = 0.0135; L5 IT: p = 0.9014; L5 ET: p = 0.8156).80 This reduction was accompanied by a significant interaction between frequency and BN in L2/3,81 which is concentrated at best frequency (interaction between frequency and BN: p = 2.5946 ×10−4;82 Tukey-Kramer post hoc test for best frequency: p = 5.1255 × 10−13). Average intensity tuning83 curves likewise decreased in L2/3 but not in L5 IT or ET (Fig 2B bottom row, two-way ANOVA,84 main effect for BN, L2/3: p = 3.1543 × 10−12, L5 IT: p = 0.8450, L5 ET: p = 0.8609). Together,85 these results indicate that, on average, L2/3 neurons tuning curves are suppressed in the presence86 of BN, while L5 IT and L5 ET curves remain unchanged.87 Although averaged tuning curves showed that L2/3 responses were altered differently from88 those in L5 IT and L5 ET, these averages do not capture how individual neurons adjusted their89 tuning curves in BN. To quantify neuron-by-neuron changes that underlie the population-level90 trends, we performed reduced major axis (RMA) regression between each neuron’s No-BN and BN91 tuning curves (Fig 2C) [32, 36]. This approach estimates how much a tuning curve shifts (ad-92 ditive/subtractive) and/or scales (multiplicative/divisive) in BN (Fig 2C). Each neuron produced93 four coefficients: a slope and intercept for its frequency tuning curve, and a slope and intercept for94 its intensity tuning curve (Fig 2D-E). Slope values greater than one indicated multiplicative scaling,95 whereas values less than one indicated divisive scaling. Similarly, intercept coefficients greater than96 zero indicated additive shifts, whereas values less than zero indicated subtractive shifts.97 The distribution of slope coefficients differed across subpopulations (Fig 2F left, Kruskal–Wallis98 test, main effect for frequency slope: p = 2.5484 × 10−6; main effect for intensity slope: p =99 5.3103 × 10−7). L2/3 neurons showed lower slope values than both L5 IT and ET populations100 for frequency and intensity tuning (Dunn– ˇSid´ ak post hoc test for frequency and intensity curves,101 respectively, L2/3-L5 IT: p = 3.1471 × 10−6 and p = 1.8054 × 10−6, L2/3-L5 ET: p = 0.0114 and102 p = 0.0021, L5 IT-L5 ET: p = 0.8289 and p = 0.9683), indicating that BN produced a stronger103 divisive scaling of responses in L2/3 than in either deep-layer subpopulation.104 Intercept coefficients for frequency tuning curves also differed across subpopulations (Fig 2F105 right, Kruskal-Wallis test, main effect for frequency intercept: p = 0.0020), with L2/3 neurons106 showing lower intercepts than L5 IT and L5 ET neurons (Dunn– ˇSid´ ak post hoc test, L2/3-L5107 IT p = 0 .0304, L2/3-L5 ET: p = 0 .0058, L5 IT-L5 ET: p = 0 .6027). However, no measurable108 differences in intensity tuning curve intercepts were recorded (Fig 2F, right, Kruskal–Wallis test,109 main effect for intensity intercept: p = 0.6599).110 To assess overall trends in tuning curve modulation across subpopulations, we computed, for111 each subpopulation, the proportion of neurons showing suppressive versus enhancing changes. De-112 viations above or below 50% indicate whether a subpopulation tends to increase or decrease its113 responses in BN. If the effect were not subpopulation-specific, these proportions would be similar114 across subpopulations.115 4 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted December 22, 2025. ; https://doi.org/10.64898/2025.12.19.695511doi: bioRxiv preprint Figure 2: ACtx broadcast subpopulations exhibit noise-invariant single-neuron re- sponses. (A) Example frequency response area (FRA) from an L2/3 neuron in the absence (No-BN, left) and presence (BN, right) of background noise. Both panels use the same color scale, with lighter colors indicating larger responses. (B) Average best-frequency-centered frequency tuning curves (top row) and average intensity tuning curves (bottom row) for each excitatory sub- population. (C) Schematic illustrating linear transformations between tuning curves across BN conditions (left) and reduced major axis (RMA) regression used to quantify these changes (right). Dotted line indicates the unity line. (D) Example frequency tuning curves for a single neuron recorded in No-BN and BN (left) and corresponding RMA regression (right). Estimated slope and intercept coefficients are shown above the plot. (E) Same as D, but for intensity tuning curves. (F) Distributions of RMA slope coefficients (left) and intercept coefficients (right) that dif- fered significantly from 1 and 0, respectively, for frequency (magenta) and intensity (cyan) tuning curves. Number of neurons with significant regressions for L2/3, L5 IT and L5 ET respectively: n = 272, 204, 88 frequency slopes, n = 222 , 140, 65 intensity slopes, n = 161, 113, 49 frequency intercepts, n = 107, 82, 43 intensity intercepts. (G) Proportions of RMA coefficients classified as additive, subtractive, multiplicative, or divisive for each subpopulation. Same n as in F. 5 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted December 22, 2025. ; https://doi.org/10.64898/2025.12.19.695511doi: bioRxiv preprint We found that the proportions of multiplicative versus divisive coefficients differed between116 subpopulations for both frequency and intensity slopes (Fig 2G left, separate chi-square tests for117 frequency and intensity slopes, respectively; p = 3.5892 ×10−6 and p = 2.971 ×10−7), as well as for118 additive versus subtractive frequency intercepts (Fig 2G right, chi-square test for frequency inter-119 cepts: p = 0.0026). In L2/3 neurons, the proportions of frequency and intensity slopes were strongly120 biased toward divisive changes (binomial test for frequency and intensity slopes: p = 4.2188×10−15121 and p = 7.6385 × 10−9). In contrast, L5 IT and L5 ET neurons showed proportions closer to 50%,122 with the only significant biases favoring enhancement of tuning curves via multiplicative or additive123 changes, specifically in the intensity slopes of L5 IT neurons (binomial test: p = 0 .0342) and in124 the frequency intercepts of L5 ET neurons (binomial test: p = 0.0094). Together, these results125 indicate that, although individual neurons in all subpopulations can either enhance or suppress126 their tuning, L2/3 neurons tended to exhibit predominantly divisive changes, whereas L5 IT and127 ET neurons maintain more balanced, and in some cases mildly enhancing, tuning shifts.128 While changes in tuning curves reveal the direction of modulation induced by BN, they do129 not capture whether a neuron’s trial-to-trial response distribution changes under BN. For exam-130 ple, two neurons may show similar suppression in mean tuning across conditions, yet differ in131 how reliably they respond to each tone. Such differences in response variability reflect how sta-132 ble a neuron’s responses remain after BN-induced changes in mean tuning. To quantify this, we133 applied an information-theoretic framework to measure how much of the variability in each neu-134 ron’s responses was explained by the auditory stimulus (Fig 3A). For each responsive neuron,135 we computed the distribution of its sound-evoked responses across trials and then quantified how136 much auditory information it encoded by computing the mutual information between responses137 and stimuli I(resp; stim) [37–40]. Mutual information approaches zero when a neuron’s activity is138 independent of sound frequency and intensity, and increases as responses become more stimulus-139 driven. To compare the influence of BN on information across conditions and subpopulations, we140 calculated an information modulation index (IMI) ranging from -1 to 1, which indicates whether141 a neuron conveyed more information about the stimuli in No-BN or BN (Fig 3A). IMIs near 1142 reflect greater encoding in BN, whereas IMIs near -1 indicate a loss of stimulus information in BN143 (Fig 3B). All subpopulations contained neurons with positive and negative IMIs, revealing diverse144 BN-induced changes in single-neuron encoding (Fig 3C, Kruskal-Wallis test, main effect for sub-145 population: p = 6.5648 × 10−12). On average, L2/3 neurons exhibited more negative IMIs than146 L5 IT and ET populations (Fig 3D; Dunn– ˇSid´ ak post hoc test; L2/3-L5 IT p = 3.5019 × 10−12;147 L2/3-L5 ET p = 0.0011), whereas the two L5 subpopulations did not differ from each other (Dunn-148 ˇSid´ ak post hoc test, L5 IT-L5 ET p = 0.1425). These findings indicate that L2/3 neurons reduce149 stimulus-related information in BN, whereas L5 IT and ET neurons maintain relatively preserved,150 more noise-invariant, response distribution. Together, these results show that BN induces stronger151 response suppression and reduced information encoding in L2/3, while L5 IT and ET neurons152 preserve balanced tuning and noise-invariant representations.153 BN reduces shared neural variability across spatial scales in IT but not ET neural154 responses155 While single-neuron analyses reveal how individual responses change with BN, they do not capture156 how activity is coordinated within the densely interconnected cortical circuit. To determine whether157 BN alters coordinated activity between neurons, we quantified pairwise functional connectivity158 using two complementary measures: noise correlations and signal correlations.159 If BN alters local circuit interactions, its effects should be reflected in the spatial organiza-160 6 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted December 22, 2025. ; https://doi.org/10.64898/2025.12.19.695511doi: bioRxiv preprint tion of pairwise noise and signal correlations. Noise correlations quantify the shared trial-to-trial161 variability between two neurons after removing stimulus-driven activity, providing an estimate of162 their functional coupling [41–48]. To test whether BN modulates this coupling, we examined how163 noise correlations varied as a function of intersomatic distance (see Methods), to test whether BN164 preferentially affects functional coupling in local subnetworks or broadly across ACtx.165 Consistent with previous studies [47, 49], the strongest noise correlations were observed at short166 intersomatic distances (Fig 4C, two-way ANOVA, main effect for intersomatic distance, L2/3: p =167 1.5101×10−51, L5 IT: p = 7.5697×10−28, L5 ET: p = 2.3240×10−11). BN reduced noise correlations168 in L2/3 and L5 IT neurons, but not in ET neurons (Fig 4C, two-way ANOVA, main effect for BN,169 L2/3: p = 0 .0089, L5 IT: p = 0 .0022, L5 ET: p = 0 .1302). In all subpopulation, this BN-170 induced reduction did not depend grossly on intersomatic distance, as indicated by non-significant171 interactions (Fig 4C, two-way ANOVA, interaction between BN and intersomatic distance, L2/3:172 p = 0.9398, L5 IT: p = 0.9768, L5 ET: 0.4861).173 However, because the BN-induced reduction in noise correlations was asymmetric across dis-174 tance, with larger effects at short distances and smaller effects at long distances (Fig 4C, right),175 we assessed whether BN differentially affected noise correlations across spatial scales by comparing176 BN-induced changes at short (≤ 200µm) and long distances ( > 200µm). L5 ET neurons showed177 a stronger BN-induced reduction in noise correlations at short distances (Fig 4D, Wilcoxon rank178 sum test, L5 ET: p = 0.0285), whereas L2/3 and L5 IT neurons showed similar reductions at short179 and long distances (Fig 4D, Wilcoxon rank sum test, L2/3: p = 0.9239, L5 IT: p = 0.8463). These180

