Attention Rhythmically Shapes Sensory Tuning

Main task

Each session was composed of four blocks of 126 trials each as well as supernumerary trials in case of fixation break (see above, Eye-tracking). Central fixation remained on the screen during the entire block. After a 1,000 ms stable fixation, a precue appeared for 60 ms followed by a 400 ms interstimulus interval (ISI). The two Landolt-Cs were displayed for 60 ms (one in each lower quadrant) along with a response cue indicating which of the two was the target. Participants were instructed to discriminate the position of the C gap (randomly presented on the left or the right) of the target Landolt-C (their response was collected at the end of the trial sequence). When the response cue corresponded to the precued location (2/3 of the trials), the trial was valid; when the response cue corresponded to the un-precued location (1/3 of the trials), the trial was invalid. Then a second, variable ISI (randomly chosen among 14 different possibilities equally distributed between 40 and 495 ms) preceded the presentation of two new stimuli: two patches of noise or noise + Gabor for 33 ms, at the same location as the previous Landolt-Cs, followed by a 60 ms delay. Participants were instructed to report the presence versus absence of the Gabor inside each patch of noise. The presence or absence of a Gabor in either patch within a trial was independent and random (50%), such that one, the other, both, or neither having a Gabor embedded were equally likely. These patches presented at multiple delays after the previous Landolt-Cs were used to probe the state of attention across time (Dubois et al., 2009; Dugué et al., 2015b; Senoussi et al., 2019), i.e., attention deployed in the Landolt-C task modulates probe report performance over the course of its sustained orienting (valid condition) or reorienting (invalid, endogenous reorienting, as its orienting, is not effective right from the onset of the response cue; Jonides, 1981; Muller and Rabbitt, 1989; Nakayama and Mackeben, 1989; Cheal and Lyon, 1991; Liu et al., 2007; Carrasco, 2011). At the end of the trial, participants were first asked to report the position of the gap in the target Landolt-C (press C key for a gap on the left with the left hand, and press B key for a gap on the right with the right hand; maximum response time of 1,500 ms) and then to report the presence or the absence of the Gabor inside each patch (for left patch, press C key for Gabor present and D key for Gabor absent with left hand; for right patch, press B key for Gabor present and H key for Gabor absent with right hand; maximum response time 2,000 ms).

Staircases

Two independent staircases were implemented via a Best PEST procedure (Pentland, 1980) using the Palamedes toolbox (Prins and Kingdom, 2018) in Matlab, to adapt first the size of the Landolt-C gap in the discrimination task and second the SNR [signal (Gabor)-to-noise ratio] for the detection task. In the first staircase (208 trials), only the discrimination task sequence was presented and the size of the Landolt-Cs gap (1.4 ± 0.4° on average across participants) was adjusted. This procedure (Best PEST) aimed at finding the gap size that maximized the likelihood of a cumulative normal distribution. The gap size selected is that which corresponded with 75% accuracy of the cumulative normal model. Here the trial sequence consisted of a neutral precue (both the left and right precue) for 60 ms, a 400 ms ISI and the Landolt-Cs together with the response cue for 60 ms. Participants had 1.5 s to report the location of the gap. In the second staircase (200 trials), the procedure was the same but only the detection task sequence was presented. The Best PEST procedure aimed at finding the Gabor-to-noise ratio that maximized the likelihood of the Weibull distribution. The Gabor-to-noise ratio selected was the one that matched 62% accuracy of the Weibull model. In this staircase, a trial was defined as correct when the response to both patches (left and right) was correct. One trial consisted of neutral precue for 60 ms, a 400 ms ISI, and the two probe patches (50% noise only and 50% noise + Gabor, at each location) with a response cue for 33 ms. Participants had 2 s to report the presence or the absence of the Gabor in each of the two patches. Note that it is the Gabor-like properties of the noise and its orientation, not the Gabor per se, that is crucial for the analysis. Specifically, participants are attempting to detect the presence of a signal composed of a specific orientation and frequency, which will happen to be strongly represented in the random noise on certain trials. During analysis, we relate the noise properties on each trial with the propensity to report the presence versus absence of the Gabor, such that both correct and incorrect responses provide information on sensory tuning (see below, Reverse correlation analysis). Therefore, we aimed for a lower percent correct in the detection task, such that the Gabor was difficult to detect and thus participants’ responses are more likely to be influenced by the orientation content of the noise on any given trial.

Attentional manipulation

To ensure that attention was successfully manipulated (Carrasco, 2011; Dugué and VanRullen, 2014; Dugué et al., 2016, 2019; Senoussi et al., 2019), we assessed performance in the discrimination task as per d-prime (Hanning et al., 2019; Eq. 1):d−prime=z(Hit)−z(FA),(1) with z(Hit) and z(FA) as the z score for hit and false alarm rates. We also checked reaction times from Response 1 screen onset (Fig. 1) to rule out speed–accuracy trade-off (Wickelgren, 1977; Bogacz et al., 2010; Heitz, 2014; Huang et al., 2017).

