Abstract
Noise correlations (rnoise) between neurons can affect a neural population's discrimination capacity, even without changes in mean firing rates of neurons. rnoise, the degree to which the response variability of a pair of neurons is correlated, has been shown to change with attention with most reports showing a reduction in rnoise. However, the effect of reducing rnoise on sensory discrimination depends on many factors, including the tuning similarity, or tuning correlation (rtuning), between the pair. Theoretically, reducing rnoise should enhance sensory discrimination when the pair exhibits similar tuning, but should impair discrimination when tuning is dissimilar. We recorded from pairs of neurons in primary auditory cortex (A1) under two conditions: while rhesus macaque monkeys (Macaca mulatta) actively performed a threshold amplitude modulation (AM) detection task and while they sat passively awake. We report that, for pairs with similar AM tuning, average rnoise in A1 decreases when the animal performs the AM detection task compared with when sitting passively. For pairs with dissimilar tuning, the average rnoise did not significantly change between conditions. This suggests that attention-related modulation can target selective subcircuits to decorrelate noise. These results demonstrate that engagement in an auditory task enhances population coding in primary auditory cortex by selectively reducing deleterious rnoise and leaving beneficial rnoise intact.
Introduction
Understanding how neural populations encode information and how changes in behavioral states affect these codes constitutes an exciting frontier in neuroscience. Many studies report that interneuronal correlations contribute to population coding fidelity (Zohary et al., 1994; Oram et al., 1998; Averbeck et al., 2006; Cohen and Maunsell, 2009). Tuning correlation (rtuning) quantifies the degree to which two neurons respond similarly to a stimulus set; noise correlation (rnoise) quantifies the degree to which two neurons' response variability to a given stimulus is correlated. When rtuning and rnoise have identical sign, sensory discrimination is impaired; and when they have opposite sign, discrimination is enhanced (Oram et al., 1998; Shadlen and Newsome, 1998; Abbott and Dayan, 1999; Nirenberg and Latham, 2003; Romo et al., 2003; Averbeck et al., 2006). Most studies report that both rtuning and rnoise between pairs of nearby neurons are on average positive in sensory cortex and that rnoise remains positive even when rtuning is negative (Smith and Kohn, 2008; Cohen and Kohn, 2011). Thus, cortex contains a mixture of rtuning/rnoise relationships: some benefit the population code and some impair it.
Behavioral states, such as attention and wakefulness, affect rnoise (Cohen and Newsome, 2008; Mitchell et al., 2009; Herrero et al., 2013; Issa and Wang, 2013; Ecker et al., 2014). Most reports show that, during increased sensory demand (e.g., when attention is required), pairs of neurons decrease rnoise, regardless of rtuning (Smith and Kohn, 2008; Cohen and Maunsell, 2009). Decreasing rnoise is expected to enhance sensory discrimination for neuron pairs with positive rtuning but weaken it for pairs with negative rtuning (e.g., Gu et al., 2011). Recently, however, Jeanne et al. (2013) found that learning decreases rnoise when rtuning is positive but increases it when rtuning is negative. This suggests that selective rnoise modulation provides a mechanism not only for learning, but also for rapid changes in sensory discrimination.
The view that primary auditory cortex (A1) is a purely sensory field is changing as recent findings show that myriad behavioral variables affect its activity (Scheich et al., 2007; Jaramillo and Zador, 2011; Niwa et al., 2012a; Bizley et al., 2013; Jaramillo et al., 2014). However, the degree to which interneuronal correlations in A1 reflect behavioral states remains unclear. We therefore measured both rtuning and rnoise between pairs of A1 neurons recorded from the same electrode during both passive listening and when the animal was engaged in threshold discrimination. We asked whether task engagement affects rnoise between A1 neurons. We find that both rtuning and rnoise are on average positive in A1 and that active engagement reduces average rnoise, between pairs with positive rtuning, but not between pairs with negative rtuning. This effect is optimal for sensory discrimination because rnoise in pairs with negative rtuning tends to be positive and a reduction would impair population coding. Moreover, we find that task engagement often modulates rtuning in individual pairs. Our findings highlight the dynamic nature of A1 population coding (Bathellier et al., 2012) and establish selective rnoise modulation as a mechanism for rapid sensory enhancement.
Materials and Methods
Subjects.
We recorded extracellular activity from right primary auditory cortex (A1) in three adult rhesus macaques (Macaca mulatta; two female, one male), weighing 6–11 kg. All procedures were approved by the University of California, Davis Animal Care and Use Committee and met the requirements of the United States Public Health Service policy on experimental animal care.