indicate that BN broadly reduces shared variability across spatial scales in L2/3 and L5 IT181 populations, but acts more locally within L5 ET networks.182 Next, we assessed pairwise functional connectivity through signal correlations, defined as the183 correlation between the tuning curves of two neurons (Fig 4B). BN induced a decrease in signal184 correlations only in L2/3 neurons (Fig 4E left, two-way ANOVA, main effect for BN, L2/3: p =185 9.3076 × 10−6), whereas L5 IT and L5 ET neurons showed no significant change between BN186 conditions (Fig 4E center and right, two-way ANOVA, main effect for BN, L5 IT: p = 0.3666,187 Figure 3: L2/3 single neurons encode less stimulus information in BN. (A) Schematic illustrating mutual information between neural responses and the auditory stimulus, I(resp; stim), in relation to the entropy of each variable. (B) Mathematical formulation of mutual information between neural activity and stimulus identity, along with the definition of the information modu- lation index (IMI). (C) Cumulative distributions of IMI values for each excitatory subpopulation (L2/3: n = 423, L5 IT: n = 290, L5 ET: n = 163). (D) Mean IMI values corresponding to the distributions in C. Same n as in C. Error bars denote mean ± s.e.m. 7 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted December 22, 2025. ; https://doi.org/10.64898/2025.12.19.695511doi: bioRxiv preprint L5 ET: p = 0.3435). For all subpopulations, the effect of BN on signal correlations did not differ188 between short and long intersomatic distances (Fig 4F, Wilcoxon rank-sum test, L2/3: p = 0.1262,189 L5 IT: p = 0.1289, L5 ET: p = 0.4635). Taken together with our prior analyses, these results190 indicate that BN not only reduces L2/3 response amplitudes, but also makes their tuning curves191 less similar to one another, whereas L5 IT and L5 ET neurons maintain stable tuning similarity192 across BN conditions.193 Figure 4: Excitatory subpopulations exhibit BN-dependent changes in pairwise cor- relations. (A) Mean-subtracted trial-by-trial responses for an example pair of simultaneously recorded neurons (left) and the corresponding noise correlation matrix for all neuron pairs within an example field of view (FoV; right). (B) Same as A, but showing signal correlations computed from tuning curves for the same example neurons. (C) Mean noise correlations plotted as a func- tion of intersomatic distance under No-BN and BN conditions (L2/3: n = 9266, L5 IT: n = 3720, L5 ET: n = 2258 neuron pairs). (D) BN-induced change in noise correlations at short ( ≤200 µm) and long (>200 µm) intersomatic distances. (E) Same as C, but for signal correlations (same n as in C). (F) Same as D, but for signal correlations (same n as in D). Shaded regions denote mean ± s.e.m. 8 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted December 22, 2025. ; https://doi.org/10.64898/2025.12.19.695511doi: bioRxiv preprint Figure 5: L5 IT neurons exhibit noise-invariant detection and discrimination of audi- tory stimuli. (A) Schematic of the artificial neural network used for binary detection decoding. (B) Example neurometric curves for the same neural population under BN (left) and No-BN (right) conditions. (C) Paired scatter plot of cross-validated detection performance (d’) for each frequency–intensity combination under No-BN and BN conditions. n denotes the number of bi- nary classifiers. (D) Mean detection performance (d’) for each excitatory subpopulation, averaged across decoding runs shown in C. (E) Schematic of the artificial neural network used for multi- nomial discrimination decoding. (F) Example confusion matrices for the same neural population under BN (left) and No-BN (right) conditions. (G) Paired scatter plot of cross-validated discrimi- nation accuracy for each stimulus intensity under No-BN and BN conditions. n denotes the number of multinomal classifiers. (H) Mean discrimination accuracy for each excitatory subpopulation, av- eraged across decoding runs shown in G. Error bars denote mean ± s.e.m. 9 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted December 22, 2025. ; https://doi.org/10.64898/2025.12.19.695511doi: bioRxiv preprint Population-level decoding of pure tones are noise-invariant in L5 IT194 Although single-neuron activity and pairwise correlations reveal how BN modulates inputs to in-195 dividual neurons and pairs, neural representations ultimately arise from the collective output of196 entire populations [50–54]. We therefore asked two key questions regarding how BN influences197 population-level stimulus representations: 1) does BN reduce the ability of neural populations to198 detect the presence of a pure tone, and 2) does BN impair their ability to discriminate between199 different pure tone frequencies?200 To evaluate the noise invariance of population-level representations, we performed two decoding201 analyses to address both detection and discrimination of pure tones. We trained classifiers on trial-202 wise population activity to test whether each subpopulation could reliably detect the presence of a203 sound and identify its frequency across BN conditions. To capture potential nonlinear interactions204 among neurons and to maintain a consistent architecture across decoding analyses, we first trained205 artificial neural networks to classify whether a sound was present on each trial based solely on206 sound-evoked activity (see Methods) (Fig 5A). Because classification performance depends on the207 number of neurons provided to the classifier, we trained all classifiers on randomly selected subsets208 of 40 neurons. This choice balanced the need for sufficient neurons to support reliable classification209 while avoiding oversampling in smaller FoVs, ensuring robust and comparable performance across210 FoVs. Each network consisted of an input layer of 40 units, two hidden layers of 16 units, and a211 single readout unit. Hidden layers used a ReLU activation function, and the readout was passed212 through a sigmoid nonlinearity, with outputs above 0.5 classified as sound-present. We cross-213 validated all decoding analyses by separating trials into independent training and testing sets and214 trained independent classifiers for each tone frequency. This design ensured that any reduction in215 classification performance under BN reflected diminished sound detection, rather than reductions216 arising from mismatches in frequency tuning between the trials used in the training and testing sets.217 For some example FoVs, detection performance in BN was markedly reduced (Fig 5B). Comparing218 average detection performance between No-BN and BN sessions revealed that L2/3 and L5 ET219 neurons exhibited reduced detection performance under BN, whereas L5 IT neurons showed noise-220 invariant detection of pure tones (Fig 5D, paired t-test, L2/3: p = 6.8345×10−7, L5 IT: p = 0.4586,221 L5 ET: p = 0.0073).222 To answer whether ACtx neural subpopulations can reliably discriminate pure tone frequencies,223 we next trained neural networks to classify each trial into one of the eight possible frequencies224 (Fig 5E). To make discrimination decoding directly comparable to detection decoding and to allow225 for the same degree of nonlinear interaction in the hidden layers, we used an identical architecture226 except for the readout: the output layer contained eight units instead of one. Each output unit227 corresponded to a single frequency, and a softmax function was applied to normalize the outputs228 and select the decoded frequency on each trial (Fig 5F). In this multinomial task, decoding accuracy229 remained stable for L5 IT neurons but decreased significantly for L2/3 and L5 ET neurons (Fig 5H,230 paired t-test, L2/3: p = 2.0953 × 10−4, L5 IT: p = 0.1521, L5 ET: p = 0.0026). Thus, L5 IT neural231 populations preserve both detection and discrimination decoding performance across BN conditions,232 while both L2/3 and L5 ET neural populations show reduced decoding performance for both sound233 detection and frequency discrimination in BN.234 L5 IT neurons maintain neural manifold geometry235 While population-level decoding provides an estimate of how well each subpopulation encodes the236 specific pure-tone features used here, it does not address how many additional features (for example,237 10 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted December 22, 2025. ; https://doi.org/10.64898/2025.12.19.695511doi: bioRxiv preprint more frequencies or stimulus dimensions) the same population could, in principle, encode. To assess238 the impact of BN on the structure and capacity of neural representations, we examined the geometry239 of population activity across conditions. Manifold geometry analysis [55] provides a way to quantify240 Figure 6: L5 ET neurons show noise-dependent manifold geometry but noise-invariant manifold size. (A) Simulated examples of manifold objects encoded by three neurons, illustrating differences in manifold geometry. Insets show the manifold capacity ( αM), radius ( RM), and dimensionality (DM) for the highlighted light pink manifold object in both examples. (B) Manifold geometry metrics for each excitatory subpopulation, with manifold objects constructed by pooling all intensities for each frequency (L2/3: n = 961, L5 IT: n = 2576, L5 ET: n = 566 neurons). (C) Manifold geometry metrics as a function of stimulus intensity, with manifold objects constructed from individual frequency–intensity combinations. Same n as in B. (D) Example sound-evoked population activity from a L2/3 FoV projected onto the first two principal components (left), and total variation of the population activity across principal components for each BN condition (right). The dotted line indicates the number of principal components required to explain 60% of the variance for this FoV.