1, and 14.3 under low, medium, and high contrast, respectively (Figure S2). These lie at approximately the same percentile (∼70%) of each stimulus distribution,
relative to their projection onto X⋅vX⋅v. Neurons in auditory cortex thus adapt their sensitivity to be most informative about stimuli within the current stimulus distribution. To fully quantify the relationship learn more between stimulus contrast and gain, we presented to a subset of these cells a larger set of DRCs with eight different σL values ranging from 1.4 dB to 11.5 dB (c = 17% to 116%). We obtained 80 units for which the above analysis could be performed over the whole contrast range. On average, these showed a clear, monotonic increase in gain as the contrast of the stimulus was reduced ( Figure 4E). The relationship between relative gain and contrast was extremely well described by a standard normalization equation ( Heeger, 1992 and Carandini et al., 1997): equation(2) G(σL)=a1+bσLnwhere G denotes the gain and a,
b, and n are constants (see Model 5 in Table S2). This model explained 99.9% of the variance in the population average of relative gain values. This model also provided a good description of the relative gain values for individual units (Figure S3H). However, in some units, the model failed at the lowest contrasts. For these units, gain increased as contrast was reduced down to a threshold, below which gain either leveled off or decreased. For 46/80 units, this threshold was σL = 2.9 dB (c = 33%); for a further 26 units, this threshold was 4.3 dB (c = ZD6474 49%); and a further four units had a threshold of σL = 5.8 dB (c = 64%). At these thresholds and above, gain was well fit on a cell-by-cell basis by Equation 2 for 76/80 units. The model produced marginally Adenosine better predictions of neural responses than fitting individual nonlinearities to each contrast condition ( Table S2). Thus, across a wide range of contrasts, gain normalization is a robust phenomenon for individual units. In the experiments presented so far, the mean SPL of each tone in the DRC, μL,
was kept fixed. To explore the effect of mean, we presented a further set of stimuli in which both the mean of the level distributions (μL) and the contrast (σL) were manipulated independently. We estimated LN models from responses to a range of mean/contrast conditions, together with curve transformations from each stimulus condition relative to the μL = 40 dB SPL, σL = 8.7 dB (c = 92%) nonlinearity. Of the 1001 units above, 56 units yielded predictive LN models across the whole range of conditions. Only data from these 56 units are analyzed below, in order to maintain the same sample set across stimulus conditions. Nevertheless, data from all units where LN models were predictive in only a subset of conditions (n = 217) yielded similar results (data not shown). At all mean levels tested, decreasing contrast caused gain to increase across the population of cells.