Contrast gain control in auditory cortex

Abstract

The auditory system must represent sounds with a wide range of statistical properties. One important property is the spectrotemporal contrast in the acoustic environment: the variation in sound pressure in each frequency band, relative to the mean pressure. We show that neurons in ferret auditory cortex rescale their gain to partially compensate for the spectrotemporal contrast of recent stimulation. When contrast is low, neurons increase their gain, becoming more sensitive to small changes in the stimulus, although the effectiveness of contrast gain control is reduced at low mean levels. Gain is primarily determined by contrast near each neuron's preferred frequency, but there is also a contribution from contrast in more distant frequency bands. Neural responses are modulated by contrast over timescales of ∼100 ms. By using contrast gain control to expand or compress the representation of its inputs, the auditory system may be seeking an efficient coding of natural sounds.

Figure 1

Stimulus Paradigm (A–C) Dynamic random chord (DRC) sequences with different spectrotemporal contrasts. The elements of each sequence are chords of pure tones, whose levels are randomly chosen from the distributions in (D). (D) Tone level distributions for the DRC sequences in (A)–(C). These all have the same mean level μ_L, but different widths. The blue line corresponds to the low-contrast (σ_L) DRC (A); the green line is the medium-contrast DRC (B); the red is the high-contrast DRC (C). (E and F) The widths of the level distribution determine both the spectral and the temporal contrast of the individual sequences, as shown by the temporal profiles of a pure tone at a fixed frequency for the three sequences in (E) and the spectral profiles of a chord at a fixed time for the three sequences in (F). (G) We characterized the relationship between stimulus and neuronal response using a linear-nonlinear model. The sound input is treated as a spectrogram, X; the (normalized) receptive field v acts as a linear filter on X, extracting the relevant features of the sound via the dot product $\mathbf{X}\cdot \mathbf{v}$. The output of the linear filter is (optionally) matched to the output spike rate through a nonlinearity that captures features such as thresholding. We parameterized these nonlinearities as sigmoids. See also Tables S1 and S2 and Figure S1.

Figure 2

Effect of Stimulus Contrast on Tuning (A and B) Example units show the same spectrotemporal selectivity in the receptive fields (STRFs) estimated from the three contrast conditions. Red denotes components of the stimuli that excite the unit, blue denotes components that inhibit. As the color scale is uniform across the plots, it is clear that the dominant variation across the conditions lies in the magnitude of the drive to the unit: this is stronger (the cortex appears to “listen harder”) under low-contrast than under high-contrast stimulus conditions. (C–E) Best frequency (C) does not vary systematically with stimulus contrast across units. Tuning bandwidth (D) shows a small, significant broadening at low contrast. In almost all units, the gain of the STRF (E) increases as contrast decreases. STRF properties under low-contrast stimulation are shown as blue circles, under medium contrast as green crosses. Filled dark blue circles and dark green pluses indicate data in these two same conditions from the awake recordings. (F) The linear STRFs fitted from DRCs with different contrasts are sufficiently similar in tuning properties that, once adjusted for differences in STRF gain, they predict responses across stimulus conditions on average 96.5% as well as they do within their own conditions. Thus, most of the contrast dependence of STRFs is captured by a change in gain. Red crosses indicate data from the awake recordings. See also Table S3.

Figure 3

Output Nonlinearities for Two Example Units Show Gain Rescaling as a Function of Stimulus Contrast (A) For each unit, we fitted a single linear STRF, then calculated a separate output nonlinearity for each contrast condition. The abscissa denotes the output of the linear STRF, $\mathbf{X}\cdot \mathbf{v}$; the ordinate is the predicted spike rate. (B) The differences between the nonlinearities were quantified by the linear transform required to convert the high-contrast nonlinearity into each of the medium- and low-contrast nonlinearities. Solid curves show the original sigmoids for the unit shown in (A); dashed lines show the result of the transformation of the high-contrast (red) curve into the low-contrast (blue) and medium-contrast (green) curves. The parameters for the high-to-low transform for this unit were G = 2.9, Δx = 17.6%, Δy = 8.9%; the high-to-medium transform parameters were G = 1.5, Δx = 2.0%, Δy = −0.6%. (C and D) Nonlinearities for a second example unit; panels equivalent to (A) and (B). Parameters for these transforms were G = 1.5, Δx = −27.6%, Δy = 2.4% (high-to-low) and G = 1.2, Δx = −3.8%, Δy = 8.5% (high-to-medium). See also Figure S2.