(E) Same as D, shown for an example L5 IT FoV. (F) Difference in total variation between No-BN and BN conditions for each FoV, shown separately for each subpopulation (L2/3: n = 10, L5 IT: n = 11, L5 ET: n = 5 FoVs). Error bars denote mean ± s.e.m. 11 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted December 22, 2025. ; https://doi.org/10.64898/2025.12.19.695511doi: bioRxiv preprint the efficiency of a population code by estimating the capacity of a neural population to represent241 multiple perceptual features.242 Recent work has shown that neural population activity can be described in terms of neural243 manifolds and their effective dimensionality, which capture the dominant structure of population244 responses [55–57]. When sound-evoked responses are embedded in an n-dimensional space (one axis245 per neuron), activity trajectories typically lie on a lower-dimensional manifold rather than filling246 the entire space [50, 58]. Within this framework, responses to different stimuli (e.g., pure tone247 frequencies) form distinct manifold objects in neural space, each consisting of the set of response248 patterns associated with a given stimulus. The geometry of these manifold objects, particularly249 their size and dimensionality, constrains how efficiently a population can represent multiple stimuli250 [55]. For example, in a simulated population of three neurons, responses to three different stimuli251 might form manifold objects that resemble a line, disk, or sphere (Fig 6A, top). Lower-dimensional252 objects (e.g., lines and disks embedded in three dimensions) can be packed more efficiently in253 neural space, which promotes separability between manifold objects and increases the number254 of objects that can be encoded simultaneously. In contrast, larger or higher-dimensional objects255 quickly constrain the population representational capacity by making manifold objects less linearly256 separable [55] (Fig 6A, bottom). Thus, smaller and lower-dimensional manifolds correspond to a257 more efficient neural code, while larger and higher-dimensional manifolds reflect poorer encoding.258 Using this framework, we tested how BN affects manifold geometry across excitatory subpop-259 ulations. We quantified three established metrics of manifold structure: capacity (α M), radius260 (RM) and dimensionality (D M) which together reflect the efficiency of population-level encoding261 [55, 56, 59]. Effective encoding is characterized by high αM, low RM, and low DM, whereas reduced262 efficiency is marked by lower capacity and larger, higher dimensional manifolds (Fig 6A). In this263 simplified example, the neural manifold is composed of three neurons, which sets the maximum264 possible manifold dimensionality DM to 3.265 We first constructed manifold objects by pooling responses across all intensities for each fre-266 quency under No-BN and BN conditions, yielding eight objects per condition. Because manifold267 metrics depend on the dimensionality of the neural population, we analyzed randomly selected268 subsets of 40 simultaneously recorded neurons and resampled across subsets to approximate trends269 in the full population, which sets the maximum manifold dimensionality DM to 40. L2/3 and270 L5 ET neurons showed significant BN-dependent changes in all three metrics (Fig 6B, Wilcoxon271 sign-rank test for αM, RM and DM, respectively, L2/3: p = 4 .2940 × 10−7, p = 4 .3582 × 10−8,272 p = 3.3680 × 10−9, L5 ET: p = 6.8053 × 10−10, p = 6.462 × 10−10, p = 2.6811 × 10−10). In contrast,273 L5 IT neurons showed no significant BN-related differences in manifold capacity, radius, or dimen-274 sionality (Fig 6B, Wilcoxon sign-rank test for αM, RM and DM, L5 IT: p = 0.1148, p = 0.1488,275 p = 0.1363).276 To examine the effects of BN on manifold geometry at a finer stimulus scale, we repeated277 this analysis using one manifold per frequency–intensity combination, yielding 48 manifold objects278 per BN condition. This allowed us to assess how BN affected the geometry of representations for279 specific acoustic stimuli. L2/3 and L5 ET manifolds shifted toward less efficient geometry (Fig 6,280 two way ANOVA, main effect of BN for αM, RM and DM, respectively, L2/3: p = 7.5707 × 105,281 p = 4.2740 × 10−6, p = 4.5530 × 10−7, L5 ET: p = 6.7232 × 10−5, p = 2.1831 × 10−7, p =282 7.7084 × 10−8). In contrast, L5 IT manifolds showed noise-invariant capacity and dimensionality,283 with BN selectively increasing manifold radius (two way ANOVA, main effect of BN for αM, RM284 and DM, respectively, L5 IT: p = 0.1097, p = 0.0329, p = 0.0546). These results demonstrate that285 the manifold geometry of L2/3 and L5 ET populations is particularly susceptible to BN, while L5286 12 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted December 22, 2025. ; https://doi.org/10.64898/2025.12.19.695511doi: bioRxiv preprint IT neurons preserve a more noise-invariant manifold structure.287 Although manifold geometry analysis characterizes the structure of individual manifold objects,288 it does not capture the global spread of all objects within the neural representational space. To289 assess whether BN alters this global dispersion of sound-evoked population activity, we first visual-290 ized the data using principal component analysis (PCA; see Methods). In some FoVs, the first two291 principal components (PCs) revealed a clear contraction of the span of stimulus representations292 under BN (Fig 6D, left), while in others the overall spread appeared unchanged (Fig 6E, left).293 To quantify these differences, we identified for each FoV the minimum number of PCs that294 explained at least 60% of the variance in sound-evoked activity under both BN conditions (Fig 6D,E,295 right). We then projected the data into this reduced space and computed the manifold size as the296 total variation (TV) of the projected activity (see Methods). TV provides a robust scalar measure297 of the effective radius of the point cloud in lower-dimensional projections of varying dimensionality,298 enabling comparisons across FoVs. This analysis revealed a significant reduction in manifold size299 for L2/3 populations under BN (Fig 6F, Wilcoxon signed rank test, L2/3: p = 0.0020), while the300 L5 IT and ET populations did not show a change in TV under BN (Wilcoxon signed rank test, L5301 IT: p = 0.5771, L5 ET: p = 0.1250). Thus, the global representational space contracted in L2/3302 under BN, while remaining effectively noise invariant in L5 IT and L5 ET populations.303 These two analyses, manifold geometry and manifold size, revealed an apparent discrepancy304 for L5 ET. In L2/3, BN both reduced the efficiency of manifold geometry and contracted overall305 manifold size, whereas L5 IT remained noise-invariant in both measures. In contrast, L5 ET306 neurons showed less efficient manifold geometry under BN, but preserved their global manifold307 size. Although this pattern is difficult to visualize directly, it implies that individual manifold308 objects became larger and more high-dimensional without increasing the overall spread of the309 full set of objects. This could occur if manifolds expand inward (toward the origin) or toward310 one another rather than outward, two possibilities that are not mutually exclusive. This scenario311 is consistent with the decoding results (Fig 5D,H), where L5 ET detection and discrimination312 performance decline under BN, as would be expected if near-boundary trials are displaced toward313 neighboring frequencies or toward the origin in neural space. Overall, BN degrades the efficiency of314 population-level encoding within manifold objects in L5 ET, while the total neural subspace used315 for population-level representations remains unchanged.316 Discussion317 We imaged sound-evoked responses in defined excitatory subpopulations of the ACtx with in vivo318 two-photon microscopy. To distinguish how different circuit elements contribute to noise invariance,319 we compared an excitatory subpopulation primarily involved in local processing (L2/3) with two320 deep-layer subpopulations that broadcast information to distant targets (L5 IT and L5 ET). All321 three subpopulations adjusted their responses when tones were embedded in BN, but noise invari-322 ance was concentrated in the broadcast pathways. L2/3 neurons showed clear noise dependence,323 including suppressed single-neuron responses (Fig 2, Fig 3), increased pairwise correlations (Fig 4),324 and reduced fidelity in population-level encoding of tone identity (Fig 4, Fig 5, Fig 6). Under iden-325 tical experimental conditions, L5 IT neurons maintained stable single-neuron and population-level326 representations, with noise-related differences appearing only in pairwise correlations (Fig 4C). L5327 ET neurons expressed a more limited form of noise invariance: their single-neuron and pairwise328 responses were largely stable across BN conditions, whereas their population decoding performance329 and manifold structure were not. Together, these results show that excitatory subpopulations in330 13 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted December 22, 2025. ; https://doi.org/10.64898/2025.12.19.695511doi: bioRxiv preprint ACtx rely on different representational levels to preserve sensory information in noise and reveal a331 functional stratification of noise invariance across the cortical microcircuit.332 Significance of noise invariance for sound processing333 ACtx plays a critical role in extracting behaviorally relevant sound features [6, 22, 60–62], and the334 canonical cortical microcircuit carries and transforms sensory information within ACtx [28]. In this335 circuit, thalamocortical inputs primarily innervate L4, which relays information to L2/3, then to336 L5 and L6, where signals are broadcast widely throughout the brain [25, 26, 29, 63]. L2/3 neurons337 are the first major cortical recipients of L4 input and provide dense, complex projections to L5338 [24], a principal output layer with extensive long-range targets [29, 64–69]. Along the ascending339 auditory pathway, sound representations become increasingly noise-invariant from the periphery to340 ACtx [1], and even more so in higher-order ACtx [19]. When considered together, the hierarchical341 organization from periphery to cortex and from L4 to L2/3 to L5 supports a model in which ACtx342 integrates noise-dependent inputs, refines them through subpopulation-specific computations, and343 broadcasts increasingly noise-invariant representations of sounds that can drive adaptive behavior344 [28, 70, 71].345 Prior studies