Abstract

Studies by cognitive psychologists, psychophysicists, neuroscientists, and economists provide ample evidence that humans use prior knowledge to bias decisions adaptively. In this study, we sought to locate and investigate the brain areas mediating this behavior. Participants viewed ambiguous abstract shapes and decided whether a shape was of Category A (smoother) or B (bumpier). The decision was made in the context of one of two prior knowledge cues, 80/20 and 50/50. The 80/20 cue indicated that upcoming shapes had an 80% probability of being of one category, for example, B, and a 20% probability of being of the other. The 50/50 cue indicated that upcoming shapes had an equal probability of being of either category. The shift in bias produced by the 80/20 cue relative to the 50/50 cue was of the predicted sign for every subject but varied in magnitude. We searched for brain regions in which activity changes correlated with the extent of the bias shift; these were dorsolateral pFC (middle frontal gyrus), inferior frontal junction, anterior insula, inferior parietal lobule, intraparietal sulcus, head of the caudate, posterior cingulate cortex, and fusiform gyrus. The findings indicate that an individual's brain activity in these regions reflects the extent to which that individual makes use of prior knowledge to bias decisions. We also created within-ROI tuning curves by binning the shape curvature levels and plotting brain activity levels at each of the nine bins. In the fronto-parietal and anterior insula ROIs, the tuning curves peaked at targets contraindicated by the prior knowledge cue (e.g., Category B targets if the 80/20 cue meant 20% probability B). The increased activity in these regions likely indicates a no-go response when sufficient perceptual evidence favored the alternative contraindicated by the 80/20 cue.

INTRODUCTION

When making decisions, people take advantage of available prior knowledge to help them make choices. For example, consider a prior knowledge condition indicating that Alternative 1 has an 80% and Alternative 2 has a 20% chance of being the correct choice—we will call this an 80/20 prior knowledge condition and a 50/50 prior knowledge condition, indicating that Alternatives 1 and 2 each have a 50% chance of being the correct choice. Unsurprisingly, human participants tend to choose Alternative 1 more often when the prior knowledge condition is 80/20 than when it is 50/50 (Healy & Kubovy, 1978, 1981; Green & Swets, 1966). This common sense behavior is adaptive, as it increases the chance that decisions will be correct. In fact, for any two-alternative prior knowledge condition, an optimal ratio of Alternative 1 choices to Alternative 2 choices may be calculated and converted to a quantity called an optimal bias (Wickens, 2002; Bohil & Maddox, 2001; Green & Swets, 1966).

It is known that the degree to which decision bias in prior knowledge tasks approximates optimality varies widely across individuals (Healy & Kubovy, 1978; Green & Swets, 1966). Here, we sought to exploit that variability to locate the brain areas mediating the use of prior knowledge to bias decisions adaptively. To address this problem experimentally, we probed the behavioral and fMRI responses of human participants viewing ambiguous abstract shapes and deciding whether a shape was of Category A (smoother) or B (bumpier). The decision was made in the context of one of two prior knowledge cues, 80/20 and 50/50. The 80/20 cue meant that upcoming shapes had an 80% probability of being of one category, for example, B, and a 20% probability of being of the other. The 50/50 cue meant that upcoming shapes had an equal probability of being of either category. Participants learned the meaning of the cues in pre-scan training runs. During training, the 80/20 and 50/50 cues were accompanied by 80/20 and 50/50 target distributions, respectively; the training distributions were created by manipulating the prior probability of occurrence of the physical targets themselves, rather than changing the category boundary. To quantify each individual's ability to adapt flexibly to different prior knowledge conditions, we measured each individual's bias shift—the difference between the decision biases adopted in the two conditions. To locate the brain regions in which individuals' brain activity reflected the extent to which that individual makes use of prior knowledge to bias decisions, we searched for brain regions in which activity changes correlated with the extent of the bias shift. We also created within-ROI tuning curves by binning the shape curvature levels and plotting brain activity levels at each of the nine bins. This approach revealed different activity patterns across the ROI set. In the fronto-parietal and anterior insula ROIs, the tuning curves peaked at targets contraindicated by the prior knowledge cue (e.g., Category B targets if the 80/20 cue meant 20% Probability B). In the head of the caudate, posterior cingulate cortex, and fusiform gyrus (FuG), the tuning curves were flat. The greater activity for contraindicated than indicated targets in the fronto-parietal and anterior insula ROIs suggests that these regions contribute to the decision using a mechanism similar to a no-go response. In this scenario, the prior knowledge induces a default behavioral response in favor of the indicated (80%) alternative; the role of the fronto-parietal and anterior ansula regions is to overcome that default response when sufficient perceptual information favors the contraindicated (20%) alternative.

