Abstract
Discrimination of sensory stimuli is fundamentally constrained by the information encoded in neuronal responses. In the barn owl, interaural time difference (ITD) serves as a primary cue for azimuthal sound localization and is represented topographically in the midbrain auditory space map in the external nucleus of the inferior colliculus (ICx). While prior studies have demonstrated a correspondence between spatial tuning and behavioral acuity, it remains unclear how changes in sensory reliability influence this relationship. Here, we examined how behavioral and neuronal ITD discrimination thresholds vary with binaural correlation (BC), which manipulates ITD cue reliability. Using the pupil dilation response as a behavioral metric in head-fixed owls of either sex, we found that ITD just-noticeable differences increased exponentially as BC decreased. In contrast, the widths of ICx ITD tuning curves increased more modestly, indicating that tuning resolution alone does not account for behavioral discrimination performance. By computing the Fisher information from ICx neuronal responses, we showed that the average neuronal discriminability predicts behavioral thresholds across BC values. A habituation-based model incorporating BC-dependent changes in tuning width, firing rate, and response variability successfully accounted for both direction and ITD discrimination. These findings support a model in which perceptual acuity is governed by the combined influence of neuronal tuning and variability and provide a unified framework for understanding how midbrain auditory representations underlie adaptive spatial hearing.
Significance Statement
Determining the relationship between neural coding and perception is a major goal in neuroscience. We studied how barn owls discriminate interaural time differences (ITDs), a primary sound localization cue, when sensory reliability is degraded. Behavioral sensitivity declined sharply with reduced cue reliability, more than expected from changes in neural tuning resolution alone. Instead, behavioral thresholds align with a population-level measure of neural information that accounts for both tuning sharpness and response variability. A computational model suggests that discrimination performance arises from the interaction between neural habituation and degraded signal quality. These findings provide a mechanistic framework for understanding how the brain adapts to noisy environments by integrating reliability into sensory coding.
Introduction
How does the neural representation of sensory information influence the ability to discriminate changes in sensory stimuli? The information present in neuronal responses is impacted by factors including the selectivity, magnitude, and variability of responses to sensory signals (Seung and Sompolinsky, 1993; Abbott and Dayan, 1999; Pouget et al., 2003; Harper and McAlpine, 2004; Averbeck et al., 2006; Moreno-Bote et al., 2014). Determining how features of neural responses influence sensory discrimination abilities is essential for understanding how sensory processing translates into perception and behavior, a key question in neuroscience. Here we investigate the relationship between neural coding and sensory discrimination in the barn owl's sound localization system.
Previous work on sound localization in barn owls supports the hypothesis that the resolution of spatial coding in the midbrain auditory space map determines the owl's ability to discriminate sound source locations (Bala et al., 2003, 2007). Behavioral and neurophysiological studies investigating direction discrimination found a correspondence between the spatial resolution of spatially selective auditory neurons in the external nucleus of the inferior colliculus (ICx) and the minimum audible angle (Bala et al., 2003, 2007). While neuronal spatial resolution accounted for the difference in discrimination performance in azimuth and elevation, these studies did not manipulate the reliability of sensory information by adding noise to the stimulus. The reliability of sensory information is known to affect the ability to discriminate stimuli and influences multiple aspects of neuronal responses (Albeck and Konishi, 1995; Saberi et al., 1998; Peña and Konishi, 2004; Cazettes et al., 2016). Thus, how the reliability of a sound localization cue relates to neuronal and perceptual discrimination in barn owls remains an open question.
Barn owls primarily use the interaural time difference (ITD) to determine the azimuthal direction of a sound source (Moiseff, 1989; Poganiatz et al., 2001; Kettler et al., 2017). The reliability of ITD can be manipulated by changing the binaural correlation (BC), achieved by adding independent noise sounds to the ears (Saberi et al., 1998). Behavioral experiments with owls localizing sounds by turning the head show that localization accuracy decreases with BC (Saberi et al., 1998). Studies in humans show a similar decrease in localization and discrimination performance as BC decreases (Trahiotis et al., 2001; Rakerd and Hartmann, 2010).