have examined sensory representations that remain invariant to factors other346 than BN. For example, work on level invariance shows that ACtx representations can remain stable347 despite changes in sound intensity [72, 73], and non-primary ACtx can form distractor-invariant348 representations of sounds when they are behaviorally relevant [74, 75]. In the visual system, a349 rich body of work has investigated how neurons in the visual pathway support object recognition350 independently of changes in rotation, position, and other variables [76–80]. A common limitation of351 many of these studies is that they treat cortical excitatory neurons as a homogeneous population,352 overlooking the heterogeneity of subpopulations within a cortical column. In light of our findings,353 this gap raises the possibility that invariant coding in other sensory modalities may also arise from354 specialized computation by distinct excitatory subpopulations.355 In this study, we used pure tones as the signal and white noise as the masker. Pure tones356 give precise control of frequency and intensity to isolate mechanisms of noise invariance and is a357 common stimulus choice in work probing noise-invariant representations in the auditory system358 [2, 3, 6, 10, 81], but they sample only a small region of the space of natural sounds. Extending359 our findings to more complex stimuli will require careful stimulus design. For example, time-360 varying stimuli, such as amplitude-modulated tones, introduce fluctuations in SNR over time and361 would require ACtx neurons to express noise invariance both for spectral content and for temporal362 envelope tracking. Human studies suggest that envelope tracking can remain relatively preserved363 in BN [82], but the contribution of ACtx neurons to time-varying noise invariance is not yet known.364 Stimuli with richer spectral structure, such as frequency sweeps and natural vocalizations, would365 demand invariance across multiple spectral and temporal channels. Although ACtx neurons exhibit366 multiplexed responses to several sound features [38, 75, 83–85], future work should test whether367 individual neurons maintain a consistent degree of noise invariance across different regions of their368 receptive fields and whether population-level codes support multiplexed forms of noise invariance369 across features and timescales.370 Differences in noise invariance between L5 subpopulations371 At the level of individual neurons, L5 IT and ET populations exhibited a similar degree of noise372 invariance in both tuning and response distributions (Fig 2B,G). However, differences become373 14 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted December 22, 2025. ; https://doi.org/10.64898/2025.12.19.695511doi: bioRxiv preprint more pronounced in pairwise and population-level representations. Pairwise analyses showed that374 L5 IT neurons exhibited reduced noise correlations across all intersomatic distances, whereas L5375 ET neurons showed reductions primarily only locally (Fig 3C-D). Noise correlations can reflect376 shared inputs, functional coupling, and the information content of population responses [86–88].377 The broad reduction in L5 IT (and L2/3) correlations suggests more global changes in correlated378 variability under BN, whereas the spatially restricted effects in L5 ET may reflect BN-dependent379 modulation of local subnetworks within this subpopulation.380 Population-level differences were even more striking. Decoding analyses showed that BN im-381 paired both sound detection and frequency discrimination in L5 ET neurons, whereas L5 IT de-382 coding performance remained stable (Fig 5D,H). Consistent with this result, manifold geometry383 analyses revealed that BN disrupted the fine-scale structure of sound representations in L5 ET384 neurons, altering manifold capacity, radius, and dimensionality, while affecting only manifold ra-385 dius in L5 IT neurons (Fig 6C). Notably, despite these changes in fine-scale geometry, the global386 structure of sound representations in both L5 IT and L5 ET populations remained noise invariant,387 as reflected by stable manifold size across BN conditions (Fig 6F).388 In L5 ET neurons, this dissociation between degraded fine-scale geometry and preserved global389 structure suggests that BN may cause individual manifold objects to expand toward one another390 or toward the origin, without changing the overall extent of the population-level representation.391 Together, these results indicate that L5 ET neurons preserve the global size of their neural man-392 ifold under BN but exhibit degraded stimulus-specific geometry, whereas L5 IT neurons maintain393 both global and fine-scale structure. This divergence may reflect differences in the functional de-394 mands of their downstream targets. If L5 IT and L5 ET projection targets differ in computational395 requirements or modularity, the corresponding cortical output pathways may differentially shape396 population responses to preserve noise-invariant representations appropriate for their target circuits.397 These findings add to a growing body of evidence that L5 IT and L5 ET neurons are function-398 ally distinct. Across cortical areas, these subpopulations differ in their projection targets as well399 as in multiple morphological and physiological properties [21, 28, 29, 65, 89, 90]. Here, we show400 that although both L5 IT and L5 ET neurons participate in broadcast pathways and exhibit noise401 invariance at multiple representational levels, they differ in the extent to which specific pairwise402 and population-level metrics remain noise invariant. One possible explanation for this differential403 modulation by BN is that L5 IT and L5 ET neurons receive distinct long-range inputs [29, 65].404 Differences in top-down modulation could lead to subpopulation-specific effects of BN by selectively405 enhancing or suppressing neuronal responses, thereby shaping the degree of noise invariance ex-406 pressed at the population level. An additional possibility is that these subpopulations differ in their407 local circuit organization, including recurrent connectivity within each group and their interactions408 with other neural populations, such as inhibitory interneurons [67, 91, 92]. Together, differences in409 long-range inputs, local recurrence, and inhibitory interactions may underlie the distinct patterns410 of noise invariance observed between L5 IT and L5 ET neurons. Future studies could directly test411 these ideas by transiently inactivating cortical regions that provide top-down input to L5 neurons,412 such as posterior parietal cortex or anterior cingulate cortex, and by using cell-type-specific optoge-413 netic perturbations to assess whether local microcircuits are differentially engaged in the presence414 of BN.415 15 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted December 22, 2025. ; https://doi.org/10.64898/2025.12.19.695511doi: bioRxiv preprint Potential mechanisms that lead to noise invariance416 Early work on noise invariance emphasized mechanisms in the auditory periphery that reduce the417 overall gain of auditory nerve responses [11, 93]. Although such mechanisms can improve sound418 coding in noisy environments, noise-dependent distortions of sound representations remain evident419 beyond the periphery, indicating that global gain adjustments alone are insufficient to account420 for noise-invariant coding. Instead, peripheral adaptations likely constitute the initial stage of a421 multistep computation that is progressively refined in downstream auditory structures.422 Cholinergic projections from the basal forebrain exhibit strong layer-, subpopulation-, and423 region-specific organization. Across cortical areas, basal forebrain inputs exert layer-specific effects424 [94, 95], and within ACtx, cholinergic innervation differs between primary and non-primary subdi-425 visions [17, 95]. At the cellular level, cholinergic signaling differentially modulates L5 IT and L5426 ET neurons [66, 96]. In parallel, sound representations become progressively more noise invariant427 across layers of the cortical microcircuit in primary ACtx and between primary and non-primary428 auditory fields, mirroring differences in cholinergic innervation and functional responses across these429 populations. Consistent with this framework, recent work has implicated cholinergic input to ACtx430 as a potential mechanism supporting noise-invariant representations, in part through its effects on431 spontaneous firing rates and local synchrony [3]. Together, these observations suggest that neu-432 romodulatory influences may contribute to subpopulation-specific differences in noise invariance433 across the auditory hierarchy and within the ACtx microcircuit.434 In addition to neuromodulatory influences, inhibitory circuits have been proposed as mech-435 anisms contributing to the construction of noise-invariant representations in ACtx. Inhibitory436 neurons play well-established roles in sensory processing, including surround suppression [5, 97,437 98], suppressive feedback [99, 100], and temporal sharpening [101, 102]. With respect to noise438 invariance, inactivation of parvalbumin (PV) or somatostatin (SOM) interneurons impairs behav-439 ioral performance in noisy conditions to a degree comparable to inactivation of ACtx itself [6].440 This result indicates that PV and SOM activity is necessary for detecting sounds in BN, but also441 suggests that inhibitory neurons alone do not fully account for the cortical mechanisms underlying442 noise-invariant perception. Consistent with this interpretation, optogenetic activation of PV neu-443 rons suppresses ACtx tuning curves in a manner similar to BN; however, combining PV activation444 with BN produces even stronger suppression, indicating that PV activity alone is insufficient to445 explain the full modulation of tuning curves observed in noisy environments [2].446 Notably, many of these studies have treated inhibitory neurons as a homogeneous population447 within ACtx. In contrast, both theoretical and experimental work suggests that feedforward inhi-448 bition may support contrast gain control within the canonical cortical microcircuit [71], potentially449 supporting noise invariance by selectively modulating the gain of individual neurons according to450 their receptive fields and bottom-up inputs. Under this framework, inhibitory neurons would them-451 selves be differentially engaged by BN and would, in turn, selectively enhance or suppress excitatory452 neurons within the same layer, thereby