| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
Center for Hearing and Balance, Department of Biomedical Engineering, Johns Hopkins School of Medicine, Baltimore, Maryland
Submitted 10 March 2005; accepted in final form 8 September 2005
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMETHODSRESULTSDISCUSSIONGRANTSACKNOWLEDGMENTSREFERENCES |
|---|
100 noise-like stimuli with pseudorandom spectral levels in 1/8- or 1/16-octave frequency bins. The model assumes that rate is determined by a linear weighting of the spectrum plus a second-order weighting of all pairs of spectrum values within a certain frequency range of best frequency. The method allows prediction of rate responses to stimuli with arbitrary wideband spectral shapes, thus providing a direct test of the degree of linearity of spectral processing Auditory-nerve fibers are shown to rely mainly on linear weighting of the stimulus spectrum; however, significant second-order terms are present and are important in predicting responses to random-spectrum shape stimuli, although not for predicting responses to noise filtered with cat head-related transfer functions. The second-order terms weight the products of levels at identical frequencies positively and the products of different frequencies negatively. As such, they model both curvature in the rate versus level function and suppressive interactions between different frequency components. The first- and second-order characterizations derived in this method provide a measure of higher-order nonlinearities in neurons, albeit without providing information about temporal characteristics. |
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTSDISCUSSIONGRANTSACKNOWLEDGMENTSREFERENCES |
|---|
; Deng and Geisler 1987; Sumner et al. 2003; Zhang et al. 2001). Models of synaptic processing have extended these results into the cochlear nucleus (Hancock and Voigt 1999; Hewitt et al. 1992; Kalluri and Delgutte 2003; Rothman and Manis 2003; Rothman et al. 1993). However, such models depend on a detailed knowledge of the circuitry and the physiological properties of the neurons; it seems unlikely that similar models will be developed for higher levels of the auditory system because of the complexity of the circuits. Instead it seems more feasible to rely on "black-box" models of signal processing, using nonlinear methods derived from Wiener and Volterra analysis (Eggermont 1993; Escabi and Read 2003; Klein et al. 2000).
The first Wiener kernel, or reverse-correlation function (de Boer and Kuyper 1968), captures the tuning of AN fibers with low best frequencies (BFs) and accurately predicts general aspects of their responses (Carney and Yin 1988; de Boer and de Jongh 1978; Møller 1977; Temchin et al. 2005). For fibers with higher BFs and neurons in the CNS that do not phase lock to the stimulus waveform, the characterization must be based on higher-order representations, like the second Wiener kernel (Lewis et al. 2002; van Dijk and Wit 1994; Recio-Spinoso et al. 2005; Yamada and Lewis 1999), the spectrotemporal receptive field (STRF) (Aertsen et al. 1981; Kim and Young 1994; van Gisbergen et al. 1975), or a method based on second-order distortion products (van der Heijden and Joris 2003). The best studied of these is the STRF, which describes the average stimulus power spectrum as a function of time preceding spikes (Aertsen and Johannesma 1981; Eggermont 1993; Klein et al. 2000; Theunissen et al. 2001) and is directly related to the second Wiener kernel (Aertsen and Johannesma 1981; Klein et al. 2000; Lewis and van Dijk 2004). When studied with STRFs, AN fibers show tuning consistent with conventional tuning curves and inhibitory regions attributable to suppression and refractory effects (Kim and Young 1994; Lewis and van Dijk 2004).
The applicability of these characterizations is mainly limited by the degree of nonlinearity of auditory neurons; essentially it is impractical to estimate nonlinearities of a sufficiently high order to capture accurately the properties of auditory neurons, especially in the CNS (Johnson 1980). As a result, STRFs of neurons in the central auditory system change significantly according to the stimulus used to derive them (Escabi and Schreiner 2002; Nelken et al. 1997; Theunissen et al. 2000) and do not accurately predict responses to arbitrary stimuli (e.g., Eggermont et al. 1983; Machens et al. 2004; Nelken et al. 1997; Sen et al. 2001; Versnel and Shamma 1998; Yamada and Lewis 1999; Yeshurun et al. 1989; Yu and Young 2000).
One approach to the problem of nonlinearity is to separate the analyses of frequency selectivity and temporal (or envelope) characteristics. This is not generally possible for arbitrary systems, but STRFs of neurons in the AN and inferior colliculus are usually separable into independent time and frequency functions (Lewis and van Dijk 2004; Qiu et al. 2003), although not in all cases and not in auditory cortex (Depireux et al. 2001; Eggermont et al. 1981; Sen et al. 2001). In fact, models of lower auditory processing commonly separate the frequency tuning from the temporal processing of stimulus modulation (e.g., Dau et al. 1997; Delgutte et al. 1998). Of course, spectral processing (tuning) implies a corresponding temporal response function, but the representation of the auditory stimulus envelope is not determined by the frequency tuning, even in the AN (Joris and Yin 1992). Thus it seems reasonable to simplify the task of characterizing lower-level auditory neurons by analyzing spectral and temporal processing separately.
Here we apply the method of spectral weight functions to AN fibers (Yu and Young 2000). Weight functions are equivalent to the STRF averaged over its time axis and so are a measure of the frequency-dependent component of a separable STRF. By ignoring temporal features of the stimulus/response relationship, the analysis can be extended to higher orders of nonlinearity than is feasible with the full STRF. We show here that both first- and second-order weight functions can be robustly estimated; these measure the gain of the fiber to stimulus energy in different frequency bands (1st order) or pairs of bands (2nd order). An advantage of weight functions is that they can be tested by attempting to predict the rate responses to wideband stimuli with various spectral shapes, thus providing a rigorous test of the completeness and accuracy of the model (Barbour and Wang 2003; Yu and Young 2000). The model accurately predicts the rate responses of AN fibers to both random-spectrum shape (RSS) stimuli and to noise signals filtered by head-related transfer functions (HRTFs) (Musicant et al. 1990; Rice et al. 1992). Second-order weights are necessary for accurate prediction of responses to RSS stimuli but not for responses to HRTF stimuli.
|
METHODS |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTSDISCUSSIONGRANTSACKNOWLEDGMENTSREFERENCES |
|---|
The data were obtained from nine experiments in healthy adult cats (3–4 kg). Cats were anesthetized with ketamine (40 mg/kg im) and xylazine (2 mg im). An intravenous line and tracheal tube were inserted. The animal was maintained at a surgical plane of anesthesia throughout the experiment with pentobarbital sodium (5–15 mg/h iv as needed to prevent withdrawal reflexes). Atropine (0.1 mg im) was given to control secretions. The bulla was vented with a length of PE-90 tubing. The cat’s temperature was maintained at 38.5°C, and intravenous fluids were given. For recording, the posterior fossa was opened and the cerebellum retracted to expose the AN. Glass micropipettes filled with 3 M NaCl (10–30 M
) were used to record from single AN fibers. The procedures were approved by the Johns Hopkins Animal Care and Use Committee.
Single fiber recording was done in a sound-proofed (IAC) chamber. The acoustic system consisted of an electrostatic or a dynamic driver coupled to the ear through a 12.5-cm earbar. The system was calibrated in situ prior to the experiment using a probe tube placed within 2 mm of the eardrum. Typical calibrations have been shown previously (Rice et al. 1995). The electrostatic driver has a bandwidth of 30–40 kHz, whereas the dynamic driver provides significant energy only 
10 kHz. The dynamic driver was used to provide stimuli at high sound levels in five experiments; only fibers with BFs <10 kHz were studied in those experiments.