Stimuli.
We presented unmodulated and sinusoidal amplitude-modulated (AM) broadband noise bursts (800 ms duration) across a range of AM frequencies and depths. For a given recording session, a single AM frequency between 2.5 and 1000 Hz was used. We selected this AM frequency based on the best modulation frequency (BMF) of the multiunit (MU) activity of the recording site; our methods for determining BMF are described in the Physiology section, below. We varied AM depth from 6% to 100%. We have previously reported our sound generation methods (O'Connor et al., 2011). Briefly, sound signals were produced using a custom MATLAB program and generated with a D/A converter (Cambridge Electronic Design, model 1401). They were then attenuated (TDT Systems, PA5 and Leader LAT-45), amplified (RadioShack, MPA-200), and passed to a speaker (RadioShack, PA-110; or Optimus, Pro-7AV) positioned 0.8 or 1.5 m in front of the subject, centered at the interaural midpoint. Sounds were generated at a 100 kHz sampling rate and cosine-ramped at the onset and offset (5 ms). Intensity was calibrated across all sounds (Bruel and Kjaer model 2231) to 63 dB at the outer ear.
Task.
The task is the same as described by Niwa et al. (2012b, 2013). We recorded extracellular activity during each of two different conditions: (1) task engagement and (2) passive listening. During task engagement (active condition), animals indicated whether a sound is AM in a Go/No Go paradigm. Animals completed a single trial by (1) waiting for a cue light to prompt trial initiation, (2) depressing a lever to initiate a trial, (3) listening to two successive sounds, an “S1” (standard), unmodulated noise burst, and an “S2” (test) stimulus, either unmodulated noise (nontarget) or AM noise (target), (4) indicating detection of target by releasing the lever within 800 ms after S2 offset or indicating the second sound was a nontarget (unmodulated) by keeping the lever depressed for 800 ms after S2 offset. S1 and S2 were both 800 ms and were separated by 400 ms of silence. Animals were rewarded with liquid (juice or water) for both hits (correctly releasing the lever after target presentation) and correct rejections (correctly withholding lever release after nontarget presentation). Animals were informed of both misses (failure to release lever after targets) and false alarms (inappropriate lever release after nontargets) and received a penalty (15–60 s timeout in which a new trial could not be initiated) for false alarms. Within a recording session, a single AM frequency was used, but depth was varied (6%, 16%, 28%, 40%, 60%, 80%, and 100% depth). Multiple stimuli (16%, 28%, and 40% AM depth) were near animals' AM detection thresholds (O'Connor et al., 2000, 2011; Niwa et al., 2012b).
During passive blocks, animals sat quietly while we presented the same stimuli as in the active condition. Animals received randomly timed liquid rewards. During a recording, animals participated in one passive and one active block, each consisting of ∼450 trials (∼50 repetitions per stimulus). For Animals V and W, we counterbalanced which condition, passive or active, came first. For Monkey X, the active condition was always followed by the passive condition. Animals were informed by cue light as to whether they were to respond to sounds (active condition) or not (passive condition).
Physiology.
The data presented here are a subset of those presented previously (Niwa et al., 2012b). Briefly, after training on the task, a craniotomy was made over the right parietal cortex. A titanium head post was implanted centrally behind the brow ridge and a CILUX recording cylinder implanted over the craniotomy. During recording, a plastic grid was attached to the cylinder to allow the passage of tungsten microelectrodes (FHC, 1–4 MΩ; Alpha-Omega, 0.5–1 MΩ) through a guide tube. The guide tube was used to puncture the dura mater; then the electrode advanced vertically via a hydraulic drive through parietal cortex to A1. During all recordings, animals sat head-restrained in an acoustically transparent chair in a sound-attenuated booth.
Extracellular signals were amplified (AM Systems, model 1800), bandpass filtered between 0.3 Hz and 10 kHz (Kron-Hite, 3382), then converted to a digital signal at a 50 kHz sampling rate (Cambridge Electronic Design, model 1401). Contributions of single neurons to the signal were determined offline. We used a cubic spline interpolation algorithm in Spike2 (Cambridge Electronic Design) to create single-unit template waveforms and then match spiking events to those templates. We then used principal components analysis to confirm that events assigned to separate single units formed separable clusters in principal component space. Thresholds for determining spiking activity above background noise were determined visually by the experimenters with the aid of Spike2's automatic trigger-setting algorithm. Spiking activity was generally 4–5 times the background noise level. Fewer than 0.2% of spike events assigned to single-neuron clusters fell within a 1 ms refractory period window. Only recordings in which >1 neuron was isolated and held during an entire recording block are analyzed in the present report.