METHODS

All procedures were approved by the National Institute of Mental Health Institutional Review Board. Some analyses of these fMRI and behavioral data were published previously (Hansen, Hillenbrand, & Ungerleider, 2011); the stimuli and subjects were described in detail in that publication. Briefly, we present data from 22 participants (nine men) with mean age of 25.6 years (range = 20–41 years). Targets were circles distorted by a sinusoidal modulation (Wilkinson, Wilson, & Habak, 1998) ranging linearly from 4% to 22% of the mean radius, with a step size 0.5%. The targets were presented one at a time with random sizes, orientations, and locations. The smoothest target was defined as the Category A prototype, and the bumpiest as the Category B prototype (Figure 1). Distributions of Categories A and B shapes were integer approximations of overlapping Gaussians (Maddox, 2002; Healy & Kubovy, 1981) with means of 10.5% (A) and 15.5% (B) and d′ = 2; for an illustration, see Figure 1A of Hansen et al. (2011).

Figure 1. 

Decision behavior during fMRI. (A) Reproduced from Hansen et al. (2011). Red and dark blue: reports from participants trained that “80/20” meant 80% probability of B and 20% probability of A. Light blue and orange: reports from participants trained that “80/20” meant 80% probability of A and 20% probability of B. The dotted horizontal line indicates chance performance. The shift in the curves from dark blue to red and from light blue to orange shows that participants reported targets as the indicated category given evidence farther from the indicated target prototype when the cue was 80/20 than when the cue was 50/50. (B) Data from A, broken into the first versus the second half in chronological order of acquisition. Responses were consistent over the course of the experiment, showing that experience with the experimental stimuli during scanning (where the true proportion was 50/50 for both cues) did not alter the bias learned during training (where the true proportion matched each cue).

Figure 1. 

Decision behavior during fMRI. (A) Reproduced from Hansen et al. (2011). Red and dark blue: reports from participants trained that “80/20” meant 80% probability of B and 20% probability of A. Light blue and orange: reports from participants trained that “80/20” meant 80% probability of A and 20% probability of B. The dotted horizontal line indicates chance performance. The shift in the curves from dark blue to red and from light blue to orange shows that participants reported targets as the indicated category given evidence farther from the indicated target prototype when the cue was 80/20 than when the cue was 50/50. (B) Data from A, broken into the first versus the second half in chronological order of acquisition. Responses were consistent over the course of the experiment, showing that experience with the experimental stimuli during scanning (where the true proportion was 50/50 for both cues) did not alter the bias learned during training (where the true proportion matched each cue).

The task was to decide whether a shape was Category A or B. The shapes were presented one at a time for 200 msec with random sizes, orientations, and locations to encourage the use of stimulus shape to make decisions and to prevent participants from relying on retinotopic location or spatial attention to perform well. No part of any shape subtended more than two radial degrees, and the location of the fixation cross was inside each shape. A 500-msec cue was presented 500 msec before each shape. The same cue was used throughout each run. To ensure that the participant would not forget which decision-making context was applied in the current run, we used the explicit numeric prior probability ratio—80/20 or 50/50—as the cue. During scanning, the cue was 50/50 on six runs and 80/20 on six runs, with the cue type alternating pseudorandomly from run to run. Before entering the scanner, the participants underwent behavioral training. Because subtle variations in instructions are known to produce spurious variations in decision bias (Healy & Kubovy, 1978), the instructions were presented in writing and included questions to confirm that the participant understood the cues and the task. Training runs included explicit prior knowledge cues—that is, cues informing participants of the probability that targets were of either category—of 80/20 and 50/50.

In this article, the terms ‘indicated category’ and ‘contraindicated category’ refer to the categories associated with 80 and 20, respectively, during the 80/20 training runs and the terms ‘indicated target’ and ‘contraindicated target’ refer to targets with curvature smoother (or bumpier) than the mean sinusoidal modulation of 13% if the indicated category was A (or B). The indicated category was A for 8 participants and B for 14 participants. The 80/20 training runs were consisted of 80% indicated and 20% contraindicated targets. The 50/50 runs were consisted of 50% of each target type. Thus, during training, the explicit prior knowledge cues reflected the implicit—that is, the true—prior probability distributions of the targets. Participants received feedback after each training trial. The scanning runs differed from training runs in three respects. First, all scanning runs were consisted of 50% indicated and 50% contraindicated targets, such that the targets in each 80/20 run were identical to the targets in a 50/50 run. This control ensured that differences between prior knowledge conditions could be attributed only to the cue and not to stimulation differences. Second, participants did not receive feedback during scanning. Third, in one third of the trials in each scanning run, a blank screen took the place of the target, and participants were instructed to make no response. The inclusion of these trials permitted us to obtain estimates of activity during decision trials versus nondecision trials.