When a fiber was isolated, its BF and threshold were determined using an automated tuning curve algorithm. A series of RSS stimuli, described in the following text, were then presented. The stimuli were 400 ms in duration and were presented once per second across a range of sound levels in 10-dB steps. In one experiment, the same set of stimuli was repeated multiple times, but usually each stimulus was presented only once. The spontaneous rate (SR) was determined as the average rate during the last 400 ms of the interstimulus interval for the lowest-level RSS presentation. The rate responses to the RSS stimuli were used to construct the weight-function model described in the following text.
In one experiment, a test stimulus set consisting of a sample of noise (bandwidth: 10 Hz to 50 kHz) filtered with cat HRTFs (from Rice et al. 1992) was presented. Stimuli filtered with 99 different HRTFs were presented along with the unfiltered noise, and the resulting rate responses were used to test the ability of the weight-function model to predict responses to stimuli other than those used to generate the model. HRTF stimuli were also 400 ms in duration and were presented across a range of sound levels in 10-dB steps; the sound levels were adjusted to cover a similar range as the RSS stimuli, although it is not possible to exactly match levels. The sampling rate of the HRTF stimuli was adjusted to shift the spectra nominally at 10 kHz to the BF of the fiber. For example, a fiber with BF 10 kHz would receive HRTF stimuli with sampling rate 100 kHz (the nominal value), but a fiber with BF 5 kHz would receive the same stimuli with sampling rate 50 kHz. This manipulation guarantees that the spectral shapes and sound levels near the fiber’s BF vary maximally across stimuli.
Weight function model
The analyses in this paper are based on the following model for the average discharge rate r of a fiber over the duration of the stimulus (Yu and Young 2000)
 | (1) |
For all calculations using Eq. 1, the dB levels of the stimulus S(fi) were corrected for the acoustic calibration by setting the 0-dB reference level for the stimuli to the SPL of the frequency bin at BF [with S(fBF) = 0] and then correcting the values in other bins according to the calibration C(f) [C(f) is the sound level of a tone of frequency f at a fixed rms voltage signal into the speaker]. In other words, the decibel values of the stimulus S(f) were corrected by adding C(f) – C(fBF), the difference between the calibration at f and the calibration at BF. In practice, this adjustment has only a small effect.
The parameters of the weight-function model were fit to data consisting of the rate responses to a set of 95–99 RSS stimuli, described in the next section. Each of the stimuli has a different spectrum, a different set of S(fi). The rates in response to the RSS stimuli were measured and the parameters R0, {wi|i = m1...m2}, and {wij|i = m1...m2, j = i...m2} were estimated by minimizing the error between the actual rates and the rates predicted by Eq. 1. The error E is the 
2 difference between the model and the rates divided by the number of degrees of freedom (Press et al. 1986)
 | (2) |
k is the actual rate in response to the kth stimulus
and r( )
is the rate predicted by Eq. 1 for the kth stimulus. The variances 
k2 should be the variances of the rates k. Because multiple repetitions of the stimuli were usually not obtained, estimates of the rate variances are not available; therefore Poisson variances were assumed, k2 = k/T, where T is the duration of the stimulus presentation. To avoid dividing by zero in Eq. 2, k2 was not allowed to be <1, corresponding to 0.4 spikes/s. df is the number of degrees of freedom of the estimate. When M parameters are being estimated, df = K –M, where K is the number of stimuli. The degrees of freedom is important in comparing fits of first- and second-order models. Because the second-order model has more parameters, it is bound to fit better; however, dividing by df partially compensates for that effect. If the model fit the data exactly except for additive Gaussian noise in the rates (i.e., k = r( ) + nk where nk has mean 0 and variance k2), then E in Eq. 2 would have mean value 1 and SD 
2/df. Estimation of the weights by minimization of error E is done by the method of normal equations (Press et al. 1986).
The SDs of the weight estimates were estimated using bootstrap, in which the calculation of weights was repeated for sets of 100 stimulus/response pairs chosen at random with replacement from the full set (Efron and Tibshirani 1993). Two-hundred bootstrapped weight sets were computed, and the mean ± SD of the weight estimates were calculated from those results. The bootstrap calculation gives an elevated estimate of SD for this problem, 
2.5 times larger than direct estimates of SD calculated in seven cases where the RSS stimulus sets were presented multiple times. For this reason, the bootstrap SDs are used here only for illustration and not to qualify weights for inclusion in the model.
Simulations show that this estimation method usually gives unbiased estimates of the weights. However, two sources of error were found. One source of bias in R0 and error in the weights is rate clipping. Stimuli that drive the discharge rate to zero would produce negative rates if that were possible. Because the model does not include rectification at zero rate, the estimate of R0 increases to better fit the clipped rates and the second-order weights are also distorted. In AN fibers, few rates are actually driven to zero, and so clipping is not a major source of error in these data. Compensation for clipping can be done by deleting zero-rate cases from the data before fitting; however, this was not necessary here.
A second source of bias in R0 is estimating fewer weights than are actually needed, i.e., making m1 too large or m2 too small in Eq. 1. If an important second-order weight is deleted from the diagonal (where fi = fj), then the fitting process changes R0 to compensate and better fit the overall average rate. The first- and second-order weights are still estimated without bias in this case. Low-BF fibers have the widest (along the log-frequency axis) weight functions, and it was sometimes difficult to include enough second-order weights in their models because the RSS stimulus sets were not large enough; this effect accounts for the fact that R0 estimates for low-BF neurons were consistently biased by a few spikes/s, which is comparable to the SD of the R0 estimates. Because of this bias, low-BF neurons (<3 kHz) are not shown in Fig. 7.
|
K/2). The number of second-order weights is proportional to the square of the bandwidth of the weight set, so the second-order weight limits were more restrictive than the first-order limits; however, the second-order weight range usually included the range over which first-order weights were significantly different from zero, from the bootstrap calculation, the major exception being low-BF neurons. In the section of RESULTS on prediction of responses (Fig. 8) a method for further refining the range m1 to m2 is described.
|
The RSS stimuli are designed to estimate the parameters of Eq. 1. Each stimulus consists of a sum of tones equally spaced logarithmically at 1/64 octave. Tones of successive frequencies are gathered into groups of either four or eight tones to form frequency bins of width 1/16 or 1/8 octave, respectively. All the tones within a bin have the same amplitude, and the energy of a bin S(f) is the decibel value corresponding to the sum of the energies of the tones contained in the bin. Because the frequencies are logarithmically spaced, the overall stimulus is not periodic and sounds noise-like. For the same reason, the overall spectrum of the stimuli has a 1/f shape, e.g., for a stimulus with 0 dB in all bins. To avoid a click at time 0, the starting phases of the tones were randomized.
The total bandwidths of the stimuli were 1, 2.13, or 8.5 octaves; the 1-octave stimuli were used in five experiments where high sound levels were studied (narrow bandwidth was needed to maximize the sound level). When 1- and 2.13-octave stimuli were used, the sampling rate of the stimulus was adjusted to place the BF of the fiber 2/3 of the way between the lowest bin frequency and the highest.