We determined which AM frequency to present during the experiment by finding the BMF of the multi-unit (MU) activity at each recording site. MU activity was defined as any clear spiking activity well above the background noise level of the recording. After an auditory-responsive site was found, we presented 800 ms AM stimuli across a range of frequencies, all at 100% depth, as well as 800 ms unmodulated stimuli. All firing rates were calculated over the entire 800 ms stimulus period only. Then, we used signal detection analyses (receiver operating characteristic area) to find the AM frequency that the MU activity best discriminated from unmodulated sounds (i.e., the BMF). We calculated both rate-based and temporal measures (spike count and vector strength, respectively) of MU activity; thus, we derived two BMFs at each recording site. Sometimes, the spike count-based BMF (BMFsc) differed from the vector strength-based BMF (BMFvs). In these cases, we alternated which BMF we used, such that we used the BMFsc whether we used BMFvs the last time the issue arose, and vice versa.
We used stereotactic coordinates to target A1, and we used physiological response properties during recording. In one animal, we anatomically confirmed our recording location (the other two are still serving as research subjects). We measured the pure-tone tuning of recorded neurons and determined their location in A1 based on the tonotopic gradient and the sharpness of their pure-tone tuning relative to neurons recorded in belt regions. Pure-tone tuning was assessed by presenting 100 ms pure-tones varying across frequency and intensity, with at least three repetitions for each frequency/intensity combination. Stimuli were presented in random order. This allowed us to measure each unit's frequency response area by finding the contour line for frequency/intensity combinations that evoked firing rates at least 2SD above the spontaneous firing rate (determined in a 75 ms window before stimulus presentation). Thus, we could measure each unit's characteristic frequency (CF) and sharpness of tuning: bandwidth (BW). In each animal, we mapped CF and BW to determine the topographic distribution of each. Recordings from a region with a high-to-low caudal-to-rostral CF gradient with narrow BW were assigned to A1.
In one animal, we confirmed our stereotactic and physiological assignments by performing postmortem histological analyses. Upon termination of recording, we inserted biotinylated dextran amine to the rostral, middle, and caudal regions of physiologically determined A1 at the border between A1 and middle-medial belt. The animal was then given an overdose of sodium pentobarbital and perfused using 4% PFA in 0.1 M phosphate buffer. The brain was removed, blocked, and sliced into 50 μm sections, and slices were stained in alternation with the following: (1) mouse anti-parvalbumin → biotinylated horse anti-mouse secondary → acetylavidin biotinylated peroxidase complex (ABC) → diamino benzadine (DAB); (2) Nissl substance; (3) Nissl substance → ABC → DAB. This histology was previously shown by O'Connor et al. (2010), and our physiologically determined A1 borders were validated with anatomical evidence (e.g., dense parvalbumin staining in the area within the biotinylated dextran amine markers).
Data analysis: selection of single neurons for analysis.
Neurons can use both rate and temporal codes to represent AM (Liang et al., 2002; Yin et al., 2011; Johnson et al., 2012). Although many neurons display both rate and temporal codes, some do not respond to AM or the carrier by changing firing rate (Yin et al., 2011). In the present analysis, we measured correlations in firing rate both within and across stimuli between A1 neurons. Thus, we excluded from analysis any neurons that do not respond to those sounds by changing firing rate. To do so, we tested whether the firing rate in response to any stimulus presented during either task condition was significantly different from the prestimulus baseline firing rate (rank-sum test, p < 0.05, corrected for multiple comparisons). In the present report, we found 221 pairs of neurons suitable for analysis: 167 pairs in the passive condition, 199 in the active condition, and 145 in both conditions.
Tuning correlation (rtuning) and noise correlation (rnoise).
Both rtuning and rnoise were computed for each pair in each condition. rtuning is the Pearson correlation between the mean firing rates of each neuron in the pair to the set of 8 stimuli (0%, 6%, 16%, 28%, 405, 60%, 80%, and 100% AM) used during a recording. How rnoise was calculated depended on the analysis being performed. When analyzing effects without collapsing across AM depths, rnoise was calculated separately for each of the 8 stimuli, that is, as the Pearson correlation between the trial-by-trial firing rates of each neuron to each of the repetitions of a given stimulus. The minimum number of repetitions of a given stimulus was 27, the maximum was 71 and the average was 50.5. When collapsing across stimuli to derive a single rnoise measure for a pair, firing rates within each stimulus were z-scored, then combined into a single vector of normalized firing rates. Then, the Pearson correlation between these vectors was calculated. To confirm the robustness of our results, we also used nonparametric Spearman correlation to calculate rtuning and rnoise (data not shown) and none of our major findings changed.