All MRI data were collected on a GE 3-T scanner with a GE whole-head eight-channel coil. For fMRI, we used an EPI sequence with repetition time of 2.5 sec/shot (2.5 sec/acquired brain volume), echo time of 30 msec, field of view of 22 cm × 22 cm, resolution of 64 × 64 voxels per slice (in-plane voxel size of 3.4 × 3.4 mm), and slice thickness of 3.0 mm. Each fMRI brain volume consisted of 38 axial slices. For anatomical images, we used an magnetization prepared rapid acquisition gradient-echo sequence with field of view of 24 × 24 cm, 128 locations per slab, and slice thickness of 1.2 mm. Unless otherwise noted, preprocessing and subsequent analysis of the MRI data were performed with the AFNI software package (Cox & Hyde, 1997; Cox, 1996). Each subject's T1-weighted anatomical data set was warped via 12-parameter affine transform to a single template volume (the N27 “Colin” brain).

We used the behavioral data (Figure 1) to calculate behavioral criterion values (Figure 2) as λ = −½ [Z(f) + Z(h)] (Wickens, 2002). The criterion values reported for the ideal observer would produce response ratios equivalent to the ratio of expected target types, that is, 80/20 or 50/50. We calculated voxelwise across-subject correlations between the magnitude of the shift of the decision bias between cue conditions and the magnitude of fMRI activity changes during the trials driving the bias shift. The first step was to define regressors from the behavioral data. Report change trials—that is, the trials that drove the bias shift—were defined for each participant by identifying those in which the participant decided that a particular target was of the indicated category when the cue was 80/20 and of the contraindicated category when the cue was 50/50. The number of report change trials from the participant who produced the fewest was 24. Therefore, to correct for the relationship between the magnitude of the bias shift and the number of report change trials, we limited the number of report change trials for all participants to the 24 trials per prior knowledge condition with targets nearest the contraindicated target prototype. The logic for using the same number of report change trials for each participant was as follows: Large and small bias shifts necessarily imply large and small numbers of report change trials, respectively. In general linear model (GLM) analysis, large numbers of trials tend to produce more accurate beta weight estimates than smaller numbers of trials. Therefore, if more trials were used for participants who produced more report changes, the accuracy of beta weight estimation could covary with behavioral bias shift magnitude. Such covariance would imply that the beta weights from participants with smaller shifts were more variable (noisier) than the beta weights from participants with larger shifts, potentially confounding the interpretation of the results.

Figure 2. 

Bias shifts in individual participants. (A) Measurements of the decision bias when the cue was 80/20 (red) and 50/50 (blue). Although the direction of every subject's bias shift between conditions was in the predicted direction, there was substantial across-subject variance in the magnitude of the bias shift. (B) Magnitudes of the bias shift, calculated as (observed 80/20 criterion − observed 50/50 criterion)/(optimal 80/20 criterion − optimal 50/50 criterion).

Figure 2. 

Bias shifts in individual participants. (A) Measurements of the decision bias when the cue was 80/20 (red) and 50/50 (blue). Although the direction of every subject's bias shift between conditions was in the predicted direction, there was substantial across-subject variance in the magnitude of the bias shift. (B) Magnitudes of the bias shift, calculated as (observed 80/20 criterion − observed 50/50 criterion)/(optimal 80/20 criterion − optimal 50/50 criterion).

Note that the report change trials used for each participant were precisely those trials in which we had the most reliance that the prior knowledge was the cause of the behavioral report change, because the nearer a target was to the contraindicated target prototype, the less likely it was that a report of the indicated category was due to noise. More of the report change trials should be due to prior knowledge—as opposed to random behavior—in participants with larger shifts than in participants with smaller shifts, assuming that the number of report change trials due to random behavioral responses is constant across participants. This assumption is reasonable, given that we observed no correlation with d-prime.

In each subject's report change trials regressor, report change trials were given a value of 1, and all other trials were given a value of 0. The sequence of 0s and 1s was convolved with a model hemodynamic function to create a regressor for a GLM analysis. Other inputs to the GLM were the matched expectation and overall decision regressors, explained below, and head motion estimates. The GLM analysis modeled the 80/20 and 50/50 data separately. Outputs were voxelwise beta weights representing the percent signal change versus baseline attributable to each regressor. A baseline was obtained by identifying activity modeled by a constant, linear, and quadratic drift and signal variability accounted for by head motion estimates.