The stimulus levels S(fi) were chosen as follows. The RSS stimulus set consists of 100 stimuli. Consider a matrix S whose columns are the dB values of individual stimuli
 | (3) |
One stimulus was generated for each column in Eq. 3. In practice, between one and five columns were set to 0 dB to provide an independent estimate of R0, as will be discussed in Fig. 7. Thus the actual RSS stimulus set consisted of 95–99 stimuli. Each stimulus was generated by adding sinusoids, in random phase, with the amplitudes set by converting the decibel values in S to linear values (as 
). Stimuli were generated to have duration 400 ms at a nominal sampling rate of 100 kHz. When the sampling rate was changed to position the spectrum relative to a fiber’s BF, stimuli were extended in time, if necessary, by making them periodic. Although the periodic stimuli have a discontinuity (click) at each period, the clicks were not audible and the fibers did not show click responses. For the 8.5-octave-wide stimuli, frequency bin centers ranged from 100 Hz to 38 kHz and the sampling rate was fixed at 100 kHz. For the variable-sampling rate stimuli, the sampling rate was always at least four times the highest frequency component in the stimulus.
|
RESULTS |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTSDISCUSSIONGRANTSACKNOWLEDGMENTSREFERENCES |
|---|
Weight functions of AN fibers show an excitatory area (positive weights) near BF that may or may not be accompanied by an inhibitory area (negative weights) at lower or higher frequencies. The first-order weight functions for the example unit in Fig. 1, B and C, were taken at sound levels separated by 20 dB. In both cases, there is a peak weight near 3 spikes/(s dB) at BF with smaller weights at adjacent frequencies. In this case, the frequency of the center of the RSS bin nearest BF (i.e., the tuning curve BF) was slightly below BF, so the peak weight, which is plotted at the bin center, appears to be below BF. The second-order weights show positive (green) values on the diagonal of the weight function, i.e., for terms like wBF,BFS(fBF)2; such terms fit the curvature of the rate response to stimulus level. Negative weights (red) appear off the diagonal. The largest negative and positive weight are usually at or near BF.
|
points. The legends in these figures give the values of error E (Eq. 2) for the fits. As is typical, E is smaller with the second-order terms.
The pattern of weights in Fig. 1, B and C, is typical of AN fibers. In this case, the stimuli had frequency bins 1/8 octave wide and the BF of the neuron was high (16.3 kHz), so significant weights only occupy one or two bins near BF. At lower BFs, the weight functions are broader, on logarithmic axes, consistent with the behavior of AN tuning curves (Evans 1977; Liberman 1978). Figure 2 shows an example for a fiber with BF 1.4 kHz. In addition to being broader, the first-order weight function has a smaller peak amplitude, 1.5 spikes/(s dB). The second-order weights show the same general patterns as in Fig. 1.
|
|
). First, the first-order weights provide a best linear fit to the rate-level function, most clearly seen for the highest two levels in Fig. 3B. Second, the second-order weights on the diagonal of the weight matrix (like wBF,BF) model departures of the rate response from linearity. This is clearly illustrated at the lowest reference level in Fig. 3A (
), where about half the data points are below threshold and the rate response has a clear upward curvature. Third, the RSS stimuli do not recapitulate the static rate-versus-level function of the fiber. Instead, there is an adjustment of the dynamic range of the fiber at each reference level. The adjustment can be seen by comparing the points for the 19 and 39 dB reference levels in Fig. 3A (x and 
). For a particular sound level on the abscissa, the points for the 39 dB reference level are systematically below those for the 19 dB level (except for 2 points). Presumably, this behavior reflects cochlear suppression. The dynamic range of the rate response is preserved, to some extent, at higher reference levels by suppressing responses to tones at all frequencies. Consistent with this hypothesis, the suppression-like effect is apparent in the low-SR fiber in Fig. 3A and not in the high-SR fiber in Fig. 3B. The large second-order terms near threshold are produced by the rectification of rate at zero in Fig. 3A and at SR in Fig. 3B. These second-order terms are not the focus of later sections on second-order weights. Instead, it is the smaller second-order weights illustrated by the curves at 11 and 31 dB in Fig. 3B, where the fiber is well away from threshold. Notice that rates do not show a hard rectification at high sound levels as they do at threshold. The data at the highest levels in Fig. 3 are typical in that saturation occurs as a decrease in the slope of the rate functions at high levels and not as a true saturation with negative curvature.
First-order weights at various sound levels
First-order weight functions are qualitatively the same in all AN fibers. The major quantitative trends are illustrated in Figs. 1 and 2: weight functions broaden, on a logarithmic frequency axis, and have lower amplitudes at lower BFs. There are three additional trends: first, weight functions saturate, meaning a decrease in weight amplitude, at higher reference sound levels; second, the frequency of the peak weight shifts toward lower frequencies at very high sound levels; third, at higher sound levels there is a consistent small inhibitory weight at frequencies above BF.
The trends in weight functions were studied by averaging weights across fibers, after gathering fibers into three groups by BF: 1–3, 3–10, and 10–30 kHz. The similarity of weights across fibers is illustrated in Fig. 4, which shows weight functions for individual medium-SR fibers (gray), along with their averages (black lines with unfilled circles). The weight functions were aligned on their tuning-curve BFs for averaging. The center frequency of the bin containing the largest weight is always near BF (except at very high sound levels) but is not always exactly at BF. However, the trends discussed in this section do not change if weight functions are aligned on the peak weight instead of the tuning curve BF. The weights plotted here were estimated from responses to RSS stimuli with 1/16-octave bins, providing better frequency resolution than in Figs. 1 and 2. The tendency of weight functions to be lower-amplitude but broader in extent at lower BFs is clear in Fig. 4.
|
). Data are available at levels up to 80 dB SPL/component; at higher levels, the mean weights continue to decrease in amplitude toward zero (not shown).
|
Finally, significant negative weights were often seen at frequencies above BF in individual fibers (e.g., Fig 1C); such negative weights were consistently seen in the averages, in 15/23 cases in Fig. 5, mostly at high sound levels (>10 dB/component). These are often statistically significantly different from zero. Similar inhibitory regions have been reported in STRFs (Kim and Young 1994; Lewis and van Dijk 2004).
The peak sensitivity to stimulus intensity is found in high SR fibers at low sound levels but shifts to medium and low SR fibers at higher levels (Fig. 6). The shift is clearest for the high BF fibers (Fig. 6A) but is evident for lower BFs as well (Fig. 6B). The weight at BF increases rapidly at low sound levels, peaks at a level that depends on the SR group, and then declines toward zero. Of course, at the highest levels the largest weight is below BF, so it is not plotted in Fig. 6; however, the decline of peak weight at high levels occurs regardless of which weight is plotted. Weights are significantly smaller in the low BF group (1–3 kHz, gray lines in Fig. 6B) than for the two higher BF groups.
|
Nonlinearity of weight functions
Second-order weights are necessary only if they improve fits to data and predictions of responses to other stimuli. However, there is a simple argument that the second-order terms are needed. In the model of Eq. 1, the rate response to the all-0-dB stimuli [i.e., those with S(f) = 0 for all f] is R0. If the second-order terms are not important and can be dropped from Eq. 1, then the average of the rates across the entire RSS stimulus set is also R0, because of the way the RSS stimuli were generated. To see this, consider the average rate 
rk
predicted by a model containing first-order terms only
 | (4) |
rk should equal R0 if the second-order terms in the model can be ignored.