Space-specific neurons are known to exhibit changes in ITD tuning such that the spatial resolution, peak firing rates, and membrane potential variability decrease as BC decreases (Albeck and Konishi, 1995; Saberi et al., 1998; Rich et al., 2015; Cazettes et al., 2016). Theoretical studies have suggested that changes in ITD tuning resolution can explain the decrease in the owl’s localization accuracy as BC decreases (Fischer and Peña, 2011; Rich et al., 2015; Cazettes et al., 2016). However, the relationship between neural coding and perceptual discrimination was unknown because neither the perceptual nor the neuronal discriminability of ITD has been tested as a function of BC.
To investigate this question, we used the pupil dilation response (PDR) to determine how the owl's ITD discrimination threshold changes with BC. The PDR allows measurement of the behavioral discrimination threshold without extensive training (Bala and Takahashi, 2000). We also determined neuronal ITD discrimination thresholds from previous recordings in the midbrain and characterized how ITD tuning and variability changed with BC (Cazettes et al., 2016). We show that neuronal ITD tuning resolution is not the sole predictor of perceptual discrimination performance. However, we find that perceptual discrimination is consistent with the average neuronal discrimination threshold in the auditory space map, which is influenced by tuning resolution, firing rate magnitude, and response variability. We show that modification of a previously proposed habituation model that explained the owl's discrimination in azimuth and elevation (Bala et al., 2003, 2007) also describes changes in discrimination of ITD as sensory reliability changes. These results extend previous studies and provide a more complete view of the relationship between neural coding and sensory discrimination in the barn owl's sound localization system.
Materials and Methods
Behavior: experimental design and statistical analyses
The subjects were five captive-bred barn owls of either sex. Headplates were surgically attached prior to experimentation, after which the owls were given at least 2 weeks to recover. Procedures were approved by the Institutional Animal Care and Use Committee of the University of Oregon.
Procedures for measuring the PDR were performed as previously described (Bala and Takahashi, 2000). Stimuli consisted of 100 ms noise bursts with 5 ms rise and fall times, presented at 50 dB of sound pressure level (SPL). Sounds were presented through earphones (Etymotic ER-2) placed in the ear canal, which were separately attenuated (Tucker-Davis Technologies PA-5). BC was varied by adding independent noise to the sound at one ear. The sound at that ear was a combination of the coherent noise carrying the ITD and the independent noise. The amplitudes of the independent noise and coherent noise were varied so that at BC = 1 the sound was only the coherent noise and at low BC the sound was dominated by the independent noise. The input was scaled to maintain a constant level of the stimulus. The BC is the measured cross-correlation between the sound signals delivered to the ears. After at least 100 presentations of the habituating stimuli at ITD = 0, test stimuli were presented at one of a few different ITDs. Test stimuli were repeated 4–6 times, separated by 40–50 presentations of the habituating stimuli.
Owls were wrapped in a tight-fitting jacket and laid in the stereotaxic device while head-fixed for sessions that lasted no longer than 90 min at a time. An infrared pupillometer was placed a few millimeters from one eye, which converted pupil size into voltage. Pupil size was measured for 1 s before and 2 s after sound onset. While the eyelid was held slightly open, the owls could still extend the nictitating membrane to moisten the cornea. Trials were discarded if the pupil was fully or partially obscured by the nictitating membrane during the 2 s gathering period. Custom MATLAB code was used to analyze PDRs. PDR magnitudes were calculated by integrating the voltage over 1 s after sound onset, relative to the mean voltage during 25 ms before and after sound onset. PDR magnitudes were then normalized into z-scores, enabling pooling across sessions varying BC within the same owl. The first 100 habituating trials were excluded from analysis, so that analysis only included fully habituated PDR trials. Data were analyzed using signal detection theory, which enabled us to convert pupil sizes into “hit” and “false alarm” rates, which were then used to construct receiver-operating characteristic (ROC) curves. The percent correct could then be calculated by integrating the hit rate over the false alarm rate (integrating the ROC curve) to determine discriminability, with a discrimination threshold set at 75% correct. This discriminability is the estimated just-noticeable difference (JND) from the PDR.