The focus of the present study is the examination of rtuning and rnoise in A1 during both passive and active task conditions, and the corresponding effects on neural discrimination. Figure 1 displays how rtuning and rnoise interact to affect neural discrimination. Figure 1, A and B, each depict two hypothetical neurons' responses to two different stimuli. The firing rate of one neuron is plotted on the x-axis, and the other on the y-axis (i.e., the plot is the joint firing rate distribution). The distribution is plotted as a dot (mean value) within an ellipse (variance). To the extent that the ellipses of the two joint distributions overlap, that pair of neurons fails to discriminate between those stimuli because joint responses within that overlapping region can arise in response to either stimulus. The pair of neurons in Figure 1A exhibits positive rtuning and negative rnoise and the pair in Figure 1B exhibits negative rtuning and positive rnoise. The overlap between the joint responses is minimal in each of these cases. Figure 1C, D represents the relationship between rnoise and neural discrimination for pairs with positive and negative rtuning, respectively. When rtuning is positive (Fig. 1, left, C), neural discrimination decreases as rnoise increases. The ellipses in the Figure 1 insets illustrate this effect: given a constant rtuning value, increases in rnoise yield increased overlap between joint response distributions to the two stimuli, and thus poorer neural discrimination. On the other hand, when rtuning is negative, neural discrimination increases as rnoise increases. Again, the Figure 1 insets illustrate this effect by showing how the joint response distribution overlap decreases as rnoise increases.
Neural discrimination.
The ability of pairs of neurons to discriminate between nontarget and target sounds was assessed using a binomial logistic regression model, similar to that used by Jeanne et al. (2013). Binomial logistic regression is useful for making binary classifications based on a set of variables. In this case, the binary classification is between nontarget and target sounds, and the variables used to make the classification are the firing rates for each of the neurons in the pair. Thus, the model takes as inputs two firing rates and outputs a single classification prediction value between 0 (0% likely to be classified as target) and 1 (100% likely to be classified as target). We fit the model parameters for each stimulus based on data from one-half of the target trials and one-half of the nontarget trials (even or odd repetitions of each stimulus) and tested classifier performance on a trial-by-trial basis for the remaining one-half of trials. For each trial, the classifier produced a single classification prediction value corresponding to the probability that the two firing rates on that trial were in response to the target stimulus. The percentage of correct classifications for a target stimulus (neural hit rate) was simply the mean prediction classification value. The neural false-alarm rate was the proportion of classifications of nontarget stimuli as targets. Therefore, the percentage of correct classifications for the nontarget stimulus was simply 1 − (neural false alarm rate). The overall percentage of correct classifications for each pair at each AM depth was the weighted mean of these two values such that: We used the MATLAB function ‘glmfit’ with a logistic link function to implement these analyses. This link function assigns coefficients to joint firing rates using the equation: When fitting the model, y is set to 1 (target stimuli) or 0 (nontarget stimuli). X1 and X2 are the vectors of trial-by-trial firing rates for neuron 1 and neuron 2, respectively. The ‘glmfit’ function assigns the coefficients “b1” and “b2” and intercept “a” using maximum likelihood estimation. After we fit these parameters using one-half of the available trials as described above, we tested the performance of each pair by using the remaining one-half of trials to derive a prediction classification value for each pair for each trial using ‘glmval’ in MATLAB.
rnoise simulation.
Because there exist multiple aspects of a pair's activity in addition to rnoise that exhibit task-related shifts (e.g., firing rate, Fano factor, and rtuning) that could contribute to observed changes in stimulus discrimination, we sought to determine the unique contribution of rnoise. To do so, we performed a simulation where we manipulated rnoise while keeping other variables constant. This simulation artificially imposed the passive rnoise values on active recordings. If task-related changes in rnoise in the real data contributed to improved stimulus discrimination, then the imposition of passive rnoise on active recordings should decrease neural discrimination. This simulation involved 5 steps for each pair, for each stimulus. (1) For each neuron in the pair, we randomly shuffled the vector of trial-by-trial firing rates in response to the stimulus in the active condition. The result of this shuffling was that firing rate, Fano factor, and rtuning were unchanged while giving a new rnoise value. (2) Both rnoise and classification performance, as described above, were calculated. (3) Steps 1 and 2 were repeated 1000 times. (4) After 1000 shuffles, we selected the simulated rnoise value that most closely matched the actual passive rnoise value. (5) If the simulated passive rnoise for a given stimulus was not within ±5% of the actual passive rnoise, we excluded that pair/stimulus from further analysis. After performing steps 1–5 for each pair at each stimulus, we calculated an average classification performance across all pairs at each stimulus. This allowed us to compare classification performance in the active condition with classification performance under simulated passive rnoise. Any differences in classification performance could thus be attributed to observed shifts in rnoise between the passive and active conditions.
