Robustness to Noise in the Auditory System: A Distributed and Predictable Property

The decrease in EI values is more pronounced in chorus noise than in stationary noise in each auditory structure. A, Raster plots of responses of the four original vocalizations, noisy vocalizations (in both noises), and noise alone recorded in CN, CNIC, MGv, A1, and VRB. The gray part of rasters corresponds to the evoked activity. For each structure, all the rasters correspond to the same recording. B, Rasters showing examples of neuronal responses in stationary noise with values of EI > 0 corresponding to a signal-like response (left, IC recording) and EI < 0 corresponding to a masker-like response (right, A1 recording). Top panels show the responses to the original vocalizations, the middle panels the responses to vocalizations at the 0-dB SNR in stationary noise and the bottom panels the responses to stationary noise alone. C, Box plots showing the EI values for the three SNRs obtained in CN (in black), CNIC (in green), MGv (in orange), A1 (in blue), and VRB (in purple) alternatively in stationary noise (SN) and chorus noise (CN). In each box plot, the red dot represents the mean value. The black lines represent significant differences between the mean values (one-way ANOVAs, p < 0.001; with post hoc paired t tests, p^a_{+10 dB,CN} = 4.03e-18, p^b_{+10 dB, CNIC} = 1.45e-07, p^c_{0dB, CNIC} = 2.47e-45, p^d_{0dB, MGv} = 2.11e-30, p^e_{0dB, A1} = 5.41e-25, p^f_0dB,VRB = 6.36e-10, p^g_{-10 dB, CN} = 5.74e-12, p^h_{-10 dB, CNIC} = 1.83e-59, pⁱ_{-10 dB, MGv} = 1.16e-36, p^j_{-10 dB, A1} = 2.84e-24, p^k_{+10 dB, VRB} = 4.62e-11).

The choice of five clusters is optimal to reveal the different behaviors in both noises. A, Mean square error of EI profile clustering as a function of the number of clusters using the K-means algorithm for the stationary and chorus noise. B, C, Population average EI profile (±SEM) of each cluster when considering six clusters to separate the data in the stationary noise (B) and in the chorus noise (C). Note that in both noises, two clusters have similar mean EI profile, i.e., the same EI evolution across the three SNRs (the two gray clusters in B and the two blue clusters in C) leading us to consider only five clusters in the following results (Fig. 4).

Robustness to noise is a distributed property in the auditory system. A, Each row corresponds to the EI profile of a given neuronal recording obtained in the five auditory structures in stationary noise with a color code from blue to red when progressing from low to high EI values. On the right, five stacked colors delineate the identity for the five categories of responses. The signal-like category is in green, the signal-dominated category in pink, the balanced category in turquoise, the insensitive category in gray and the masker-like category in yellow. The names of the categories used in the study by Ni et al. (2017) are provided for comparison. B–E. 3D representation of the five categories in stationary noise (B), mean EI values (±SEM) of the five categories (C), relative proportions of each category in stationary noise (D), and proportion of each category in the five auditory structures from CN to VRB (E). F–J, Same representations as in A–E for the responses collected in the chorus noise. See Table 1, selection type (b), for referring to the number of selected recordings in each structure.

The noise-type sensitivity is found at each stage of the auditory system. A1, Group-switching matrix representing the percentage of recordings in a given category in chorus noise depending on the category originally assigned in the stationary noise. The abscissas indicate the category identity in the stationary noise, and the ordinates represent the category identity in the chorus noise. For example, signal-like responses in stationary noise are also 50% signal-like in chorus noise but 10% are reclassified as signal-dominated, 35% balanced, 1.5% insensitive, and 3.5% masker-like. Note that, in stationary noise, the number of recordings in each category were 139, 346, 83, 540, and 159 in signal-like, signal-dominated, balanced, insensitive, and masker-like category, respectively. A2, Mean percentages of recordings changing category from the stationary noise to the chorus noise, first in each category and second in each structure [VRB, (in purple), A1 (in blue), MGv (in orange), CNIC (in green), and CN (in black)]. B1, Group-switching matrix representing the percentages of recordings changing category from the stationary noise to the chorus noise based only on recordings considered as reliable with the bootstrap procedure in the two types of noise (with a confidence interval ⩾95%). B2, Mean percentages of recordings changing category from the stationary noise to the chorus noise, first in each category and second in each structure [VRB, (in purple), A1 (in blue), MGv (in orange), CNIC (in green), and CN (in black)].

Descriptors of the categories in stationary noise used by the classifiers. A–C, Three TFRP parameters were chosen as descriptors: the BF firing rate, the bandwidth and the response duration. D, E, Two signal descriptors were selected corresponding to the firing rate and the CorrCoef values obtained in original conditions. F, G, Two main masker descriptors were presented corresponding to the firing rate and the CorrCoef values obtained in stationary noise alone. H, I, The three other masker descriptors are: (H) the ratio between the masker firing rate taken at the time the signal should have occurred and the initial masker firing rate during the first 200 ms of the masker and the number of action potentials emitted during the first (initial; I) and last (final, I) 50 ms of the masker alone over a 564-ms period. J, K, Two descriptors of the signal-to-masker ratio are presented and taken into account the firing rate of responses to the signal and to the masker; the two differ only on which part of the response to the masker is taken into account (see Materials and Methods). For each violin plot, the red dot represents the median value and the black lines represent significant differences between the median values (Kruskal–Wallis test, p < 0.05).

Descriptors of the categories in chorus noise used by the classifiers. A–C, Three TFRP parameters were chosen as descriptors: the BF firing rate, the bandwidth and the response duration. D, E, Two signal descriptors were selected corresponding to the firing rate and the CorrCoef values obtained in original conditions. F, G, Two main masker descriptors were presented corresponding to the firing rate and the CorrCoef values obtained in chorus noise alone. H, I, The three other masker descriptors are: (H) the ratio between the masker firing rate taken at the time the signal should have occurred and the initial masker firing rate during the first 200 ms of the masker and the number of action potentials emitted during the first (initial; I) and last (final; I) 50 ms of the masker alone over a 564-ms period. J, K, Two descriptors of the signal-to-masker ratio are presented and taken into account the firing rate of responses to the signal and to the masker; the two differ only on which part of the response to the masker is taken into account (see Materials and Methods). For each violin plot, the red dot represents the median value and the black lines represent significant differences between the median values (Kruskal–Wallis test, p < 0.05).

The neuronal behaviors in stationary and chorus noise are predictable based on response parameters obtained in quiet. A, All tested combinations (1–15) based on four types of descriptors (TFRP, signal, masker, and signal/masker) of the categories in stationary noise and their respective percentages of accuracy of the classifier. The gray part means that the descriptor is included in the classifier and the white part means that the descriptor is excluded from the classifier. B, C, Example of the confusion matrix obtained with all descriptors (combination 1) in stationary (B) and chorus (C) noise. Each row corresponds to a true category and each column corresponds to a predicted category. The numbers in the confusion matrix correspond to the percentage of recordings of a given true category which have been predicted to belong to a given predicted category.