stabilizing sound representations in noisy conditions. A key453 prediction of this mechanism is that noise invariance should increase across the cortical microcircuit,454 from L2/3 to L5, a pattern that is consistent with our findings. Future experiments could directly455 test this hypothesis by identifying interneuron populations whose activity is selectively modulated456 by BN and determining whether their trial-by-trial influence on local excitatory neurons adjusts457 gain in a manner that promotes noise-invariant sound representations.458 Our results demonstrate that excitatory subpopulations in ACtx make distinct and complemen-459 tary contributions to constructing noise-invariant representations. We speculate that a complete460 16 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted December 22, 2025. ; https://doi.org/10.64898/2025.12.19.695511doi: bioRxiv preprint mechanism for noise invariance within ACtx must include components that differentially influence461 the three excitatory subpopulations examined here and can account for the mixed pattern of in-462 variance across single-neuron, pairwise, and population levels. Such components may include finely463 tuned gain control by inhibitory interneurons, as well as neuromodulatory inputs that are selectively464 engaged in the presence of BN.465 Conclusions466 Disentangling sensory signals from background noise is a fundamental process that enables animals467 to represent stimuli accurately and generate appropriate behavioral responses. We show that exci-468 tatory subpopulations in ACtx respond differentially to sounds in BN, depending on their laminar469 position and projection class within the cortical microcircuit. This subpopulation-specific orga-470 nization supports the idea that deep-layer broadcast pathways preferentially carry noise-invariant471 representations, whereas superficial populations remain more noise-dependent. Our findings bridge472 the noise-dependent representations observed in earlier stages of the auditory pathway with the473 more noise-invariant representations reported in higher auditory areas. Together, they underscore474 the role of excitatory subpopulations in implementing the computations that give rise to noise-475 invariant coding.476

and Methods477 Mice478 All procedures were approved by the University of Pittsburgh Animal Care and Use Committee and479 follow the National Institute of Health guidelines for the care and use of laboratory animals. Data480 were collected from 17 mice (10-16 weeks old, both male and female). For L2/3 recordings, we used481 two C57BL/6 mice (#000664, Jackson Labs), and five Emx1-Cre mice (#005628, Jackson Labs).482 For L5 IT recordings, we used six Tlx3-Cre mice (B6.FVB(Cg)-Tg(Tlx3-Cre)PL56Gsat/Mmucd,483 MMRRC). For L5 ET mice, we used four C57BL/6 mice. All mice were housed on a 12 h light/dark484 cycle with ad libitum access to food and water. All imaging was conducted during the dark cycle.485 Surgical Procedures486 Virus-mediated gene delivery487 Mice were anesthetized with 4% isoflurane and positioned in a stereotaxic frame (model 1900,488 Kopf). Throughout the procedure, a surgical plane of anesthesia was maintained using a continuous489 infusion of isoflurane (2%) in oxygen. Mice lay atop a homeothermic blanket system (Fine Science490 Tools) that maintained core body temperature at approximately 36.5°C. The scalp was shaved and491 sterilized with alternating applications of iodine and ethanol, followed by subcutaneous injection492 of lidocaine hydrochloride (5 mg/ml) for local analgesia.493 For ACtx injections, a ∼1 cm incision was made between the right eye and ear to expose the494 temporalis muscle, which was then retracted. Two burr holes (∼0.3 mm diameter each) were drilled495 along the right temporal ridge, spanning a region 1.5–2.5 mm rostral to the lambdoid suture. For496 inferior colliculus (IC) injections, a midline incision was made to expose bregma and lambda. The497 skull was leveled such that the vertical difference between bregma and lambda was less than 100498 µm, and a single burr hole was drilled at 4.8 mm caudal and 0.9 mm lateral to bregma.499 17 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted December 22, 2025. ; https://doi.org/10.64898/2025.12.19.695511doi: bioRxiv preprint Viral injections were performed using a motorized stereotaxic injector (Nanoject III, Drum-500 mond Scientific). For ACtx injections, 250 nl of either a non-conditional GCaMP8s virus (pGP-501 AAV-syn-jGCaMP8s-WPRE, Addgene, titer: 3.5 x 10 12 vg/mL) or a Cre-dependent GCaMP8s502 virus (pGP-AAV-syn-FLEX-jGCaMP8s-WPRE, Addgene, titer: 6 x 10 12 vg/mL) was delivered at503 a depth of approximately 450 µm below the pial surface at each injection site. For IC injections,504 250 nl of retrograde GCaMP8s virus (pGP-AAV-syn-jGCaMP8s-WPRE, Addgene, titer: 4 x 10 12505 vg/mL) was delivered at depths of 900 µm and 400 µm below the pial surface. Following injections,506 the surgical sites were closed, antibiotic ointment was applied, and postoperative analgesia was ad-507 ministered subcutaneously (carprofen, 5 mg/ml). Mice were provided with ad libitum access to a508 carprofen MediGel and were closely monitored for three days following surgery.509 Cranial window implantation510 Mice were brought to a surgical plane of anesthesia using the same anesthesia and temperature-511 control procedures described above. The dorsal surface of the skull was exposed, and the periosteum512 was removed. The skull was cleaned with 70% ethanol and chemically etched before affixing a513 custom titanium head plate (eMachineShop). The head plate was secured to the skull with opaque514 dental cement (C&B Metabond) and allowed to fully cure. After head-plate attachment, the515 temporalis muscle was retracted to expose the temporal ridge. A circular outline (3 mm diameter)516 centered over the temporal ridge approximately 0.5 mm above the lambdoid suture was marked517 using a biopsy punch. The skull within and around this outline was thinned using a hand drill to518 create a flat surface. Once sufficiently thinned, the outlined bone was carefully removed with a519 scalpel to expose the underlying cortex. A cranial window was constructed by placing a stack of520 glass coverslips (two 3 mm diameter and one 4 mm diameter) over the exposed brain. The edges521 of the craniotomy were sealed with silicone elastomer (Kwik-Sil) to create an airtight seal, and522 the window was secured with opaque dental cement applied around the perimeter of the 4 mm523 coverslip. All remaining exposed skull was covered with dental cement, and the surrounding skin524 was affixed to the cement using Vetbond (3M) tissue adhesive. Mice recovered under the same525 postoperative analgesia and monitoring conditions used following viral injections.526 Acoustic Stimulation527 Stimuli were generated with a 24-bit digital-to-analog converter (National Instruments model PXI-528 4461) using custom scripts written in MATLAB (MathWorks) and LabVIEW (National Instru-529 ments). Acoustic stimuli were delivered via a free-field speaker (PUI Audio) facing the left ear and530 calibrated using a free-field prepolarized microphone (377C01, PCB Piezotronics).531 Calcium Imaging532 Light-reversed mice were awake and head-fixed for all recording sessions. Prior to imaging, mice533 were habituated to head-fixation and the recording chamber for several days. Neural activity534 in response to four pure tones (4, 8, 16, and 32 kHz) were captured by widefield fluorescence535 imaging (Bergamo, ThorLabs) and used to functionally confirm the location of the right primary536 ACtx. Two-photon calcium imaging was conducted using an InSightX3 (Spectra Physics) Laser537 tuned to 940 nm and a water-immersion objective (Nikon 16x). All two-photon imaging (Bergamo,538 ThorLabs) was of the right ACtx. Mice were head-fixed upright with the microscope rotated to be539 parallel to the cranial window (approximately 40 to 50 ◦ tilt). Images were collected at 30 Hz. The540 depth below pial surface used for recordings depended on neuron subtype (L2/3: 150-250 µm, L5541 IT: 350-500 µm, L5 ET: 450-600 µm). Separate FoVs from the same mouse were at least 50 µm542 18 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted December 22, 2025. ; https://doi.org/10.64898/2025.12.19.695511doi: bioRxiv preprint above or below the original imaging plane. All two-photon calcium imaging was conducted within543 a dark, sound-attenuating chamber.544 Two imaging sessions were performed for each FoV: one without background noise (No-BN)545 and one with background noise (BN). Each session consisted of 960 trials, comprising 20 repetitions546 of 50 ms pure tones presented in pseudo-random order. Tones varied in frequency (4–45 kHz)547 and intensity (20–70 dB SPL), yielding 48 unique frequency–intensity combinations. Stimulation548 parameters were identical across sessions, with the sole difference being the presence of continuous549 white noise at 50 dB SPL delivered from a secondary speaker positioned directly below the primary550 stimulation speaker during BN sessions. Each mouse underwent as many imaging sessions as FoVs551 available. Each session lasted approximately 48 minutes, corresponding to a total of ∼88,000552 imaging frames per session.553 Data Analysis554 Image processing555 Two-photon imaging data were processed using the open-source software Suite2P [34]. Image stacks556 were motion-corrected by rigid registration, and regions of interest (ROIs) were automatically de-557 tected with neuropil subtraction. All ROIs were manually curated to ensure that they corresponded558 to individual neurons. Calcium fluorescence traces were deconvolved to estimate spike rates and559 then z-scored within each session by subtracting the baseline mean and dividing by the baseline560 standard deviation. Imaging sessions from the same field of view were aligned between conditions561 using the ROICaT cell-matching algorithm. Suite2P outputs from paired sessions were registered562 to each other, and ROIs with overlapping spatial footprints were identified as matching neurons563 across sessions.564 Responsiveness565 Neuronal responsiveness was determined using an approach adapted from Kato et al. [103]. A566 neuron was considered responsive to a given stimulus if it met two criteria: (1) sound-evoked567 responses exceeded 0.5 z-scores above baseline in at least 50% of trials, and (2) the mean sound-568 evoked response across all trials exceeded 1 z-score. Responsiveness was assessed separately for each569 unique frequency–intensity combination