Neurophysiology
Neurophysiological data analyzed in this study were collected and partially analyzed in a previous study (Cazettes et al., 2016). Briefly, extracellular recordings of isolated ICx neurons were made from two adult female barn owls under ketamine and xylazine anesthesia. Experiments were performed in a double-walled sound-attenuating chamber. Sounds were delivered over calibrated earphones inserted in the ear canal. ITD tuning was measured by varying ITD in 10 µs steps over 10 repetitions using broadband noise signals (0.5–10 kHz) at BC values ranging from 0 to 1 in 0.1 steps. BC was varied by adding independent noise signals to the sounds at the left and right ears. The value of BC is given by
Neural discriminability
Behavioral ITD discriminability measured using PDR was compared with neural ITD discriminability from responses of single ICx neurons to ITD and BC. The Fisher information
We computed the linear Fisher information
Habituating population model
A habituating population model was used to examine how responses of midbrain space-specific neurons can lead to the behavioral ITD discrimination results of the PDR experiments. We used a version of the habituating population model that previously described results of PDR experiments determining the minimum audible angle in azimuth and elevation (Bala et al., 2003, 2007). The previous model was modified to include the nonuniform distribution of preferred azimuth and elevation that is observed in the midbrain auditory space map (Knudsen, 1982; Fischer and Peña, 2011, 2017) and to include model responses to ITD and BC.
The habituating population model proposes that the PDR response is mediated by a population of neurons that habituate to repeated input from midbrain space-specific neurons. The habituating neurons inducing PDR receive topographic input from the midbrain space-specific neurons. The midbrain space-specific population contains neurons with a range of preferred directions so that sound sources at any direction in the frontal hemisphere will be represented by a local pattern of activity. The responses of the habituating neurons to a novel stimulus are determined by comparing the responses of space-specific neurons to the novel stimulus to the mean responses of space-specific neurons to the habituating stimulus, relative to the standard deviation of responses of space-specific neurons to the habituating stimulus.
Specifically, let
To simulate midbrain space-specific responses to ITD at different BC values, the stimulus ITD was corrupted with Gaussian noise with a standard deviation that increased exponentially as BC decreased (Fischer and Peña, 2011) and the width of the Gaussian tuning response function increased as BC decreased (Cazettes et al., 2016). The responses of space-specific neurons (RSSN) on each trial were drawn from a Poisson distribution
Let
It is assumed that discrimination is based on the average activity in the habituating population,
Code accessibility
Python code to run the simulation is available at https://github.com/brian-fischer/ITD-Discrimination.
Results
Behavioral acuity
We first determined the barn owl's discriminability of ITD at different BC values. We used the PDR (Bala et al., 2003, 2007) to measure discriminability in five owls. The PDR is a robust measure of discriminability that does not rely on the owl turning its head. Furthermore, the owl's head-aim errors in a head pointing task (Knudsen et al., 1979) are strongly similar to the just-noticeable difference (JND) estimated with the PDR (Bala et al., 2003). Thus, henceforth, we refer to the PDR-estimated JND as simply the JND. Herein, the JND for perfectly correlated sounds (BC = 1) ranged from 6 to 14 µs, which is consistent with the previously measured discrimination threshold in azimuth of 3° (Bala et al., 2003). The dependence of the JND on BC was consistent across owls, where the owls showed a small increase in JND as BC decreased to 0.5 and then a large increase for BCs below 0.5 (Fig. 1). The dependence of the JND on BC was well described by an exponential curve in each of five owls tested (Fig. 1; r > 0.9 for each).
Interaural time difference discrimination thresholds. The behavioral just-noticeable difference (JND) in ITD estimated from the PDR was computed for a range of binaural correlation (BC) levels in five owls. For each owl the JND versus BC data were well-fit by an exponential curve (solid line; r > 0.9 for each).