Figure 7 shows a comparison of the rates in response to the all-0-dB stimuli (abscissa) to the overall average rates (x) and the estimates of R0 from the second-order model (
). The latter are close to the all-0 dB rates, as expected. However the average rates are systematically larger than both R0 and the rates to the all-0-dB stimuli. The differences are statistically significant (see the caption) and are caused by the contribution of the second-order terms. For this analysis, only fibers with BFs from 3 to 30 kHz were used; lower BF fibers yield a biased estimate of R0 as discussed in METHODS. Fibers with rates near threshold or saturation were also excluded (see caption) to eliminate nonlinearities from those causes. The result does not change qualitatively if low BF fibers and fibers at threshold and saturation are included.
In addition to the argument of Fig. 7, the importance of second-order terms is suggested by the quality of fits to the responses to RSS stimuli. Figure 8 shows the results of a cross-validation test in which the RSS stimulus set was divided into two groups: the first 60 stimuli were used to derive a weight-function model and the remaining 40 stimuli were used to test the adequacy of the model. Testing with stimuli that were not included in calculating the model provides a test of the ability of the model to generalize and allows identification of model overfitting, i.e., adjusting model parameters to fit noise in the data. The frequency range covered by the weights included in the model (i.e., the range m1 to m2 in Eq. 1) was chosen using the algorithm illustrated in Fig. 8, A and B. Beginning with only one weight at BF, weights were added alternately on the low- and high-frequency side of BF. The number of second-order weights was also increased, by using a square of second-order weights the sides of which cover the same frequency range as the first-order weights. Both first- and second-order models were computed for each number of weights and the errors of the fits (Eq. 2) were plotted as in Fig. 8B. The left plot shows errors in the fit to the model data for first- and second-order models; these are typical in that the error is smaller for the second-order model and both fits improve over the first few weights and then saturate, usually with a shallow local minimum around four first-order weights (arrow). The right plot shows errors for the test data (i.e., for the 40 RSS stimuli not included in computing the model). Here the error is also smaller for the second-order model, but the error of the second-order model increases rapidly beyond three first-order weights (typically 3–6 weights). The increase suggests that the second-order model is overfitting the data with the larger number of weights.
For each set of responses, the number of weights included was set by the first local minimum in the error function for the second-order model fit (the arrow in Fig. 8B, left) or, if there was no local minimum, by the smallest number of weights giving an error within 20% of the overall minimum. This method usually yields an error for the test data (Fig. 8B, right) near the minimum error, so models chosen in this way are called "best-fitting" models in the following text. Figure 8A compares the best-fitting model to a model computed with a wider bandwidth. As is typical, the differences between the weights in the two models are small. Best-fitting models had between three and five first-order weights, for RSS stimuli with 1/8-octave bins.
The second-order best-fitting models provided a clearly better fit to the test data, compared with the first-order models. Figure 8C shows a comparison of error E for first- and second-order models, meaning best-fitting models evaluated with test data. The second-order models have less error in the majority of cases and the difference is significant (P 0, see legend).
The quality of the fit was also better using fraction of variance (fv) as the error measure, where fv is defined as
 | (5) |
k is the rate in response to the kth stimulus, rk(Sk) is the value predicted by the model, and 
is the mean value of k. fv has a maximum of 1 when the model fit is perfect and decreases as the error increases; it can go negative for large error, where the mean rate is a better fit than the model. The fraction of variance captured by the second-order model is significantly better than that captured by the first-order model (Fig. 8D, see legend for statistics). However, the quality of the fit varies. There are a number of cases of poor fits in Fig. 8D (fv values <0.2). These cases are mostly responses at high sound levels where the rates are close to saturation. In this case, the noise in the rate estimates is a significant portion of the rate variation and the model cannot fit this random component.
In practice, the maximum likely value of fv when fitting test data is [1 – 
n2/(r2 + n2)], where n2 is the variance in estimating the rates (i.e., the variance of the noise in the k) and r2 is the variance of the rates (the denominator sum in Eq. 5, divided by K – 1, assuming no noise in the rates). As a test of the noise limit in these data, fv was corrected by subtracting an estimate of n2 from both numerator and denominator in Eq. 5 (using the Poisson estimate k/T for n2). With the correction, the peak in the distribution of fv for the second-order model (i.e., the peak at 0.6–0.8 in Fig. 8D) is centered on 1. Thus evaluating the fits of the second-order model appears to be limited by the noise in the data for fv values around 0.6–0.8.
Properties of second-order weight functions
Second-order weights have a particular shape that is illustrated by the examples in Figs. 1, 2, and 8 and the averaged second-order weights in Fig. 9. Weights are positive on the diagonal of the weight matrix (green squares), i.e., for terms like wBF,BFS(fBF)2. The positive diagonal is surrounded by negative weights (red squares) for terms involving the product of two different frequencies. This shape is observed in all BF ranges and SR groups. As sound level changes, the behavior shown in Fig. 9 is typical. At low levels, positive weights on the diagonal dominate. At higher levels, the negative weights off-diagonal increase in magnitude. At very high levels, the weights become smaller, like the first-order weights, reflecting saturation.
|
1/16 octave in width. The actual shape of the second-order weights near the diagonal is probably not adequately revealed in the present data. Prediction of rate responses to HRTF spectral shapes
The responses to HRTF stimuli were used to test the ability of the weight-function model to predict response rates. These stimuli were computed by filtering a sample of broadband noise with transfer functions of the cat head and ear for sound sources at varying azimuth and elevation (Musicant et al. 1990; Rice et al. 1992). HRTF spectra are different from the RSS spectra in that they are smoother in shape except at spectral notches. HRTF and RSS spectra do have the common feature of a general low-pass shape, so that the average overall spectral levels behave roughly as 1/f in both stimulus sets.
For the weight-function modeling, the energy in the HRTF spectra was gathered into frequency bins corresponding to the RSS bins. The resulting spectra were corrected for the acoustic calibration and expressed as dB relative to the reference level for the weight-function model. In general the RSS and HRTF stimuli were presented at a number of different reference sound levels. For each HRTF data set, a weight function model was computed from the two RSS data sets with reference levels nearest (within 10 dB) the average HRTF sound level at the fiber’s BF. The reference level for the whole calculation was the reference level of the lower-level RSS stimulus set. In some cases, only one RSS set within the 10-dB window was available. The weight-function model was computed using the number of weights determined by the analysis of Fig. 8. The model was then applied to compute the discharge rates for the HRTF stimuli. In general, the two-level models did a better job of predicting HRTF responses than the one-level models (not shown).