and aggregated between both BN conditions. A neuron570 was classified as sound responsive if it met these criteria for at least one stimulus.571 Tuning curves572 Each recorded neuron was tested with 20 trials of each unique stimulus, yielding a total of 960573 trials per session (48 frequency–intensity combinations). Neural activity within a fixed response574 window (0.5 s following sound onset) was averaged across trials to construct a frequency response575 area (FRA) for each neuron. FRAs were represented as matrices with eight rows corresponding576 to frequencies and six columns corresponding to intensities (Fig 2A). Frequency tuning curves577 were derived by averaging responses across all intensities for each frequency, and intensity tuning578 curves were derived by averaging responses across all frequencies for each intensity. To construct579 population-averaged frequency tuning curves, individual neuron tuning curves were centered on580 their best frequency (BF) and then averaged within each subpopulation. Intensity tuning curves581 were averaged across neurons without re-centering.582 19 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted December 22, 2025. ; https://doi.org/10.64898/2025.12.19.695511doi: bioRxiv preprint RMA regression583 To compare tuning curves of individual neurons across BN conditions, we used reduced major axis584 (RMA) regression [18, 36]. Unlike ordinary least-squares regression, RMA accounts for measure-585 ment noise in both variables. For each neuron, RMA was applied separately to frequency and586 intensity tuning curves to estimate slope and intercept parameters. Only slope and intercept esti-587 mates that differed significantly from their null values (slope = 1, intercept = 0) were included in588 subsequent analyses. Statistical significance was assessed using one-sample t-tests comparing the589 estimated coefficients to their respective null values, with standard errors derived from the resid-590 ual variance of the fitted model [104]. Intercepts significantly greater than or less than zero were591 classified as additive or subtractive shifts, respectively, whereas slopes significantly greater than or592 less than one were classified as multiplicative or divisive scaling.593 Mutual information594 Before computing mutual information, we empirically estimated the marginal and joint probability595 distributions P (X), P (S), and P (X, S). Neural responses X were discretized by binning sound-596 evoked activity using a fixed bin width of 0.2 z-scored responses. Stimulus identity S was encoded597 as an integer index ranging from 1 to the total number of unique stimuli (48) and had a uniform598 distribution by experimental design. The joint distribution P (X, S) was obtained by histogramming599 sound-evoked responses separately for each stimulus and normalizing across all trials.600 Mutual information between the neural response and stimulus identity was computed as: I(X; S) = X xϵX X sϵS P (x, s)log2  P (x, s) P (x)P (s)  This calculation yielded a single non-negative scalar value quantifying how much information a neuron’s responses conveyed about stimulus identity. Mutual information was computed separately for the BN and No-BN conditions. Because the stimulus set represented only a limited sampling of the auditory space encoded by ACtx neurons, we applied a bias correction to each mutual informa- tion estimate. For each neuron, bias was estimated using the analytically derived approximation from Panzeri et al. [105]: bias = P sϵS |RS − 1| − ||R| − 1| 2N ln(2) where |R| denotes the number of response bins with nonzero probability in P (X), |Rs| denotes601 the number of response bins with nonzero probability in the conditional distribution P (X | S = s),602 and N is the total number of stimulus presentations. The estimated bias was subtracted from each603 mutual information value.604 To quantify how background noise altered stimulus-related information on a neuron-by-neuron basis, we computed an information modulation index (IMI): IM Ij = Ij,BN − Ij,N o−BN Ij,BN + Ij,N o−BN Each neuron was thus assigned an IMI value ranging from −1 to +1. Positive IMI values605 indicate greater stimulus-related information in BN than in No-BN, values near zero indicate little606 difference between conditions, and negative values indicate greater information encoding in No-BN.607 20 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted December 22, 2025. ; https://doi.org/10.64898/2025.12.19.695511doi: bioRxiv preprint Pairwise correlations608 Noise correlations (spike-count correlations) were computed by subtracting, for each neuron, its609 mean response across trials for a given stimulus and then calculating the Pearson correlation coeffi-610 cient between the resulting trial-by-trial residuals for each pair of simultaneously recorded neurons.611 Signal correlations (tuning correlations) were computed as the Pearson correlation between the612 corresponding entries of the FRAs of pairs of simultaneously recorded neurons. All correlation613 analyses were restricted to neurons classified as sound responsive.614 Detection and discrimination decoding615 We evaluated stimulus-encoded information at the population level using two decoding approaches:616 discrimination and detection. For detection decoding, we constructed training datasets consisting617 of sound-on and sound-off trials. Sound-on trials were drawn from responses to tones presented at618 60 and 70 dB SPL for a single frequency, whereas sound-off trials consisted of randomly sampled619 neural activity measured 2 s after sound onset, when no stimulus was present. Decoders were620 trained separately for each frequency to avoid confounds arising from frequency tuning differences621 between training and testing data. We trained a neural network with the same architecture used622 for frequency discrimination decoding, except that the output layer consisted of a single artificial623 neuron. The output neuron was normalized to values between 0 and 1 using a sigmoid activation624 function, and trials were classified as sound-present if the output exceeded a threshold of 0.5.625 Detection performance was quantified using d’ (d-prime), which incorporates both hit rate (HR) and false alarm rate (F AR ): d′ = Z(HR) − Z(F AR) HR = #hits #hits + #misses F AR = #f alse alarms #f alse alarms + #correct rejections For discrimination decoding, we trained a neural network to classify stimulus frequency based626 on the activity of 40 simultaneously recorded neurons. Neural activity was passed through two627 hidden layers of 16 artificial neurons each and then into an output layer of eight artificial neurons,628 one corresponding to each stimulus frequency. Hidden layers used rectified linear unit (ReLU)629 activation functions, and the output layer was normalized using a softmax function. Network per-630 formance was assessed using ten-fold cross-validation, and decoding accuracy was quantified as631 the mean classification accuracy across folds. To control for intensity-dependent effects, discrimi-632 nation decoding was performed using trials from a single sound intensity at a time and repeated633 independently for each intensity.634 Manifold geometry635 To characterize the population-level geometry of sound-evoked neural activity, we applied a mean-636 field-theoretic manifold geometry analysis [55]. This framework quantifies the classification capacity637 of a neural population by describing the geometric structure of population responses to different638 stimulus categories. For each frequency, we extracted population activity from a single FoV and639 represented each trial as a vector in Rn, where each dimension corresponded to the activity of640 one neuron. To standardize dimensionality across FoVs and subpopulations, we analyzed randomly641 21 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted December 22, 2025. ; https://doi.org/10.64898/2025.12.19.695511doi: bioRxiv preprint sampled subsets of n = 40 neurons and resampled with replacement to approximate the full recorded642 population. Trials corresponding to the same stimulus category formed a point cloud in neural643 space, which was treated as a single neural manifold. For each manifold, we extracted three644 geometric metrics: manifold radius (RM), effective dimensionality (DM), and classification capacity645 (αM). We performed this analysis in two complementary ways. First, we pooled trials across all646 intensities for each frequency, yielding one manifold per frequency (Fig 6B). Second, we constructed647 separate manifolds for each frequency-intensity combination, yielding a finer-grained analysis of648 population geometry (Fig 6C).649 Manifold geometry analysis quantifies how efficiently population responses to different stimuli650 can be separated in high-dimensional neural space. The manifold radius RM measures the spatial651 extent of trial-to-trial variability within a stimulus category, whereas the effective dimensionality652 DM reflects the number of dimensions required to capture this variability. Together, these metrics653 describe how compact and low-dimensional a stimulus representation is, properties that facilitate654 separability from other stimulus manifolds. For example, if responses to a given frequency be-655 come more consistent across trials, the corresponding point cloud contracts toward a single point,656 resulting in smaller values of RM and DM.657 The manifold capacity αM provides an integrated measure of coding efficiency by quantifying658 how many such manifolds can be linearly separated by the same neural population. In the mean-659 field framework, αM is inversely related to both RM and DM; thus, reductions in capacity indicate660 increased manifold size, increased dimensionality, or both. For our purposes, αM serves as a661 compact summary metric linking population geometry to the efficiency of stimulus encoding.662 Manifold size663 To quantify the global size of the neural manifold, we first projected population activity into a664 lower-dimensional subspace that captured a substantial fraction of the variance in neural responses.665 For each neuron, we computed the peri-stimulus time–averaged activity for each unique stimulus666 (48 total), pooling trials from both No-BN and BN sessions. The activity of all simultaneously667 recorded neurons was then aggregated into a t × n matrix, where t denotes time points and n668 denotes the number of neurons. We reduced the dimensionality of this matrix using principal669 component analysis (PCA), projecting the data into a t × p subspace, where