Neuronal interaural time difference tuning
We then tested whether the behavioral JND is predicted by the spatial resolution of ITD responses of midbrain auditory space map neurons, as was found previously for azimuth and elevation (Bala et al., 2003, 2007). Previous studies demonstrated that the widths of ITD tuning curves of midbrain auditory space map neurons in ICx (Cazettes et al., 2016) and optic tectum (OT; Saberi et al., 1998) increase exponentially as BC decreases. While this exponential pattern is consistent with an exponential relationship between the behavioral ITD JND and BC, is the change in ITD tuning curve width proportional to the change in behavioral ITD JND? In contrast to the prediction of Bala et al. (2003, 2007), we found that decreasing BC caused a much greater increase in behavioral ITD JND than the increase in the widths of ITD tuning curves in ICx (Fig. 2). For example, at BC = 0.6, behavioral ITD JNDs were more than two times the JNDs for perfectly correlated sound (BC = 1), whereas ITD tuning widths of ICx neurons only increased by 1.06 times. In fact, average ITD tuning width of ICx neurons never reached twice the tuning width for BC = 1. Example cases of single ICx neurons where the ITD tuning width reached double the tuning width found for perfectly correlated sounds only occurred when BC was 0.2 or lower. At these lowest BCs the behavioral ITD JND was between 4 and 12 times the value found for perfectly correlated sounds (7.5 times on average). Thus, the ratio of behavioral ITD JND found at different BC values is not predicted by the ITD tuning widths of ICx neurons at those BCs.
Change in just-noticeable difference (JND) and neuronal resolution. The ratio of the behavioral ITD JND at each BC relative to the JND at BC = 1 for the five owls (grayscale circles). The mean ratio of the single-neuron ITD tuning curve widths for ICx neurons at each BC relative to BC = 1 (blue diamonds). The upper and lower blue lines show the first and third quartiles [adapted from Fig. 3D in Cazettes et al. (2016)]. The mean ratio of the OT multielectrode array (MEA) recorded population activity width at each BC relative to BC = 1 (red squares; from Ferger et al., 2021).
In addition to the analysis of single-neuron ITD tuning curves' width, we examined the tuning width of population activity in the auditory space map in OT at different BC values (Ferger et al., 2021). Ferger et al. (2021) estimated the width of population activity in OT from simultaneously recorded neurons using multielectrode arrays covering wide areas of the OT auditory space map. Consistent with single-neuron responses in ICx, the rate of increase in behavioral ITD JND as BC decreased greatly exceeded the rate of increase in the width of population activity in OT (Fig. 2). Thus, contrary to the observation that the ratio of the behavioral JND for elevation and azimuth (with BC = 1) matches the ratio of tuning curve widths for elevation and azimuth in ICx (Bala et al., 2007), the ratio of behavioral ITD JNDs at decreased BCs does not match that of the spatial resolution of ITD tuning in ICx and OT at these same decreased BCs. This difference between behavioral discriminability and neural ITD tuning resolution is expected because decreasing BC causes an increase in sensory noise and discriminability is dependent on both changes in the mean and the variability of neural responses (Pouget et al., 2003).
Neuronal interaural time difference discrimination
We further compared the behavioral ITD JND to the optimal discriminability of ITD from single ICx neurons (Cazettes et al., 2016). Perhaps, the behavioral ITD JND may correspond to optimal discrimination when both tuning resolution and variability are considered. We used the Fisher information computed from ICx responses to estimate the single-neuron JND (Seriès et al., 2009). The Fisher information was computed as
Interaural time difference (ITD) tuning and variability. A, B, Examples of the mean spike count response (black circles) of ICx neurons to ITD demonstrating the accurate fit with a Gaussian function (blue curve) at different BC values. The error bars represent the standard deviation of the spike count. C, The median over ITD of the relative absolute error in the fit of the Gaussian function to the mean spike count at each BC. The median relative absolute error did not vary significantly with BC (p = 0.11, Kruskal–Wallis, n = 46). D, The median over ITD of the Fano factor in ICx responses at each BC. The median Fano factor did not vary significantly with BC (p = 0.99, Kruskal–Wallis, n = 46). The box plots in C and D have boxes that show the first quartile, median, and third quartile and whiskers that extend to the minimum and maximum values that are not outliers. Outliers are points more than 1.5 times the interquartile range below the first quartile and above the third quartile. The strip plots show individual data points with random jitter in the azimuthal direction to aid visibility.