• Download figure
• Open in new tab
• Download powerpoint

Figure 1.

Attentional manipulation. A, Trial sequence. Trials began with a central fixation period of 1,000 ms. An endogenous precue appeared for 60 ms. After 400 ms, Landolt-Cs were displayed for 60 ms together with a response cue indicating which Landolt-C was the target (2/3 of valid trials, i.e., target at precued location). Each Landolt-C could be oriented to the right or to the left randomly. After a variable delay from 40 to 495 ms (35 ms step; only the fixation circle remained on the screen), two patches were presented for 33 ms at the same location as the Landolt-C. Each patch could either show noise only (1/2 of trials) or noise + Gabor (1/2). After a third ISI of 60 ms, participants were instructed to report: (1) the orientation of the target Landolt-C (right/left; discrimination task) and (2) the presence or absence of a Gabor in each of the two patches. B, Behavioral results of the discrimination task. Individual (gray) and averaged (black) d-primes and median reaction times (RT) in the valid and invalid conditions. Error bars: standard error of the mean. C, Behavioral result of the Gabor detection in noise patches. Individual (gray) and averaged (black) d-primes. Error bars: standard error of the mean.

The validity effect (comparing performance between valid and invalid trials) was first assessed for each individual participant. D-prime was computed for each session and nonparametric statistics were performed: Wilcoxon signed rank test for paired-sample statistical comparison of valid and invalid d-prime. Out of the 15 participants, four participants did not show significant difference between valid and invalid trials, suggesting that the attentional manipulation was not successful. The following analyses were thus conducted on the remaining 11 participants (all correct and incorrect trials were included). We report the following effect sizes (Eq. 2):r=zscoren,(2) computed following Rosenthal recommendations (Rosenthal, 1994) with z score from the Wilcoxon signed rank test and n the number of observations.

Behavioral rhythms

For the detection task, d-primes were examined for each cueing (valid and invalid) and ISI (delay between the Landolt-C and the probe patches) condition (note that d-primes were normalized for visualization purposes only; Fig. 2; all analyses were performed on non-normalized d-prime). Responses to both patches were combined in order to increase the number of trials for reverse correlation analysis (and because no difference was found between the two patches locations; analysis not shown). In total, there were 720 trials per ISI and participant. Fast Fourier transform decompositions (FFT) were further performed on averaged d-primes to assess possible periodicity. Amplitude spectra for valid and invalid trials were plotted from 2.04 to 14.29 Hz with steps of 2.04 Hz (eight different frequencies). Phase angles were extracted from FFTs performed on each participant’s data for 12.2 Hz in valid and 6.1 Hz in invalid. We selected high- and low-performance trials by taking local maxima (above d-prime average) and minima (below d-prime average) for each of the d-prime time courses (ISI valid high: 145, 285, 390 and 460 ms; low: 75, 180, 320, 425 and 495 ms; ISI invalid high: 110, 320 and 425 ms; low: 40, 145, 355 and 460 ms). Two local maxima were removed, one in each condition, because they were below the averaged d-primes. All analyses described below were done separately for the high- and low-performance trials.

• Download figure
• Open in new tab
• Download powerpoint

Figure 2.

Behavioral rhythms. A, Individual (light traces) and averaged (dark traces) d-prime time courses (n = 11) for valid (blue) and invalid (orange) conditions separately. Shaded area: 95% confidence interval. B, Amplitude spectra obtained from FFT of individual (light traces) d-prime time courses and from FFT of averaged d-prime time courses across participants (dark traces). Dotted line: p = 0.001 (permutation test for averaged d-prime time course). Polar histogram of individual phases (light bins) and weighted phase averages (arrows) at 12.2 Hz in the valid condition and 6.1 Hz in the invalid condition. Bar plots: number of participants at each frequency with the highest peak in individual amplitude spectra.

Note that valid and invalid conditions do not have the same number of trials. Also, the ISI selection detailed above shows differences between validity condition and high/low performance. To assess the potential impact of such differences in trial number, all previous analyses for both discrimination and detection tasks were also performed on subsampled trials (100 repetitions, random selection), thus equalizing trial numbers across all conditions. The results were the same as without subsampling (analysis not shown). We thus present the results from nonsubsampled data.

All previous analyses were similarly performed on the decision criterion to assess whether the observed periodicity is exclusively on sensitivity or can also be observed on decision bias. Since in our task criterion is not independent of d-prime, we calculated criterion-prime (Eq. 3) as follows:criterion−prime=(z(Hit)+z(FA)−2)z(Hit)−z(FA),(3) which consider d-prime variation to keep the decision boundaries stationary. No periodicity was observed in criterion-prime (analysis not shown).