p was chosen as the670 minimum number of principal components required to explain at least 60% of the total variance.671 After projecting the neural activity into this reduced space, we quantified the overall size of the672 population representation by computing its total variation (TV), defined as673 T V = trace(Σ) where Σ is the covariance matrix of the projected neural activity. TV provides a scalar measure674 of the overall spread of population responses in the reduced neural space and thus serves as an675 estimate of global manifold size.676 Statistical Analysis677 All statistical analyses were performed in MATLAB (MathWorks). Data are reported as mean ±678 s.e.m. unless otherwise noted. When data did not meet assumptions of normality, nonparametric679 statistical tests were used. For multi-group comparisons, post hoc tests were performed using680 22 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted December 22, 2025. ; https://doi.org/10.64898/2025.12.19.695511doi: bioRxiv preprint Tukey’s or Dunn’s procedures, as appropriate. Statistical significance in figures is denoted as *681 p < 0.05, ** p < 0.01, and *** p < 0.0001.682 Acknowledgments683 We thank current and former members of the Williamson Lab for helpful feedback and discus-684 sions. This work was supported by NIH/NIDCD grants R21DC018327 and R01DC020459, and the685 Klingenstein-Simons Fellowship in Neuroscience to RSW.686 Author Contributions687 TSO and RSW conceptualized all experiments. TSO collected and analyzed all data. TSO and688 RSW prepared the figures and wrote the manuscript.689 Declaration of Competing Interests690 The authors declare no competing interests.691 References692 1. Rabinowitz, N. C., Willmore, B. D. B., King, A. J. & Schnupp, J. W. H. Constructing noise-693 invariant representations of sound in the auditory pathway. PLoS Biology 11, e1001710694 (2013).695 2. Christensen, R. K., Lind´ en, H., Nakamura, M. & Barkat, T. R. White noise background696 improves tone discrimination by suppressing cortical tuning curves. Cell reports 29, 2041–697 2053.e4 (2019).698 3. Christensen, R. K., Studer, F. & Barkat, T. R. Background white noise increases neuronal699 activity by reducing membrane fluctuations and slow-wave oscillations in auditory cortex.700 Progress in Neurobiology 246, 102720 (2025).701 4. Carruthers, I. M. et al. Emergence of invariant representation of vocalizations in the auditory702 cortex. Journal of Neurophysiology 114, 2726–2740 (2015).703 5. Lakunina, A. A., Nardoci, M. B., Ahmadian, Y. & Jaramillo, S. Somatostatin-Expressing704 Interneurons in the Auditory Cortex Mediate Sustained Suppression by Spectral Surround.705 The Journal of Neuroscience 40, 3564–3575 (2020).706 6. Lakunina, A. A., Menashe, N. & Jaramillo, S. Contributions of distinct auditory cortical707 inhibitory neuron types to the detection of sounds in background noise. eNeuro 9 (2022).708 7. Souffi, S., Nodal, F. R., Bajo, V. M. & Edeline, J.-M. When and how does the auditory709 cortex influence subcortical auditory structures? new insights about the roles of descending710 cortical projections. Frontiers in Neuroscience 15, 690223 (2021).711 8. Souffi, S, Lorenzi, C, Huetz, C & Edeline, J.-M. Robustness to noise in the auditory system:712 A distributed and predictable property. eNeuro 8 (2021).713 9. Schneider, D. M. & Woolley, S. M. N. Sparse and background-invariant coding of vocalizations714 in auditory scenes. Neuron 79, 141–152 (2013).715 10. Reiss, L. A. J., Ramachandran, R. & May, B. J. Effects of signal level and background716 noise on spectral representations in the auditory nerve of the domestic cat. Journal of the717 Association for Research in Otolaryngology 12, 71–88 (2011).718 23 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted December 22, 2025. ; https://doi.org/10.64898/2025.12.19.695511doi: bioRxiv preprint 11. Sachs, M. B., Voigt, H. F. & Young, E. D. Auditory nerve representation of vowels in back-719 ground noise. Journal of Neurophysiology 50, 27–45 (1983).720 12. Mokri, Y., Worland, K., Ford, M. & Rajan, R. Effect of background noise on neuronal721 coding of interaural level difference cues in rat inferior colliculus. The European Journal of722 Neuroscience 42, 1685–1704 (2015).723 13. Liang, F., Bai, L., Tao, H. W., Zhang, L. I. & Xiao, Z. Thresholding of auditory cortical724 representation by background noise. Frontiers in Neural Circuits 8, 133 (2014).725 14. Brugge, J. F., Reale, R. A. & Hind, J. E. Spatial receptive fields of primary auditory cortical726 neurons in quiet and in the presence of continuous background noise. Journal of Neurophys-727 iology 80, 2417–2432 (1998).728 15. Ruggero, M. A., Narayan, S. S., Temchin, A. N. & Recio, A. Mechanical bases of fre-729 quency tuning and neural excitation at the base of the cochlea: comparison of basilar-730 membrane vibrations and auditory-nerve-fiber responses in chinchilla. Proceedings of the731 National Academy of Sciences of the United States of America 97, 11744–11750 (2000).732 16. Davis, R. L. & Liu, Q. Complex primary afferents: What the distribution of electrophysiologically-733 relevant phenotypes within the spiral ganglion tells us about peripheral neural coding. Hear-734 ing Research 276, 34–43 (2011).735 17. Bajo, V. M., Leach, N. D., Cordery, P. M., Nodal, F. R. & King, A. J. The cholinergic736 basal forebrain in the ferret and its inputs to the auditory cortex. The European Journal of737 Neuroscience 40, 2922–2940 (2014).738 18. Mesgarani, N., David, S. V., Fritz, J. B. & Shamma, S. A. Mechanisms of noise robust739 representation of speech in primary auditory cortex. Proceedings of the National Academy of740 Sciences of the United States of America 111, 6792–6797 (2014).741 19. Kell, A. J. E. & McDermott, J. H. Invariance to background noise as a signature of non-742 primary auditory cortex. Nature Communications 10, 3958 (2019).743 20. Weible, A. P., Liu, C., Niell, C. M. & Wehr, M. Auditory cortex is required for fear potenti-744 ation of gap detection. The Journal of Neuroscience 34, 15437–15445 (2014).745 21. Schneider, N. A., Malina, M. I. & Williamson, R. S. A cortical output channel for perceptual746 categorization. BioRxiv (2025).747 22. Krall, R. F. et al. Primary auditory cortex is necessary for the acquisition and expression of748 categorical behavior. Current Biology 35, 4061–4074.e4 (2025).749 23. Kaufman, K. J., Krall, R. F. & Williamson, R. S. Pupil-linked arousal differentially modulates750 cell-type-specific sensory processing. BioRxiv (2025).751 24. Meng, X., Winkowski, D. E., Kao, J. P. Y. & Kanold, P. O. Sublaminar subdivision of752 mouse auditory cortex layer 2/3 based on functional translaminar connections. The Journal753 of Neuroscience 37, 10200–10214 (2017).754 25. Llano, D. A. & Sherman, S. M. Differences in intrinsic properties and local network connec-755 tivity of identified layer 5 and layer 6 adult mouse auditory corticothalamic neurons support756 a dual corticothalamic projection hypothesis. Cerebral Cortex 19, 2810–2826 (2009).757 26. Williamson, R. S. & Polley, D. B. Parallel pathways for sound processing and functional758 connectivity among layer 5 and 6 auditory corticofugal neurons. eLife 8 (2019).759 27. Tasic, B. et al. Shared and distinct transcriptomic cell types across neocortical areas. Nature760 563, 72–78 (2018).761 28. Harris, K. D. & Shepherd, G. M. G. The neocortical circuit: themes and variations. Nature762 Neuroscience 18, 170–181 (2015).763 29. Kim, E. J., Juavinett, A. L., Kyubwa, E. M., Jacobs, M. W. & Callaway, E. M. Three Types764 of Cortical Layer 5 Neurons That Differ in Brain-wide Connectivity and Function. Neuron765 88, 1253–1267 (2015).766 24 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted December 22, 2025. ; https://doi.org/10.64898/2025.12.19.695511doi: bioRxiv preprint 30. Winer, J. A. Decoding the auditory corticofugal systems. Hearing Research 207, 1–9 (2005).767 31. Harris, K. D. & Mrsic-Flogel, T. D. Cortical connectivity and sensory coding. Nature 503,768 51–58 (2013).769 32. Asokan, M. M., Williamson, R. S., Hancock, K. E. & Polley, D. B. Sensory overamplification770 in layer 5 auditory corticofugal projection neurons following cochlear nerve synaptic damage.771 Nature Communications 9, 2468 (2018).772 33. Chen, X. et al. High-Throughput Mapping of Long-Range Neuronal Projection Using In Situ773 Sequencing. Cell 179, 772–786.e19 (2019).774 34. Pachitariu, M. et al. Suite2p: beyond 10,000 neurons with standard two-photon microscopy.775 BioRxiv (2016).776 35. Friedrich, J., Zhou, P. & Paninski, L. Fast online deconvolution of calcium imaging data.777 PLoS Computational Biology 13, e1005423 (2017).778 36. Phillips, E. A. & Hasenstaub, A. R. Asymmetric effects of activating and inactivating cortical779 interneurons. eLife 5 (2016).780 37. Nelken, I. & Chechik, G. Information theory in auditory research. Hearing Research 229,781 94–105 (2007).782 38. Walker, K. M. M., Bizley, J. K., King, A. J. & Schnupp, J. W. H. Multiplexed and robust783 representations of sound features in auditory cortex. The Journal of Neuroscience 31, 14565–784 14576 (2011).785 39. Zan, P., Presacco, A., Anderson, S. & Simon, J. Z. Mutual information analysis of neural786 representations of speech in noise in the aging midbrain. Journal of Neurophysiology 122,787 2372–2387 (2019).788 40. Nocon, J. C. et al. A robust and compact population code for competing sounds in auditory789 cortex. Journal of Neurophysiology 130, 775–787 (2023).790 41. Cohen, M. R. & Kohn, A. Measuring and interpreting neuronal correlations. Nature Neuro-791 science 14, 811–819 (2011).792 42. Kohn, A. & Smith, M. A. Stimulus dependence of neuronal correlation in primary visual793 cortex of the macaque. The Journal of Neuroscience 25, 3661–3673 (2005).794 43. Zohary, E, Shadlen, M. N. & Newsome, W. T. Correlated neuronal discharge rate and its795 implications for psychophysical performance. Nature 370, 140–143 (1994).796 44. Abbott, L. F. & Dayan, P. The effect of correlated variability on the accuracy of a population797 code. Neural Computation 11, 91–101 (1999).798 45. Smith, M. A. & Kohn, A. Spatial and temporal scales of neuronal correlation in primary799 visual cortex. The Journal of Neuroscience 28, 12591–12603 (2008).800 46. Smith, M. A. & Sommer, M. A. Spatial and temporal scales of neuronal correlation in visual801 area V4. The Journal of Neuroscience 33, 5422–5432 (2013).802 47. Smith, G. B. et al. The development of cortical circuits for motion discrimination. Nature803 Neuroscience 18, 252–261 (2015).804 48. Yu, Y., Stirman, J. N., Dorsett, C. R. & Smith, S. L. Visual