The responses shown in Figure 3A,B illustrate that the spike count variability changed with the mean spike count. We used the Fano factor, defined as the ratio of the variance to the mean of the spike count, to analyze spike count variability. We observed Fano factors below and above 1, indicating the presence of under- and overdispersion in ICx responses (Fig. 3D). The Fano factor did not vary significantly with BC (p = 0.99, Kruskal–Wallis, n = 46), and the overall median value was 1.0.
The Gaussian fits to the ITD tuning curve and the spike count variances were used to compute the Fisher information and then estimate the ITD discrimination performance of ICx neurons. The optimal ITD discriminability of single ICx neurons showed a range of values around the behavioral JND (Fig. 4). Similarly to previous results in azimuth and elevation with BC = 1 (Bala et al., 2003, 2007), there were ICx neurons with better or worse performance than the behavioral JND over the full range of BC values. Additionally, the median of the JNDs from ICx neuronal responses was consistent with the behavioral JND values. These results are consistent with previous investigations of the relationship between behavioral and neuronal stimulus discriminability showing that behavioral discrimination does not reach the optimal value supported by the most sensitive neurons but is consistent with the average performance of ICx neurons.
Neuronal discrimination. Comparison of ITD discrimination from behavior (grayscale circles from Fig. 1) and single-neuron responses in ICx. The boxplots show the distribution of single ICx neuron responses where box ranges from the first to third quartiles and the whiskers extend to points that are not outliers. The large blue diamonds are the median JND over all neurons.
Habituating population model of discrimination performance
We then tested whether the behavioral ITD discrimination across BC values can be explained by a habituating population model that has been used to explain the behavioral JND for azimuth and elevation at BC = 1 (Bala et al., 2003, 2007). The habituating population model proposes that the PDR reflects the average activity of a population of neurons that are driven by, and habituate to, activity from the auditory space map. The model assumes that the amount of habituation is higher for activity patterns that are less variable in time. So, habituating layer neurons will have the largest responses to input from the auditory space map that deviate from their recent inputs and when the recent inputs have shown little variability. With this construction, the habituating layer neural activity resembles the magnitude of the z-score of the current auditory space map activity, relative to the mean and standard deviation of previous auditory space map activity. It is assumed that the mean activity of the habituating population determines the JND. Once habituated, the owl will detect a difference in the stimulus when the mean activity of the habituating population crosses a fixed threshold.
We first determined the discrimination threshold so that the model produced responses consistent with behavioral discriminability in azimuth and elevation at BC = 1, replicating previous results (Bala et al., 2003, 2007). Here, the model space-specific neurons have tuning curve half widths in elevation (41°) that are approximately twice the tuning curve half widths in azimuth (20°; Fig. 5A). Correspondingly, the JND in elevation (8.2°) is approximately twice the JND in azimuth (3.9°; Fig. 5B,C).
Habituating population model. A, Example tuning curves (mean and standard deviation) of model space map neurons to azimuth (red) and elevation (blue). B, Discriminability statistic from the habituating population as a function of the angular change in the stimulus for azimuth (red) and elevation (blue). The discrimination threshold (black) was selected to reproduce previously reported results, BC = 1 (Bala et al., 2007). C, Model JND in azimuth and elevation using the discrimination threshold shown in B. D, Example ITD tuning curves of an ICx neuron and (E) a model space map neuron for different BC values. Mean spike count responses were fit by Gaussian curves (blue). F, Discriminability statistic from the habituating population as a function of the change in ITD for BC values ranging from low BC (BC = 0.15; light blue) to high BC (BC = 1; dark blue) in steps of 0.05. The discrimination threshold is shown in the black line. G, Comparison of ITD discrimination from behavior (grayscale circles) and the habituating population model (blue squares). H, Comparison of ITD discrimination from behavior (grayscale circles) and the habituating population model when the firing rate of model neurons is constant at the value for BC = 1 (blue squares).