Reverse correlation analysis

To address behavioral feature tuning, we first transformed the noise from luminance intensity into the energy of different orientation components. We converted the noise image (either noise only or noise + Gabor) of each trial from pixel space (luminance intensity of each pixel) to a 1D space defined by the noise energy of components responding to different orientations (varied across all of orientation space, −90° to 90°, in steps of 7.5°, for 25 points on a linear scale). To compute the energy of each component, we took the noise image S of each trial and convolved it with two grating filters (gθ,sin in sine phase and gθ,cos in cosine phase) with the corresponding orientation. The energy was computed as follow:Eθ=(S*gθ,sin)2+(S*gθ,cos)2,(4) in which * represents the cross-correlation operator. We took the energy centered at the test stimulus (vertical orientation) for analysis. For each component with preferred orientation θ, we estimated the correlation between the energy of that component and behavioral responses using a GLM to predict the binomial dependent variable:p(yes)=ψ(βθEθ*+bθ),(5) in which p(yes) is the percentage of yes responses in the detection task and ψ is a cumulative normal distribution. Two free parameters βθ and bθ are fitted. βθ represents the correlation between the energy and the behavioral response. A zero βθ indicates that the energy of that component did not influence participants’ responses. b_θ represents a baseline tendency of the participants to respond “present” (i.e., false alarm rate), which was not related to the stimuli. Eθ* represents the centered and normalized energy of each component. Before applying the GLM, the energy of each component was first sorted into two groups based on whether the target signal was present or absent in each trial within the noise image, and the mean of the energy was removed for each group. Subtracting the mean stimulus energy for target Gabor present and target Gabor absent trials independently serves two purposes: (1) it removes Gabor stimulus energy in present trials, so that the noise energy can be assessed independent of Gabor energy; (2) it centers stimulus energy for all trials to the mean, thus allowing straightforward correlation of the responses (1 for present, −1 for absent) with stimulus energy—where positive values indicate greater-than-average stimulus and negative values indicate lower-than-average stimulus energy for each orientation channel. Reverse correlation aims to quantify which aspects of the stimulus during any given trial contribute to a higher or lower likelihood of reporting “present” irrespectively of the presence or the absence of a Gabor. Therefore, it is important to assess the difference of the noise on a given trial from the mean across all trials and correlate that with behavior. If stimulus energy of a particular orientation is higher than average, and “present” responses are more likely, then the correlation is positive, and the opposite is true if high energy is associated with less likely “present” responses. To let the estimated βθ be comparable across components, we further normalized the energy across all the trials in each component to have a standard deviation of 1. The estimated sensitivity kernel was a 1D matrix K in which K(θ)=βθ . Before Gaussian fitting, positive and negative orientations’ energy were averaged. A Gaussian tuning curve was fit to the data (Eq. 6) for each condition, where θ represented all noise orientation channels (25 orientations).x[1]*(exp(−(θ−x[2])22*x[3]2))x[4]+x[5],(6) As we were agnostic as to the precise shape of the tuning function, in addition to the single Gaussian fit (Eq. 6), we also performed a double Gaussian fit (Eq. 7). The single fit corresponds to a simple, standard model of feature tuning, whereas the double fit has been associated with a “Mexican hat” shape that sometimes arises due to attention (Müller et al., 2005).x[1]*(exp(−(θ−x[2])22*x[3]2))x[4]+x[5]+x[6]*(exp(−(θ−x[2])22*x[7]2))x[8]+x[9].(7) AIC corrected (AICc) for small data samples were performed using Python (3.8.1; Van Rossum and Drake, 1995) independently for each condition to compare single and double Gaussian fits. AICc criterion was computed as shown in Equation 8 where k was the number of parameters, L the maximized value of the likelihood function, and n the number of participants.2k−2ln(L)+2k(k+1)n−k−1.(8) For each condition, the single Gaussian fit better explained the results [valid high-performance AICc difference (single fit − double fit) = −16.03; valid low-performance AICc difference = −37.34; invalid high-performance AICc difference = −37.34; invalid low-performance AICc difference = −37.33]. All further analyses were thus performed with the tuning curve from the single Gaussian fits.

We assessed the statistical differences between high- and low-performance trials’ tuning curves using linear mixed model with R (4.2.2; (R Core Team, 2022)), and specifically, differences in the standard deviation, amplitude, and baseline of the tuning curves. Linear mixed models were fitted by maximizing the restricted log likelihood and using a normal distribution function. Performance was used as predictor with two levels (high and low) and participants as random effect. Nonparametric, paired-sample Wilcoxon signed rank tests were further performed to compare, for each orientation valid, high- and low-performance decision weights (FDR correction for multiple comparisons was applied).

Similar to the d-prime analyses, we assessed the potential impact of differences in trial number using a subsampling procedure (100 repetitions, random selection). The results were the same as without subsampling (analysis not shown). We thus present the results from nonsubsampled data.