Results
Task engagement modulates rtuning and rnoise
We measured both rtuning and rnoise between pairs of A1 neurons recorded simultaneously from a single electrode during both passive and active conditions. Intuitively, rtuning can be thought of as the similarity of tuning between neurons, whereas rnoise can be thought of as the degree to which two neurons' trial-by-trial firing rate variability is correlated. Although we expected rtuning to exhibit no change between task conditions, we found that a large proportion of pairs exhibited a change in rtuning, with many even shifting sign between passive and active conditions. Among 145 pairs tested in both conditions, 27 (19%) exhibit a significant shift in rtuning, with 20 of these changing rtuning sign (examples in Fig. 2, summary in Fig. 3C). The statistical significance of shifts in rtuning was assessed via a bootstrap analysis performed on each pair. For this analysis, we estimated confidence intervals for rtuning in the passive condition and determined whether the active rtuning was outside of this interval. This analysis was performed as follows: (1) For each neuron in a pair, we resampled (with replacement) trial-by-trial firing rates in response to each of the 8 stimuli in the passive condition. (2) Based on these simulated joint firing rates, a new rtuning value was calculated. (3) This process was repeated 1000 times for each pair to create a distribution of simulated passive-condition rtuning values. Active rtuning values that were <0.1% (p < 0.001) likely to occur via random shifts in passive rtuning were considered to have significantly shifted.
This effect was surprising to us, given that studies of the behavioral effects on rnoise have not reported rtuning shifts. However, it is worth noting that Winkowski et al. (2013) found rapid shifts in rtuning with frontal cortex microstimulation. Some neurons had exceptionally large shifts in rtuning (Fig. 2A–D). In Figure 2A, two neurons' responses as a function of AM depth are shown. In the passive condition, neuron 1's firing rate (solid line) increases with increasing AM depth (increasing rate-depth function), whereas neuron 2's firing rate (dashed line) decreases with increasing AM depth (decreasing rate-depth function). Thus, the pair exhibits negative rtuning. In the active condition, however, both neurons exhibit increasing rate-depth functions, leading to positive rtuning. Figure 2B plots the joint mean firing rate distribution for each of the eight stimuli in active and passive conditions (black-edged circles represent responses to unmodulated noise). The large shift from −0.82 (passive) to 0.89 (active) is clear by observing the least-squares lines slopes. In the example shown in Figure 2C, D, both neurons exhibit increasing rate-depth functions in the passive condition, but in the active condition, the neurons have opposite rate-depth functions. Figure 2D, like Figure 2B, shows each neuron's mean firing rates in each condition and illustrates this pair's large shift in rtuning.
Across the entire population, however, we observed no significant change in median rtuning (Fig. 3A,B). However, even though there may be no statistically significant shifts in median rtuning, shifts in rtuning between individual pairs can critically impact population coding. Although Figure 3B shows no significant shift in population median rtuning, individual pairs still do exhibit shifts. Figure 2B, D provides examples of this and the magnitudes of their rtuning shifts are highlighted (purple boxes) in Figure 3C. Because many of our analyzable pairs exhibited shifts in rtuning sign (Fig. 3C, top left, bottom right), we analyzed rtuning/rnoise relationships separately in each task condition. Thus, when we analyzed rnoise effects based on rtuning sign, we grouped each pair's rnoise value independently in each task condition based on rtuning from that condition only rather than averaging rtuning across conditions.