information is broadcast among805 cortical areas in discrete channels (2024).806 49. Khoury, C. F., Fala, N. G. & Runyan, C. A. The spatial scale of somatostatin subnetworks807 increases from sensory to association cortex. Cell reports 40, 111319 (2022).808 50. Semedo, J. D., Zandvakili, A., Machens, C. K., Yu, B. M. & Kohn, A. Cortical Areas Interact809 through a Communication Subspace. Neuron 102, 249–259.e4 (2019).810 51. Semedo, J. D. et al. Feedforward and feedback interactions between visual cortical areas use811 different population activity patterns. Nature Communications 13, 1099 (2022).812 25 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted December 22, 2025. ; https://doi.org/10.64898/2025.12.19.695511doi: bioRxiv preprint 52. Russo, A. A. et al. Neural Trajectories in the Supplementary Motor Area and Motor Cortex813 Exhibit Distinct Geometries, Compatible with Different Classes of Computation. Neuron814 107, 745–758.e6 (2020).815 53. MacDowell, C. J. et al. Multiplexed subspaces route neural activity across brain-wide net-816 works. Nature Communications 16, 3359 (2025).817 54. Srinath, R., Ruff, D. A. & Cohen, M. R. Attention improves information flow between neu-818 ronal populations without changing the communication subspace. Current Biology 31, 5299–819 5313.e4 (2021).820 55. Chung, S., Lee, D. D. & Sompolinsky, H. Classification and geometry of general perceptual821 manifolds. Physical Review X 8, 031003 (2018).822 56. Paraouty, N. et al. Sensory cortex plasticity supports auditory social learning. Nature Com-823 munications 14, 5828 (2023).824 57. Avitan, L. & Stringer, C. Not so spontaneous: Multi-dimensional representations of behaviors825 and context in sensory areas. Neuron 110, 3064–3075 (2022).826 58. Stringer, C., Pachitariu, M., Steinmetz, N., Carandini, M. & Harris, K. D. High-dimensional827 geometry of population responses in visual cortex. Nature 571, 361–365 (2019).828 59. Cohen, U., Chung, S., Lee, D. D. & Sompolinsky, H. Separability and geometry of object829 manifolds in deep neural networks. Nature Communications 11, 746 (2020).830 60. Drieu, C. et al. Rapid emergence of latent knowledge in the sensory cortex drives learning.831 Nature 641, 960–970 (2025).832 61. Blackwell, J. M. et al. Stable encoding of sounds over a broad range of statistical parameters833 in the auditory cortex. The European Journal of Neuroscience 43, 751–764 (2016).834 62. Kuchibhotla, K. & Bathellier, B. Neural encoding of sensory and behavioral complexity in835 the auditory cortex. Current Opinion in Neurobiology 52, 65–71 (2018).836 63. Steinfeld, R., Tac˜ ao-Monteiro, A. & Renart, A. Differential representation of sensory infor-837 mation and behavioral choice across layers of the mouse auditory cortex. Current Biology838 34, 2200–2211.e6 (2024).839 64. Douglas, R. J. & Martin, K. A. C. Neuronal circuits of the neocortex. Annual Review of840 Neuroscience 27, 419–451 (2004).841 65. Baker, A. L., O’Toole, R. J. & Gulledge, A. T. Preferential cholinergic excitation of cortico-842 pontine neurons. The Journal of Physiology 596, 1659–1679 (2018).843 66. Moberg, S. & Takahashi, N. Neocortical layer 5 subclasses: From cellular properties to roles844 in behavior. Frontiers in synaptic neuroscience 14, 1006773 (2022).845 67. Onodera, K. & Kato, H. K. Translaminar recurrence from layer 5 suppresses superficial846 cortical layers. Nature Communications 13, 2585 (2022).847 68. Frezel, N et al. In-Depth Characterization of Layer 5 Output Neurons of the Primary So-848 matosensory Cortex Innervating the Mouse Dorsal Spinal Cord. Cerebral Cortex Communi-849 cations 1, tgaa052 (2020).850 69. Tang, L. & Higley, M. J. Layer 5 circuits in V1 differentially control visuomotor behavior.851 Neuron 105, 346–354.e5 (2020).852 70. Bastos, A. M. et al. Canonical microcircuits for predictive coding.Neuron 76, 695–711 (2012).853 71. Willmore, B. D. B., Cooke, J. E. & King, A. J. Hearing in noisy environments: noise invariance854 and contrast gain control. The Journal of Physiology 592, 3371–3381 (2014).855 72. Sadagopan, S. & Wang, X. Level invariant representation of sounds by populations of neurons856 in primary auditory cortex. The Journal of Neuroscience 28, 3415–3426 (2008).857 73. Sun, W., Marongelli, E. N., Watkins, P. V. & Barbour, D. L. Decoding sound level in the858 marmoset primary auditory cortex. Journal of Neurophysiology 118, 2024–2033 (2017).859 26 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted December 22, 2025. ; https://doi.org/10.64898/2025.12.19.695511doi: bioRxiv preprint 74. Heller, C. R., Hamersky, G. R. & David, S. V. Task-specific invariant representation in860 auditory cortex. eLife 12 (2024).861 75. Town, S. M., Wood, K. C. & Bizley, J. K. Sound identity is represented robustly in auditory862 cortex during perceptual constancy. Nature Communications 9, 4786 (2018).863 76. DiCarlo, J. J. & Cox, D. D. Untangling invariant object recognition. Trends in Cognitive864 Sciences 11, 333–341 (2007).865 77. Logothetis, N. K., Pauls, J, B¨ ulthoff, H. H. & Poggio, T. View-dependent object recognition866 by monkeys. Current Biology 4, 401–414 (1994).867 78. Riesenhuber, M & Poggio, T. Hierarchical models of object recognition in cortex. Nature868 Neuroscience 2, 1019–1025 (1999).869 79. Mocz, V., Vaziri-Pashkam, M., Chun, M. M. & Xu, Y. Predicting Identity-Preserving Object870 Transformations across the Human Ventral Visual Stream. The Journal of Neuroscience 41,871 7403–7419 (2021).872 80. Benucci, A. Motor-related signals support localization invariance for stable visual perception.873 PLoS Computational Biology 18, e1009928 (2022).874 81. Resnik, J. & Polley, D. B. Cochlear neural degeneration disrupts hearing in background noise875 by increasing auditory cortex internal noise. Neuron 109, 984–996.e4 (2021).876 82. Souffi, S. et al. Reduction in sound discrimination in noise is related to envelope similarity877 and not to a decrease in envelope tracking abilities. The Journal of Physiology 601, 123–149878 (2023).879 83. Giroud, J., Tr´ ebuchon, A., Mercier, M., Davis, M. H. & Morillon, B. The human auditory880 cortex concurrently tracks syllabic and phonemic timescales via acoustic spectral flux.Science881 Advances 10, eado8915 (2024).882 84. Angeloni, C & Geffen, M. N. Contextual modulation of sound processing in the auditory883 cortex. Current Opinion in Neurobiology 49, 8–15 (2018).884 85. Downer, J. D., Verhein, J. R., Rapone, B. C., O’Connor, K. N. & Sutter, M. L. An emergent885 population code in primary auditory cortex supports selective attention to spectral and886 temporal sound features. The Journal of Neuroscience 41, 7561–7577 (2021).887 86. Panzeri, S., Moroni, M., Safaai, H. & Harvey, C. D. The structures and functions of correla-888 tions in neural population codes. Nature Reviews. Neuroscience 23, 551–567 (2022).889 87. Kohn, A., Coen-Cagli, R., Kanitscheider, I. & Pouget, A. Correlations and neuronal popu-890 lation information. Annual Review of Neuroscience 39, 237–256 (2016).891 88. Averbeck, B. B., Latham, P. E. & Pouget, A. Neural correlations, population coding and892 computation. Nature Reviews. Neuroscience 7, 358–366 (2006).893 89. Li, N., Chen, T.-W., Guo, Z. V., Gerfen, C. R. & Svoboda, K. A motor cortex circuit for894 motor planning and movement. Nature 519, 51–56 (2015).895 90. Lur, G., Vinck, M. A., Tang, L., Cardin, J. A. & Higley, M. J. Projection-Specific Visual896 Feature Encoding by Layer 5 Cortical Subnetworks. Cell reports 14, 2538–2545 (2016).897 91. Naka, A. & Adesnik, H. Inhibitory circuits in cortical layer 5. Frontiers in Neural Circuits898 10, 35 (2016).899 92. Sun, Y. J., Kim, Y.-J., Ibrahim, L. A., Tao, H. W. & Zhang, L. I. Synaptic mechanisms900 underlying functional dichotomy between intrinsic-bursting and regular-spiking neurons in901 auditory cortical layer 5. The Journal of Neuroscience 33, 5326–5339 (2013).902 93. Wagner, W., Frey, K., Heppelmann, G., Plontke, S. K. & Zenner, H.-P. Speech-in-noise903 intelligibility does not correlate with efferent olivocochlear reflex in humans with normal904 hearing. Acta Oto-Laryngologica 128, 53–60 (2008).905 94. Verhoog, M. B. et al. Layer-specific cholinergic control of human and mouse cortical synaptic906 plasticity. Nature Communications 7, 12826 (2016).907 27 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted December 22, 2025. ; https://doi.org/10.64898/2025.12.19.695511doi: bioRxiv preprint 95. Chavez, C. & Zaborszky, L. Basal Forebrain Cholinergic-Auditory Cortical Network: Primary908 Versus Nonprimary Auditory Cortical Areas. Cerebral Cortex 27, 2335–2347 (2017).909 96. Baker, A. et al. Specialized Subpopulations of Deep-Layer Pyramidal Neurons in the Neocor-910 tex: Bridging Cellular Properties to Functional Consequences. The Journal of Neuroscience911 38, 5441–5455 (2018).912 97. Adesnik, H., Bruns, W., Taniguchi, H., Huang, Z. J. & Scanziani, M. A neural circuit for913 spatial summation in visual cortex. Nature 490, 226–231 (2012).914 98. Ozeki, H., Finn, I. M., Schaffer, E. S., Miller, K. D. & Ferster, D. Inhibitory stabilization of915 the cortical network underlies visual surround suppression. Neuron 62, 578–592 (2009).916 99. Hishida, R. et al. Feedback inhibition derived from the posterior parietal cortex regulates917 the neural properties of the mouse visual cortex. The European Journal of Neuroscience 50,918 2970–2987 (2019).919 100. Nassi, J. J., Lomber, S. G. & Born, R. T. Corticocortical feedback contributes to surround920 suppression in V1 of the alert primate. The Journal of Neuroscience 33, 8504–8517 (2013).921 101. Morabito, A. et al. Distinct dendritic integration strategies control dynamics of inhibition in922 the neocortex. Neuron 113, 2962–2978.e10 (2025).923 102. Cai, D. et al. A critical role of inhibition in temporal processing maturation in the primary924 auditory cortex. Cerebral Cortex 28, 1610–1624 (2018).925 103. Kato, H. K., Asinof, S. K. & Isaacson, J. S. Network-Level Control of Frequency Tuning in926 Auditory Cortex. Neuron 95, 412–423.e4 (2017).927 104. Allen, M. P. in Understanding Regression Analysis (1997).928 105. Panzeri, S., Senatore, R., Montemurro, M. A. & Petersen, R. S. Correcting for the sampling929 bias problem in spike train information measures. Journal of Neurophysiology 98, 1064–1072930 (2007).931 28 (which was not certified by peer review) is the author/funder. All rights reserved. No reuse allowed without permission. The copyright holder for this preprintthis version posted December 22, 2025. ; https://doi.org/10.64898/2025.12.19.695511doi: bioRxiv preprint