The models did a good job of predicting rate responses to HRTF stimuli (Fig. 10). The examples in Fig. 10, A and B, show comparisons of predicted and actual rates for two fibers with good fits; results are shown for both the first- and second-order models, identified in the legend. In all cases, the points lie near the line of equality between actual and model rates. Unlike the result with RSS stimuli, the first- and second-order models performed about equally well, although there was a slight advantage for the second-order models, depending on how the error was measured. Figure 10C shows the chi-square error E (Eq. 2) for the first- versus second-order models. The second-order models are significantly better (see legend). However, the second-order models were not significantly better when evaluated using fv (Fig. 10D). The fv values for both fits in Fig. 10D are comparable to the results for the second-order model with RSS test stimuli in Fig. 8D. Again, the fv values that cluster around 0.6–0.8 are at the maximum allowed by the variance of the data. For these fibers, the second-order model cannot produce an improvement over the first-order model.
|
|
DISCUSSION |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTSDISCUSSIONGRANTSACKNOWLEDGMENTSREFERENCES |
|---|
These results show that AN fibers can be approximated to have a nearly linear dependence of discharge rate on the spectral content of a broadband stimulus. "Nearly linear" means that a first-order weighting of the dB spectrum captures much of the rate variation, but a small second-order correction can improve the fit (Figs. 8 and 10). The second-order correction consists of excitatory and inhibitory terms. The excitatory (diagonal) terms provide a quadratic model of the curvature of rate functions away from a strictly linear dependence on stimulus level. The inhibitory (off diagonal) terms model the inhibitory interactions of differing frequencies. The negative coefficients of the second-order weights vary with sound level (Fig. 9) and presumably reflect cochlear suppressive interactions (Delgutte 1990; Sachs and Abbas 1976; Sachs and Young 1980; Wong et al. 1998).
The near-linearity of AN weight functions provides an explanation for previous results in which the rate responses to vowel stimuli were studied (Le Prell et al. 1996; May et al. 1996, 1998). In those data, discharge rate was approximately linearly proportional to the dB level of the frequency component nearest BF. This behavior is expected if the first-order weights account for most of the rate response and if the largest weight is at BF. Moreover, the levels of adjacent stimulus components vary slowly with frequency in vowel spectra, so the effects of summing over a range of frequency components near BF would produce a small effect and would not be noticeable in the analysis done by May and colleagues.
First-order weight functions show small negative weights above BF. Presumably these negative weights also reflect cochlear suppression. As is the case for second-order weights (Fig. 9), the negative first-order weights are seen at higher sound levels. However, the effects of cochlear suppression are present in the first-order weights in more ways than as negative weights. As was pointed out in discussing Fig. 3A, the rate response can be smaller, for a given S(BF), when the overall stimulus energy is higher. This is especially so in low SR fibers, which show the effects of suppression more strongly. This effect was previously reported for vowel responses in AN fibers (May et al. 1996; Wong et al. 1998). Presumably the higher overall level of the components of the broadband stimulus suppresses responses to individual frequency components, with the result that a dynamic response is observed over a wider range of levels.
The dynamic range of AN fibers can be defined as the range over which weights are substantially nonzero. Figure 6, A and B, shows that this range is quite wide. For the high SR fibers, the range extends over 80 dB, much wider than the range usually quoted for rate versus level functions for tones (Evans and Palmer 1980; Sachs and Abbas 1974). The fact that AN fibers have wider dynamic ranges for broadband stimuli than for tones has been shown previously in experiments with broadband noise (Heinz and Young 2004; Schalk and Sachs 1980) and envelope modulation (Joris and Yin 1992; Smith and Brachman 1982). Again this effect seems to be a result of cochlear suppression.
Weight-function model
The weight-function model has two important features: first, it uses a logarithmic measure of frequency, and second, it is based on a logarithmic (dB) measure of stimulus energy. The logarithmic spacing of bins on the frequency axis is justified by the approximately logarithmic spacing of frequencies on the basilar membrane (Greenwood 1961; Keithley and Schreiber 1987; Liberman 1982). The use of logarithmic frequency bins also means that the RSS stimuli have amplitude spectra with an average 1/f shape. This is a reasonable stimulus design as many natural sounds, like speech and HRTF-filtered noise, have such 1/f spectra on average.
The justification of the decibel measure of stimulus energy (as opposed to, say, a linear measure of sound pressure in pascals or sound intensity in watts) was originally that it works well as an empirical description of AN rate responses: a plot of discharge rate versus the decibel stimulus spectrum level at BF produces an approximately linear relationship over 20–30 dB for vowels (Conley and Keilson 1995; May et al. 1996) and for HRTF-filtered noise (May and Huang 1997). Further support is provided by the recent result that neurons in the inferior colliculus respond more strongly to pseudorandom stimuli based on a logarithmic, as opposed to linear, intensity scale (Escabi and Schreiner 2002). The results shown here demonstrate that the dependence is not simply linear, that higher-order correction terms are needed to obtain the best fit. However, the results of prediction experiments with the HRTF stimuli show that the second-order corrections can be a relatively small effect in AN fibers.
The weight-function model takes the form of a Taylor-series approximation of the nonlinear rate response function of the neuron. As such, it is valid only over a limited, but unknown, range near the center-point of the approximation, which in the present case is an all-zero-dB RSS stimulus of a particular sound level. Although the model works well in AN fibers with a 12-dB SD of the bin levels, such is not the case in neurons in the DCN (Yu and Young 2000), where good approximation can be achieved only with a smaller stimulus range (down to 3 dB) (L. J. Reiss, S. Bandyopadhyay, and E. D. Young, unpublished data). Ultimately, the usefulness of the weight-function model is limited to those neurons for which a low-order Taylor approximation is sufficient. In fact, one use of the weight-function method, which is enabled by the ease with which it can be used to predict rate responses to arbitrary wideband spectra, is to classify neurons according to their degree of linearity of spectral processing (Yu and Young 2000).
There is a risk in using log amplitudes for S(f) in that doing so departs from the theoretical foundations provided by the Wiener-Volterra theory. The use of logarithmic bin amplitudes is empirically motivated in this work by the studies quoted in the preceding text and by the fact that the weight model works well in fitting and predicting AN rate data. One speculation about the logarithmic model is that it represents an approximation to the compressive input/output function of the basilar membrane. In this view, the input signal to the AN fiber is a compressed function of the stimulus, so that a logarithmic or other compressive measure of stimulus energy is appropriate in a model of rate responses. This question remains as a matter for future work.
Importance of second-order terms
Although the need for the second-order terms in modeling AN fibers seems clear, they were not important in predicting rate responses to the HRTF spectra. There are two possible reasons for the difference between the RSS and HRTF spectra in this regard. One reason is that the range of variation of the RSS stimuli around the reference level is larger (SD 12 dB) than for the HRTF spectra (SD 9 dB). Smaller-amplitude stimulus variations produce rate responses that are better fit by low-order spectral weight models (L. J. Reiss, S. Bandyopadhyay, and E. D. Young, unpublished data). A second possible reason is that the HRTF stimuli have much smoother spectra than the RSS stimuli. Whereas the RSS stimuli have uncorrelated values in adjacent frequency bins, the HRTF spectra generally vary slowly with frequency except at notches. In addition, the notches are substantially reduced in amplitude by the filtering inherent in grouping HRTF stimuli into bins for the weight-function model. Smoother spectra are likely to lead to a cancellation of positive and negative terms in the second-order summation, reducing its importance. Thus RSS spectra are a more challenging test of the model, even though it is computed from them in the first place.