Figure 4

Output Gain Is Inversely Related to Stimulus Contrast (A) Histogram of gain (relative to the reference condition) under low- (blue) and medium-contrast (green) conditions, for n = 315 units. Colored dashed lines are population medians. Red dashed line indicates G = 1, i.e., no scaling of the nonlinearity. (B) As in (A), for changes in x-offset. Positive values denote rightward shifts in nonlinearities. (C) As in (A), for changes in y-offset. Positive values denote upward shifts in nonlinearities. (D) For both the high-to-low-contrast (blue) and the high-to-medium-contrast (green) transformations, changes in gain (abscissa) and x-offset (ordinate) were positively correlated across the population of units. Shaded regions show the (bootstrapped) 99% confidence intervals about the mean regression line (solid). (E) The population median gain over a wide range of contrasts was well modeled by a standard gain normalization function, Equation 2. Gray bars indicate 99% confidence intervals on the median, via bootstrapping. See also Figure S3.

Figure 5

Contrast Gain Control Is Modulated by μ_L Only at Low Mean Levels (A) Sigmoid nonlinearities were fitted to units' responses to a range of DRCs with different contrast and μ_L statistics and compared with a reference curve. Population median gain factors are plotted against contrast, with different symbols/colors for each mean level. Colored lines show independent fits of the model in Figure 4E. (B) As in (A), for x-offset. (C) As in (A), for y-offset. (D) Values of b fitted to the model given in Equation 8, as shown in (A). These measure the relative sensitivity of neural gain to the stimulus contrast, as a function of mean level. At mean levels ≥35 dB SPL, $b\left({\mu }_L\right)$ is relatively independent of μ_L, while it becomes sensitive to μ_L at lower mean levels. Solid line denotes an exponential fit to $b\left({\mu }_L\right)$. (E) Illustration of the full model $G\left({\sigma }_L,{\mu }_L\right)$ for gain normalization (colored surface). A contour plot of this surface is projected below. Colored dots on the contour plot show the population median data used to constrain the model, from (A). Their position is a function of both σ_L and μ_L; their color denotes the median measured gain in these stimulus conditions; vertical arrows denote the residual between the measured gain and that described by the model. See also Figure S4.

Figure 6

Responses to Fixed Sounds Are Modulated by the Spectrotemporal Contrast of their Context (A) During each electrode penetration, the STRF of a representative unit was used as a basis for a test sound. This was inserted at random times into special DRCs in which stimulus contrast switched every 1 s between low (σ_L = 2.9 dB, c = 33%) and high contrast (σ_L = 8.7 dB, c = 92%). The test sound itself was identical within each stimulus regime; only the contrast of its context differed. (B) Mean response to the test sound for an example unit, when presented in high-contrast (top row) or low-contrast context (bottom row). Columns delineate responses by the time since the last switch in context at which the test sounds were presented. (C) Response to the test sound for the unit in (B), averaged within each contrast context over all postswitch delays from 150–800 ms. (D) Peak responses to the test sound across n = 63 units, during the low- and high-contrast contexts. Red dashed line shows expected response relationship if contrast-context was irrelevant. Green circle indicates the unit in (B) and (C). Shaded region shows the confidence intervals as in Figure 4D. (E) Peak response for unit in (B) and (C) as a function of the time after context switch at which the test sound was presented. Solid lines show exponential fits to these data, with time constant τ_L→H = 62 ms after an increase and τ_H→L = 85 ms after a decrease in the contrast of the context. (F) Time constants for context adaptation, as in (E), for 18 units for which both τ_L→H and τ_H→L could be reliably estimated. Data are plotted both as a scatter plot and as marginal histograms of τ_L→H (red) and τ_H→L (blue). Green circle denotes unit in (B), (C), and (E). See also Figure S5.

Figure 7

Gain Is Affected by the Contrast of Stimuli Lying both Within and Outside the Tuning of a Unit (A) During each electrode penetration, the STRF of a representative unit was mapped and the BF determined. A 0.5–1.2 octave band centered around the BF was designated the test and the remainder the mask. The contrast within these (σ_test, σ_mask) was independently varied. The example stimuli shown here were used for a unit with a BF of 9 kHz. (B–E) Relative gain from varying either σ_test or σ_mask, with the other kept constant. Color grids above each plot illustrate which stimulus conditions from (A) are being compared; the red box with the white dot is the reference curve used to calculate the transform. (F) Population median gain for 24 units with their responsive frequency range lying within the test (dots with dashed lines). Error bars denote 99% confidence interval on the median. Contrast within the test is a stronger determinant of the gain than that in the mask. Solid lines show the fit of Equation 3 to these data. See also Table S4.