We hypothesized that task engagement enhances population coding by selectively reducing average rnoise for pairs with positive, but not negative, rtuning. Two representative examples demonstrate this. One example shows the rnoise reduction for a pair of neurons with positive rtuning (Fig. 4), and one example (Fig. 5) shows the rnoise increase for a pair of neurons with negative rtuning. For the pair with positive rtuning both neurons have increasing rate-depth functions (Fig. 4A), and active engagement leads to a decrease in rnoise in this pair (Fig. 4C). rnoise reduction was the more common effect in our data, although many pairs increase rnoise with task engagement (85 of 145 pairs decrease rnoise with task engagement, p = 0.04, χ2 test). For the neuron pair with negative rtuning active engagement leads to an increase in rnoise (Fig. 5C). In Figure 4C, we show scatter plots for the unmodulated sound (black-edged circles) and 100% modulation (no-edged circles). The schematic ellipses are meant to demonstrate how the distributions go from more oblong (positive rnoise, left), to more circular (rnoise close to 0). Figure 5 shows an example with negative rtuning, where rnoise increases during the active condition. We assessed the significance of shifts in rnoise for individual pairs by conducting analysis of covariance (ANCOVA) for each pair using MATLAB's ‘aoctool’ function. In essence, this analysis tests for differences in the relationship between two variables (e.g., rnoise) between two conditions (e.g., task engagement). A total of 28 of 145 pairs tested in both conditions exhibit a significant decrease, whereas 16 exhibit a significant increase in rnoise (p < 0.05, corrected using false discovery rate) (Benjamini and Yekutieli, 2001). We further asked whether rtuning sign could predict the direction of significant rnoise shifts in individual pairs. Grouping pairs by either passive or active rtuning, we find that significant rnoise shifts are approximately evenly distributed for pairs with negative rtuning (grouped by passive rtuning, 8 of 18 decrease; grouped by active rtuning, 7 of 16 decrease). Significant rnoise shifts for pairs with positive rtuning seem more likely to decrease than increase (grouped by passive rtuning, 18 of 26 decrease; grouped by active rtuning 19 of 28 decrease). A χ2 test reveals this effect to be insignificant, although the direction of the trend is consistent with a selective rnoise decrease that depends on rtuning. We summarize rnoise effects across all pairs (167 passive and 199 active) in Figure 6.
We see a global decrease in rnoise with task engagement (ANOVA, p < 0.001), as well as a significant decrease in rnoise with increasing AM depth (ANOVA, p < 0.001) (Fig. 6A). Here rnoise is plotted as a function of AM depth. When analyzing the effect of task engagement on rnoise at each AM depth separately, we find the effect is only significant at 6% AM depth (p < 0.05), as assessed by a rank-sum test, corrected for multiple comparisons using false discovery rate (Benjamini and Yekutieli, 2001). When collapsing across all depths, we find that rnoise decreases from a mean value of 0.0812 in the passive condition to a mean value of 0.0472 in the active condition. The finding that task engagement reduces rnoise is consistent with previous reports that rnoise decreases globally as sensory demands increase. However, because reducing rnoise can impair population coding when rtuning is negative, we analyzed the interaction between task condition and rtuning on rnoise by including rtuning in the model. Figure 6C, D shows the effect of task engagement on rnoise for both positive and negative rtuning, respectively, for each tested stimulus. Because rtuning sign often changes between conditions, we treated the rnoise distributions for positive and negative rtuning sign as independent between conditions in our analyses. Using a 2 × 2 ANOVA (task condition × rtuning sign; ‘aov’ function in R), we see a global decrease in rnoise for positive rtuning pairs (p < 0.001; Fig. 6C,E), but no effect on rnoise for negative rtuning pairs (p = 0.53; Fig. 6D,E). Moreover, our analyses reveal a significant interaction effect (p = 0.02) wherein task engagement reduces rnoise only for positive rtuning pairs. Figure 6E illustrates how the effect of task condition on rnoise depends on rtuning sign, collapsed across tested AM depths.
Figure 7A illustrates the overall effect of task engagement on pairs of neurons' ability to discriminate target (AM) from nontarget stimuli (unmodulated noise). We used a binomial logistic regression (described in detail in Materials and Methods) to classify joint firing rates on each trial as either target or nontarget sounds. In both conditions, classifier performance increases with increasing AM depth (ANOVA, p < 0.0001; ‘aov’ function in R). Moreover, active engagement increases pairs' performance at each depth relative to passive listening (Fig. 7A). Statistical significance of this effect was determined using ANOVA with pairwise post hoc tests at each depth (Tukey's HSD, p < 0.05; ‘TukeyHSD’ function in R). Because we have previously shown that task engagement increases single A1 neuron's AM sensitivity based on firing rate alone (Niwa et al., 2012b), we sought to quantify to what extent the observed improvement in sensitivity for pairs could be uniquely accounted for by changes in rnoise. We did so by imposing passive rnoise values on the joint firing rate distributions in the active condition. To do so, we repeatedly shuffled the trial order of active recordings (1000 repetitions) to obtain new rnoise values. Of these 1000 simulated rnoise values, we selected for further analysis the shuffle that most closely approximated the passive rnoise value (hereafter known as the “simulated passive rnoise”) and calculated classifier performance for that shuffle. Thus, since shuffling does not affect rtuning, firing rate or Fano factor, we could directly assess how differences between passive and active rnoise uniquely contribute to changes in classifier performance. The simulated passive rnoise values are shown in Figure 7B–D. It is important to note that simulated passive rnoise is virtually equal to passive rnoise (0.0074% different on average); but as we have graphed them, simulated passive rnoise appears to only approximate the true passive rnoise. The ostensible difference arises from the fact that we group simulated passive rnoise by active rtuning, while we group true passive rnoise by passive rtuning. rtuning often shifts between conditions. Thus, for a given pair, simulated passive rnoise may be identical to the true passive rnoise value but will be grouped separately if rtuning shifts for that pair. We group simulated passive rnoise by active rtuning because we derive simulated passive rnoise from active recordings. If we grouped simulated passive rnoise by passive rtuning, or if rtuning was constant between conditions, the simulated passive and true passive rnoise graphs would appear virtually identical in Figure 7B–D.