It is instructive to relate the first- and second-order weights to the orders of nonlinearity defined in the Wiener-Volterra theory. First-order representations are those that respond to the stimulus waveform like a linear filter. They are generally estimated by reverse correlation, which gives the first Wiener kernel directly (de Boer and de Jongh 1978; de Boer and Kuyper 1968). Second-order representations are computations like the STRF that model responses to the stimulus envelope (Theunissen et al. 2001) or power spectrum (Aertsen and Johannesma 1981). The STRF can be derived from the second Wiener kernel, so that it is a second-order characterization in the Wiener-Volterra sense (Aertsen and Johannesma 1981; Klein et al. 2000; Lewis and van Dijk 2004). It can be shown that the first-order weights used here are equivalent to the STRF integrated along its time axis, i.e., to the STRF collapsed on its frequency axis. Of course, the STRF uses linear stimulus energy, not logarithmic as in the weight model. However, STRFs computed from stimuli with logarithmically defined stimulus levels have similar properties to those defined from linear stimulus levels (Escabi et al. 2003). The difference in the definition of the stimulus, while important mathematically, does not change the qualitative nature of empirical STRFs by much. Thus first-order weights are analogous to a second-order representation in the Wiener-Volterra theory, with the reservations stated above. In the same way, the second-order weights are analogous to a fourth-order representation in the Wiener-Volterra theory.
In the central auditory system, neurons become noticeably nonlinear. As discussed in the INTRODUCTION, methods such as the STRF that weight the stimulus spectrum linearly have difficulty in predicting responses to arbitrary stimuli (e.g., Eggermont et al. 1983; Machens et al. 2004; Yeshurun et al. 1989) and give different receptive fields for different stimulus sets (e.g., Theunissen et al. 2000). These failures are important because if a receptive field model does not predict the responses to stimuli accurately, it is not clear to what extent it is a meaningful description of the responses of the neuron.Moreover even in the inferior colliculus, there are units that do not have a significant STRF (Escabi and Schreiner 2002), suggesting that such units must be characterized with higher-order models. Results in auditory cortex, where responses are not well-fit with first-order weight functions, also suggest the potential of further work with higher-order characterizations (Barbour and Wang 2003). By allowing higher-order model construction, the weight function method may be useful for studying the spectral selectivity of such units.
Measures of frequency selectivity
The first-order weights in Fig. 5 provide a measure of the frequency selectivity of AN fibers responding to a broadband stimulus. This measure differs from those derived using tones (i.e., tuning curves or response maps) because it includes interactions among the frequency components of the broadband stimulus used to obtain it. The first-order weights share with revcor filters and STRF measures of AN tuning the property that the bandwidth of the tuning increases with sound level (Fig. 6C) (Carney and Yin 1988; Kim and Young 1994; Lewis and Henry 1994; Møller 1977; Recio-Spinoso et al. 2005).
It is useful to compare the increase with that measured for basilar membrane tuning using tones (Rhode 1984; Ruggero et al. 1997). Figure 11 shows a comparison of the halfwidths of weight functions (from Fig. 6C) with the halfwidths of two published gain functions for the chinchilla basilar membrane (Ruggero et al. 1997). The basilar membrane gain functions are comparable to weight functions in that both are gains in units of response per stimulus level and the halfwidths are computed in the same way (described in the caption of Fig. 6C). The difference between response/decibel in one case and response/pascal in the other can be shown not to change the shape of the gain functions (i.e., gain vs. frequency) as long as the basilar membrane data for each gain function are taken at a fixed sound level. The basal difference in bandwidth between cat and chinchilla cochleas is removed by plotting the widths in relative terms, i.e., in terms of octaves. On these log-log axes, the rate of widening of basilar membrane tuning with sound level (the slope of the functions in Fig. 11) is more than twice that of the weight functions (see the legend).
|
), the basilar membrane tuning is presumably mostly first order in the Wiener sense. Thus the comparison in Fig. 11 is between a first- and second-order characterization of peripheral tuning, and the difference in the growth of bandwidth with level may reflect some aspect of first- versus second-order components of the system. Consistent with this result, the bandwidths of STRFs for AN fibers are narrower than those of revcor functions and increase less with sound level (Kim and Young 1994). Taken together, these data suggest that tuning is different for responses that are predominantly first order, such as low-BF AN fibers or basilar membrane responses at all BFs, and responses that are predominantly second-order, such as high BF AN fibers. For this to be true, there must be a narrowing of effective tuning between the basilar membrane and AN responses in high-BF fibers. The weight function method provides a view of AN spectral sensitivity that complements other approaches. Its principal advantages are the possibility of predicting rate responses to arbitrary wideband spectra and the possibility of studying higher order nonlinearities than is feasible with reverse-correlation methods. The corresponding disadvantage is that no information is provided about temporal characteristics of neurons. As long as neurons’ responses are separable, however, the temporal and spectral sensitivities can be studied separately.
|
GRANTS |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTSDISCUSSIONGRANTSACKNOWLEDGMENTSREFERENCES |
|---|
|
ACKNOWLEDGMENTS |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTSDISCUSSIONGRANTSACKNOWLEDGMENTSREFERENCES |
|---|
Present address of B.M. Calhoun: Scientific Learning, Oakland, CA.
|
FOOTNOTES |
|---|
Address for reprint requests and other correspondence: E. D. Young, Dept. of Biomedical Engineering, 505 Traylor Bldg. Johns Hopkins School of Medicine, 720 Rutland Ave., Baltimore, MD 21205 (E-mail: eyoung{at}jhu.edu)
|
REFERENCES |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTSDISCUSSIONGRANTSACKNOWLEDGMENTSREFERENCES |
|---|
Aertsen AMHJ, Olders JHJ, and Johannesma PIM. Spectro-temporal receptive fields of auditory neurons in the grassfrog. III. Analysis of the stimulus-event relation for natural stimuli. Biol Cybern 39: 195–209, 1981.[Medline]
Barbour DL and Wang X. Auditory cortical responses elicited in awake primates by random spectrum stimuli. J Neurosci 23: 7194–7206, 2003.
Bruce IC, Sachs MB, and Young ED. An auditory-periphery model of the effects of acoustic trauma on auditory nerve responses. J Acoust Soc Am 113: 369–388, 2003.[Medline]
Calhoun BM, Miller RL, Wong JC, and Young ED. Rate encoding of stimulus spectra by auditory nerve fibers. In: Psychophysical and Physiological Advances in Hearing, edited by Palmer AR, Rees A, Summerfield Q, and Meddis R. London: Whurr, 1998, p. 170–177.
Carney LH and Yin TCT. Temporal coding of resonances by low-frequency auditory nerve fibers: single-fiber responses and a population model. J Neurophysiol 60: 1653–1677, 1988.
Conley RA and Keilson SE. Rate representation and discriminability of second formant frequencies for /
/-like steady-state vowels in cat auditory nerve. J Acoust Soc Am 98: 3223–3234, 1995.[CrossRef][Web of Science][Medline]
Dau T, Kollmeier B, and Kohlrausch A. Modeling auditory processing of amplitude modulation. I. Detection and masking with narrow-band carriers. J Acoust Soc Am 102: 2892–2905, 1997.[CrossRef][Web of Science][Medline]
de Boer E and de Jongh HR. On cochlear encoding: potentialities and limitations of the reverse-correlation technique. J Acoust Soc Am 63: 115–135, 1978.[CrossRef][Web of Science][Medline]
de Boer E and Kuyper P. Triggered correlation. IEEE Trans BME 15: 169–179, 1968.
Delgutte B. Two-tone rate suppression in auditory-nerve fibers: dependence on suppressor frequency and level. Hear Res 49: 225–246, 1990.[Medline]
Delgutte B, Hammond BM, and Cariani PA. Neural coding of the temporal envelope of speech: relation to modulation transfer functions. In: Psychophysical and Physiological Advances in Hearing, edited by Palmer AR, Rees A, Summerfield AQ, and Meddis R. London: Whurr, 1998, p. 595–603.