We modified the habituating population model of Bala et al. (2003, 2007) to allow for model space-specific neurons to respond to ITD and BC. Model space-specific neurons were driven by a stimulus ITD that was corrupted by random Gaussian noise with a standard deviation that increased exponentially as BC decreased (Fischer and Peña, 2011). The tuning widths increased and firing rates decreased as BC decreased, consistent with experimental observations (Fig. 5D,E; Albeck and Konishi, 1995; Saberi et al., 1998; Cazettes et al., 2016; Ferger et al., 2021). Model space-specific neuron spike counts had a Poisson distribution, consistent with the observation that the median Fano factor for ICx neurons was one (Fig. 3D).
We tested the model's discrimination of ITD at different BC values using the discrimination threshold found (at BC = 1) so that the model produced responses consistent with behavioral discriminability in azimuth and elevation. This model was consistent with the owls' dependence of JND on BC (Fig. 5F,G). The consistency of the model with behavior under decreasing BC depends on a decrease in space map firing rates and thus a decrease in spike count variability. When firing rates of model space map neurons were held constant at their maximum value (BC = 1), the model performance exceeded the owl's discrimination performance, decreasing more slowly with BC (Fig. 5H). The model shows that a habituating population model can explain the owl's discrimination in azimuth, elevation, and ITD at different BCs. Moreover, these results show that both response resolution and variability of midbrain auditory space map neuronal responses can determine the owl's discrimination of sound location.
Discussion
ITD reliability describes how corruptible the ITD cue for sound source localization is when faced with concurrent sounds (Cazettes et al., 2014). In the barn owl, ITD reliability derives from the responses of interaural phase difference-tuned neurons in the inferior colliculus and their target ITD-tuned neurons in the ICx (Cazettes et al., 2014). We studied the relationship between ITD-sensitive neuronal responses and behavioral ITD discrimination while presenting varying levels of binaural correlation, which directly changed ITD reliability (Licklider, 1948; Jeffress et al., 1962). We show that perceptual sensitivity to ITD degrades exponentially as binaural correlation decreases. This decline in behavioral performance, however, could not be accounted for by concomitant changes in neuronal tuning width alone. ITD tuning broadened in ICx neurons with decreasing BC; however, the rate of broadening was insufficient to explain the exponential increase in behavioral JNDs. Behavioral JNDs were more closely aligned with the average ITD discriminability predicted by Fisher information from single-unit ICx responses, thus integrating tuning resolution, response magnitude, and response variability. This suggests that perceptual acuity reflects the average information content across the neural population, rather than the optimal performance of the most sensitive neurons.
A habituating population model based on that of Bala et al. (2003, 2007) provided additional support for this interpretation. Only after incorporating empirically observed changes in tuning width, firing rate, and response variability did the model accurately capture behavioral ITD discrimination across BC values. In contrast, when model firing rates were held constant, performance deviated from behavioral data, underscoring the importance of reduced response magnitude and increased variability limiting sensory discriminability in degraded sensory environments. Additionally, decreasing binaural correlation resulted in linked decreases in both mean neuronal spike count and spike count variability, with a median Fano factor near one, suggesting that regulation of spike count variability is important for behavioral discrimination.
This work advances our understanding of the relationship between neural coding and perceptual stimulus discrimination. We provide additional support to the previous conclusion that the response properties of midbrain auditory space maps in ICx and OT determine the ability to discriminate changes in sound location by incorporating the influence of changes in sensory reliability in our understanding of how midbrain auditory space maps in ICx and OT determine the ability to discriminate changes in sound location. We showed that a habituating population model can explain how ICx responses to azimuth, elevation, and ITD as BC changes are translated into perceptual stimulus discrimination limits. This provides a more comprehensive understanding of how neural coding in the midbrain auditory space map is related to sound location perception.
Footnotes
This work was funded by National Institute on Deafness and Other Communication Disorders (R01DC007690), National Institute of Neurological Disorders and Stroke (R01NS132812-01), and CRCNS-US-Israel R01NS135851 grants.
The authors declare no competing financial interests.
- Correspondence should be addressed to Brian J. Fischer at fischer9{at}seattleu.edu.