Figure 7E, F represents the average difference between classifier performance for active recordings with active rnoise and active recordings with simulated passive rnoise. Values that significantly deviate from 0 (rank-sum test, p < 0.05, corrected) indicate that shifts in rnoise due to active engagement provide a unique contribution to observed shifts in classifier performance. Because we find that rnoise decreases with active engagement for pairs with positive rtuning but stays the same for pairs with negative rtuning, we predicted that the rnoise should contribute to shifts in classifier performance only for pairs with positive rtuning. Consistent with this prediction, we find that classifier performance is higher on average with active rnoise than with simulated passive rnoise, but only for pairs with positive rtuning. This effect is significant only at AM depths near or below psychophysical threshold (6%–40% AM), which may point to a selective role for rnoise reduction during especially difficult trials.
Discussion
Summary
We show that actively discriminating between AM and unmodulated sounds reduces rnoise between pairs of similarly tuned (positive rtuning) A1 neurons while leaving rnoise unaffected for pairs with dissimilar tuning (negative rtuning). Because decreases in rnoise should enhance population coding only when rtuning is positive, this result suggests that sensory systems can selectively target specific subpopulations within larger neural networks to rapidly and dynamically gate the transmission of sensory information. Although multiple studies have shown that increasing sensory demands lead to rapid, global decreases in rnoise, to our knowledge this finding constitutes the first report of a dynamic, rtuning specific reduction in rnoise. Moreover, in contrast to previous studies that assume rtuning is constant, we directly measured rtuning in multiple behavioral conditions. Because rtuning changes with behavioral state, it will be important to account for possible rtuning changes in future studies. Although the present analyses focus only on the neural coding of AM, the effect we observe wherein deleterious rnoise is reduced and beneficial rnoise is unaffected during task engagement could generalize to any stimulus feature encoded by firing rate.
Basic properties of neural correlations in auditory cortex (AC)
Our report of an average bias for similar tuning in A1 neuron pairs agrees with other reports that neighboring neurons in AC tend to exhibit similar response properties (Rothschild et al., 2010; Issa et al., 2014). It is worth noting that our observed rtuning distribution differs from that reported by both Rothschild et al. (2010) and Winkowski et al. (2013). Namely, our median rtuning values (0.49, active; 0.36 passive) are larger than those reported by Winkowski et al. (2013) (mean rtuning = 0.10) or Rothschild et al. (2010) (mean rtuning = 0.08). This could be explained by several factors. First, each used tonal stimuli (Winkowski et al., 2013; AM tones) (Rothschild et al., 2010; 50 ms tone-pips), where cells with different pure-tone best frequencies should be less correlated. Second, they recorded over distances of cortex, whereas we only recorded very close neighbors. Third, there could be species differences. Further studies on neural correlations in AC will benefit from testing multiple acoustic features with different recording techniques to glean a clearer picture of network-level feature processing in AC.
Our average rnoise values (0.08 passive, and 0.05 active), on the other hand, generally agree with those reported by others (i.e., they are small and positive). Those mentioned above also reported small, slightly positive, rnoise values (0.18, Rothschild et al., 2010; no value given for Winkowski et al., 2013). Issa and Wang (2013) also report average rnoise values between 0.08 and 0.20. Moreover, this result agrees with reports across sensory systems (for review, see Cohen and Kohn, 2011).