Deng L and Geisler CD. A composite auditory model for processing speech sounds. J Acoust Soc Am 82: 2001–2012, 1987.[Medline]
Depireux DA, Simon JZ, Klein DJ, and Shamma SA. Spectro-temporal response field characterization with dynamic ripples in ferret primary auditory cortex. J Neurophysiol 85: 1220–1234, 2001.
Efron B and Tibshirani RJ. An Introduction to the Bootstrap. New York: Chapman and Hall, 1993.
Eggermont JJ. Wiener and Volterra analyses applied to the auditory system. Hear Res 66: 177–201, 1993.[CrossRef][Web of Science][Medline]
Eggermont JJ, Aertsen AMHJ, Hermes DJ, and Johannesma PIM. Spectrotemporal characterization of auditory neurons: redundant or necessary. Hear Res 5: 109–121, 1981.[CrossRef][Web of Science][Medline]
Eggermont JJ, Aertsen AMHJ, and Johannesma PIM. Prediction of the responses of auditory neurons in the midbrain of the grass frog based on the spectro-temporal receptive field. Hear Res 10: 191–202, 1983.[CrossRef][Web of Science][Medline]
Escabi MA, Miller LM, Read HL, and Schreiner CE. Naturalistic auditory contrast improves spectrotemporal coding in the cat inferior colliculus. J Neurosci 23: 11489–11504, 2003.
Escabi MA and Read HL. Representation of spectrotemporal sound information in the ascending auditory pathway. Biol Cybern 89: 350–362, 2003.[Medline]
Escabi MA and Schreiner CE. Nonlinear spectrotemporal sound analysis by neurons in the auditory midbrain. J Neurosci 22: 4114–4131, 2002.
Evans EF. Frequency selectivity at high signal levels of single units in cochlear nerve and nucleus. In: Psychophysics and Physiology of Hearing, edited by Evans EF and Wilson JP. London: Academic, 1977, p. 185–196.
Evans EF and Palmer AR. Relationship between the dynamic range of cochlear nerve fibers and their spontaneous activity. Exp Brain Res 40: 115–118, 1980.[Web of Science][Medline]
Greenwood DD. Critical bandwidth and the frequency coordinates of the basilar membrane. J Acoust Soc Am 33: 1344–1356, 1961.[CrossRef]
Hancock KE and Voigt HF. Wideband inhibition of dorsal cochlear nucleus type IV units in cat: a computational model. Ann Biomed Eng 27: 73–87, 1999.[Medline]
Heinz MG and Young ED. Response growth with sound level in auditory-nerve fibers after noise-induced hearing loss. J Neurophysiol 91: 784–795, 2004.
Hewitt MJ, Meddis R, and Shackleton TM. A computer model of a cochlear-nucleus stellate cell: responses to amplitude-modulated and pure-tone stimuli. J Acoust Soc Am 91: 2096–2109, 1992.[Medline]
Johnson DH. Applicability of white-noise nonlinear system analysis to the peripheral auditory system. J Acoust Soc Am 68: 876–884, 1980.[CrossRef][Web of Science][Medline]
Joris PX and Yin TCT. Responses to amplitude-modulated tones in the auditory nerve of the cat. J Acoust Soc Am 91: 215–232, 1992.[CrossRef][Web of Science][Medline]
Kalluri S and Delgutte B. Mathematical models of cochlear nucleus onset neurons. II. Model with dynamic spike-blocking state. J Comput Neurosci 14: 91–110, 2003.[CrossRef][Web of Science][Medline]
Keithley EM and Schreiber RC. Frequency map of the spiral ganglion in the cat. J Acoust Soc Am 81: 1036–1042, 1987.[CrossRef][Web of Science][Medline]
Kim PJ and Young ED. Comparative analysis of spectro-temporal receptive fields, reverse correlation functions, and frequency tuning curves of auditory-nerve fibers. J Acoust Soc Am 95: 410–422, 1994.[CrossRef][Web of Science][Medline]
Klein DJ, Depireux DA, Simon JZ, and Shamma SA. Robust spectrotemporal reverse correlation for the auditory system: optimizing stimulus design. J Comput Neurosci 9: 85–111, 2000.[CrossRef][Web of Science][Medline]
Le Prell G, Sachs MB, and May BJ. Representation of vowel-like spectra by discharge rate responses of individual auditory-nerve fibers. Aud Neurosci 2: 275–288, 1996.
Lewis ER and Henry KR. Dynamic changes in tuning in the gerbil cochlea. Hear Res 79: 183–189, 1994.[Medline]
Lewis ER, Henry KR, and Yamada WM. Tuning and timing of excitation and inhibition in primary auditory nerve fibers. Hear Res 171: 13–31, 2002.[CrossRef][Web of Science][Medline]
Lewis ER, and van Dijk P. New variations on the derivation of spectro-temporal receptive fields for primary auditory afferent axons. Hear Res 189: 120–136, 2004.[Medline]
Liberman MC. Auditory-nerve response from cats raised in a low-noise chamber. J Acoust Soc Am 63: 442–455, 1978.[CrossRef][Web of Science][Medline]
Liberman MC. The cochlear frequency map for the cat: labeling auditory-nerve fibers of known characteristic frequency. J Acoust Soc Am 72: 1441–1449, 1982.[CrossRef][Web of Science][Medline]
Machens CK, Wehr MS, and Zador AM. Linearity of cortical receptive fields measured with natural sounds. J Neurosci 24: 1089–1100, 2004.
May BJ, Huang A, LePrell G, and Hienz RD. Vowel formant frequency discrimination in cats: comparison of auditory nerve representations and psychophysical thresholds. Aud Neurosci 3: 135–162, 1996.
May BJ and Huang AY. Spectral cues for sound localization in cats: a model for discharge rate representations in the auditory nerve. J Acoust Soc Am 101: 2705–2719, 1997.[CrossRef][Web of Science][Medline]
May BJ, LePrell GS, and Sachs MB. Vowel representations in the ventral cochlear nucleus of the cat: effects of level, background noise, and behavioral state. J Neurophysiol 79: 1755–1767, 1998.
Møller AR. Frequency selectivity of single auditory-nerve fibers in response to broadband noise stimuli. J Acoust Soc Am 62: 135–142, 1977.[CrossRef][Web of Science][Medline]
Musicant AD, Chan JCK, and Hind JE. Direction-dependent spectral properties of cat external ear: new data and cross-species comparisons. J Acoust Soc Am 87: 757–781, 1990.[CrossRef][Web of Science][Medline]
Nelken I, Kim PJ, and Young ED. Linear and non-linear spectral integration in type IV neurons of the dorsal cochlear nucleus. II. Predicting responses using non-linear methods. J Neurophysiol 78: 800–811, 1997.
Press WH, Flannery BP, Teukolsky SA, and Vetterling WT. Numerical Recipes: The Art of Scientific Computing. Cambridge, UK: Cambridge Univ. Press, 1986.
Qiu A, Schreiner CE, and Escabi MA. Gabor analysis of auditory midbrain receptive fields: spectro-temporal and binaural composition. J Neurophysiol 90: 456–476, 2003.
Recio A, Narayan SS, and Ruggero MA. Wiener-kernel analysis of basilar-membrane responses to white noise. In: Diversity in Auditory Mechanics, edited by Lewis ER, Long GR, Lyon RF, Narins PM, Steele CR and Hecht-Poinar E. Singapore: World Sci. Publ, 1997, p. 325–331.
Recio-Spinoso A, Temchin AN, van Dijk P, Fan YH, and Ruggero MA. Wiener-kernel analysis of responses to noise of chinchilla auditory-nerve fibers. J Neurophysiol 93: 3615–3634, 2005.
Rhode WS. Cochlear mechanics. Annu Rev Physiol 46: 231–246, 1984.[Medline]
Rice JJ, May BJ, Spirou GA, and Young ED. Pinna-based spectral cues for sound localization in cat. Hear Res 58: 132–152, 1992.[CrossRef][Web of Science][Medline]
Rice JJ, Young ED, and Spirou GA. Auditory-nerve encoding of pinna-based spectral cues: rate representation of high-frequency stimuli. J Acoust Soc Am 97: 1764–1776, 1995.[CrossRef][Web of Science][Medline]
Rothman JS and Manis PB. The roles potassium currents play in regulating the electrical activity of ventral cochlear nucleus neurons. J Neurophysiol 89: 3097–3113, 2003.
Rothman JS, Young ED, and Manis PB. Convergence of auditory nerve fibers onto bushy cells in the ventral cochlear nucleus: implications of a computational model. J Neurophysiol 70: 2562–2583, 1993.
Ruggero MA, Rich NC, Recio A, and Narayan SS. Basilar-membrane responses to tones at the base of the chinchilla cochlea. J Acoust Soc Am 101: 2151–2163, 1997.[CrossRef][Web of Science][Medline]
Sachs MB and Abbas PJ. Rate versus level functions for auditory-nerve fibers in cats: tone-burst stimuli. J Acoust Soc Am 56: 1835–1847, 1974.[CrossRef][Web of Science][Medline]
Sachs MB and Abbas PJ. Phenomenological model for two tone suppression. J Acoust Soc Am 60: 1157–1163, 1976.[CrossRef]
Sachs MB and Young ED. Effects of nonlinearities on speech encoding in the auditory nerve. J Acoust Soc Am 68: 858–875, 1980.[CrossRef][Web of Science][Medline]
Schalk T and Sachs MB. Nonlinearities in auditory-nerve fiber response to bandlimited noise. J Acoust Soc Am 67: 903–913, 1980.[CrossRef][Web of Science][Medline]
Sen K, Theunissen FE, and Doupe AJ. Feature analysis of natural sounds in the songbird auditory forebrain. J Neurophysiol 86: 1445–1458, 2001.
Smith RL and Brachman ML. Adaptation in auditory-nerve fibers: a revised model. Biol Cybern 44: 107–120, 1982.[Medline]
Sumner CJ, O’Mard LP, Lopez-Poveda EA, and Meddis R. A nonlinear filter-bank model of the guinea-pig cochlear nerve: rate responses. J Acoust Soc Am 113: 3264–3274, 2003.[CrossRef][Web of Science][Medline]
Temchin AN, Recio-Spinoso A, van Dijk P, and Ruggero MA. Wiener kernels of chinchilla auditory-nerve fibers: verification using responses to tones, clicks and noise and comparison with basilar-membrane vibrations. J Neurophysiol 93: 3635–3648, 2005.
Theunissen FE, David SV, Singh NC, Hsu A, Vinje WE, and Gallant JL. Estimating spatio-temporal receptive fields of auditory and visual neurons from their responses to natural stimuli. Network Comput Neural Syst 12: 289–316, 2001.[CrossRef]
Theunissen FE, Sen K, and Doupe AJ. Spectral-temporal receptive fields of nonlinear auditory neurons obtained using natural sounds. J Neurosci 20: 2315–2331, 2000.
van der Heijden M and Joris PX. Cochlear phase and amplitude retrieved from the auditory nerve at arbitrary frequencies. J Neurosci 23: 9194–9198, 2003.
van Dijk P and Wit HP. Wiener kernel analysis of inner ear function in the American bullfrog. J Acoust Soc Am 95: 904–919, 1994.[Medline]
van Gisbergen JAM, Grashuis JL, Johannesma PIM, and Vendrik AJH. Neurons in the cochlear nucleus investigated with tone and noise stimuli. Exp Brain Res 23: 387–406, 1975.[Web of Science][Medline]
Versnel H and Shamma SA. Spectral-ripple representation of steady-state vowels in primary auditory cortex. J Acoust Soc Am 103: 2502–2514, 1998.[CrossRef][Web of Science][Medline]
Wong JC, Miller RL, Calhoun BM, Sachs MB, and Young ED. Effects of high sound levels on responses to the vowel // in cat auditory nerve. Hear Res 123: 61–77, 1998.[Medline]
Yamada WM and Lewis ER. Predicting the temporal responses of non-phase-locking bullfrog auditory units to complex acoustic waveforms. Hear Res 130: 155–170, 1999.[CrossRef][Web of Science][Medline]
Yeshurun Y, Wollberg Z, and Dyn N. Prediction of linear and non-linear responses of MGB neurons by system identification methods. Bull Math Biol 51: 337–346, 1989.[Medline]
Yu JJ and Young ED. Linear and nonlinear pathways of spectral information transmission in the cochlear nucleus. Proc Natl Acad Sci USA 97: 11780–11786, 2000.
Zhang X, Heinz MG, Bruce IC, and Carney LH. A phenomenological model for the responses of auditory-nerve fibers. I. Nonlinear tuning with compression and suppression. J Acoust Soc Am 109: 648–670, 2001.[CrossRef][Web of Science][Medline]
This article has been cited by other articles:
|
|
 |
|
  M. Pienkowski, G. Shaw, and J. J. Eggermont Wiener-Volterra Characterization of Neurons in Primary Auditory Cortex Using Poisson-Distributed Impulse Train Inputs J Neurophysiol, June 1, 2009; 101(6): 3031 - 3041. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 M. B. Ahrens, J. F. Linden, and M. Sahani Nonlinearities and Contextual Influences in Auditory Cortical Responses Modeled with Multilinear Spectrotemporal Methods J. Neurosci., February 20, 2008; 28(8): 1929 - 1942. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
S. Bandyopadhyay, L. A. J. Reiss, and E. D. Young Receptive Field for Dorsal Cochlear Nucleus Neurons at Multiple Sound Levels J Neurophysiol, December 1, 2007; 98(6): 3505 - 3515. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
L. Holmstrom, P. D. Roberts, and C. V. Portfors Responses to Social Vocalizations in the Inferior Colliculus of the Mustached Bat Are Influenced by Secondary Tuning Curves J Neurophysiol, December 1, 2007; 98(6): 3461 - 3472. [Abstract] [Full Text] [PDF]
|
|
|
|
|
|
L. A. J. Reiss, S. Bandyopadhyay, and E. D. Young Effects of Stimulus Spectral Contrast on Receptive Fields of Dorsal Cochlear Nucleus Neurons J Neurophysiol, October 1, 2007; 98(4): 2133 - 2143. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||
| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
| Visit Other APS Journals Online |
20% of rates within 1 SD of SR) or saturation (<10% of rates <0.67*mean rate +0.33*SR). Eliminating responses near threshold and saturation eliminates the curvature in rate responses like that shown in 