Behavioral modulation of rnoise in visual and AC
Although the majority of reports on the effect of attention on rnoise are in visual cortex, we provide the first report of attention-related effects on auditory cortical rnoise. Our major finding that behavior selectively reduces rnoise based on rtuning has not been reported in visual cortex, at any level of the system. Instead, most have reported that the shift from an inattentive to attentive state leads to global reductions in rnoise, regardless of rtuning (Cohen and Maunsell, 2009). However, Cohen and Newsome (2008) have previously reported that pairs of neurons in area MT exhibit task-related shifts in rnoise when animals detect dot motion relative to when they fixate, and these shifts depend on both the tuning similarity between the neurons and whether each neuron contributed to the same, or different, perceptual decisions. Namely, they found that pairs most often decreased rnoise during behavior, but when pairs had very dissimilar tuning (which would correspond to very negative rtuning), rnoise increased during behavior, but only if the neurons of the pair contributed to opposite dot motion decisions. During another task condition in which both neurons contributed to the same dot motion decision, they observed a decrease in rnoise. Although We did not directly measure rtuning, their result suggests that rtuning-dependent shifts in rnoise are present in visual cortex as well as AC. However, a dearth of studies in which rtuning is directly measured contributes to an incomplete picture of how behavior affects rnoise.
Behavioral modulation of rtuning
Here we provide perhaps the first report of behavioral shifts in rtuning. Although we observe no shift in the median rtuning value between task conditions, we observe that pairs commonly exhibit shifts in the sign of rtuning in the transition between passive listening and active sound detection. However, although others have not reported on this phenomenon, we think it is likely that behavioral variables commonly shift rtuning, across sensory cortical fields. Attention has been shown to modulate single neuron tuning in AC and VC not just via gain modulation, but also by shifts in receptive fields (Fritz et al., 2003; David et al., 2008). Given two single neurons' receptive fields shifting in response to task demands, it follows that rtuning can often change, sometimes dramatically. Thus, our finding that attention can shift rtuning values is supported by changes in tuning of single neurons. Given that the effect of rnoise depends on the sign of rtuning, it will be worthwhile for future studies to investigate how attention affects each of these measures.
Issues with global rnoise decreases
Theoretical studies have established that positive rnoise can aid in neural discrimination (Oram et al., 1998; Averbeck et al., 2006; Ecker et al., 2011). Empirical studies have affirmed these ideas. Romo et al. (2003) observed that positive rnoise reduces neural discrimination thresholds for pairs with opposite tuning (which would yield a negative rtuning value). They recorded pairs of S2 neurons, a cortical field in which single neurons commonly exhibit either increasing or decreasing firing rate functions in response to increasing tactile vibration frequency. Similarly, neurons in A1 (and more so in middle lateral belt) exhibit this “dual coding” scheme for AM depth (Niwa et al., 2013). A potential goal of future research will be to characterize differences in the behavioral effects of rnoise for populations with dual coding relative to those with Gaussian feature tuning, as different mechanisms may be at play in shifting rnoise. Jeanne et al. (2013) have also reported that pairs with negative rtuning benefit from positive rnoise. They trained starlings to map specific behavioral responses to specific starling vocalization. Recording from the caudolateral mesopallium (CLM) they found that learning not only reduced rnoise when rtuning is positive but increased rnoise when rtuning is negative. The stimuli used in their study (starling vocalizations) are considerably more complex than those used here or by Romo et al. (2003). CLM has been hypothesized to specialize at representing learned vocalization stimuli (Gentner and Margoliash, 2003; Gill et al., 2008). It may be that Jeanne et al. (2013) observed a more powerful effect in CLM than we do in A1 because, although CLM is specialized for learned vocalizations, AM is one of many sound features to which A1 neurons respond (deCharms, 1998; Wang et al., 2005; Chambers et al., 2014). Future studies may manipulate stimulus conditions orthogonally to behavioral variables while recording from neural populations to build a more complete picture of the functional characteristics of neural correlation structure.
Footnotes
This work was supported by National Institutes of Health National Institute on Deafness and Other Communication Disorders Grant DC002514 to M.L.S., National Institutes of Health National Research Service Award fellowship F31DC008935 to M.N., and National Science Foundation GRFP fellowship 1148897 to J.D.D. We thank Kevin O'Connor, Doug Totten, and Jessica Verhein for comments on previous versions of the manuscript; and Brittany Rapone for assistance in figure creation.
The authors declare no competing financial interests.
- Correspondence should be addressed to Dr. Mitchell L. Sutter, Center for Neuroscience, University of California, Davis, 1544 Newton Court, Davis, CA 95618. mlsutter{at}ucdavis.edu