Matched expectation trials were identified for each participant as those in which the curvature levels of the target were identical to those on that subject's report change trials, but the behavior was not as predicted. Nonpredicted behaviors could include reporting the indicated category in both cue conditions, reporting the contraindicated category in both cue conditions, or reporting the indicated category at 50/50 and the contraindicated category at 80/20. The GLM analysis and across-subject multiple regression were run as described above.

We also performed a simple subtraction analysis to test for effects of the 80/20 vs. 50/50 prior knowledge cue on the fMRI signal when participants' behavior was not taken into account. For each subject, the decision trials that had not been used in the report change or matched expectation regressors were used to create the overall decision regressor to the GLM analysis described above. The resulting decision activations at 80/20 and 50/50 were used as input in a two-factor ANOVA (random effect of participants, fixed effect of prior knowledge cue). No 80/20 versus 50/50 prior knowledge cue effects survived cluster correction.

The differences between the beta weights from 80/20 runs versus 50/50 runs for each participant were used as inputs to a multiple regression analysis. The other inputs to the multiple regression were the 80/20 versus 50/50 differences in behavioral bias. The analysis calculated voxelwise correlations between the size of the behavioral bias shift across prior knowledge cue conditions and the size of the fMR signal change across prior knowledge cue conditions during the report change trials. From the correlation results, a mask was derived identifying voxels where the correlation and the mean activity levels at either 50/50 or 80/20 each exceeded uncorrected p < .05. Taking account of the mean activity levels ensured that the results would reflect differences between activations, not differences between deactivations. A cutoff for significant cluster size (corrected p = .05) was determined based on an uncorrected p value of <.0025, the intersection of the correlation and mean activity p values, and took into account the smoothness of the original fMRI data sets. Talairach coordinates for the surviving clusters were determined by affine registration to the TT-N27 brain template, and Brodmann's area equivalents were derived from the Talairach–Tournoux atlas (TT-Daemon).

To derive tuning curves from the within-ROI data, we sorted the trials by curvature level into nine bins ranging from smoothest to bumpiest. We performed a separate GLM analysis for each participant and prior knowledge condition, estimating fMRI responses to the presentation of targets within each bin. Nine sequences of 0s and 1s, wherein the 1s represented targets in a given bin, were convolved with a model hemodynamic function to create the regressors for the analysis. Other inputs to the GLM were the estimates of head motion. Outputs were voxelwise beta weights representing the percent signal change versus baseline attributable to each regressor. Signal variability attributable to head motion estimates was assigned to the baseline.

For each subject, the ROIs derived from the report change trial correlation were converted to individual brain space. The betas corresponding to each subject's fMRI responses to each of the nine curvature bins at 80/20 were sampled from and averaged within each individual ROI. The grand means and standard errors across subjects were calculated for each bin and prior knowledge condition, and the results were plotted as tuning curves across the dimension of curvature bins.

RESULTS

Behavior

As reported earlier (Hansen et al., 2011), the behavioral data acquired during fMRI data acquisition indicate that the prior knowledge cues learned during training (see Methods) biased participants' decisions as predicted. Namely, participants responded “B” (or “A”) for a given shape more often when an 80/20 cue indicated Category B (or A) than when the cue was 50/50. This was despite the fact that the fMRI runs with the 80/20 cue used the identical shapes as the fMRI runs with the 50/50 cue, such that the actual proportions of category B/A in the fMRI runs were always 50/50. In Figure 1A, this bias effect is shown as a shift to the left or to the right; that is, for a given curvature, the participants were more likely to classify a given stimulus as one category when the cue was 80/20 indicating that category than when the cue was 50/50. The size of this shift did not decrease from early to late trials (Figure 1B, solid vs. dashed lines), demonstrating that experience with the stimuli during fMRI scanning did not alter the bias learned during training. Participants skipped very few trials; the average percent of reports out of total decision trials was 99.6%.

To quantify the leftward and rightward shifts in the 80/20 versus 50/50 decision reports of individual participants, we used the 480 trials per cue condition (960 total) collected for each participant during scanning to calculate a decision criterion for that cue condition. The decision criterion is a quantitative metric of bias, representing the amount of supporting evidence required for the participant to make a decision in favor of one alternative over another. The resulting individual criterion measurements (Figure 2A) showed that the decision criterion shifted in the predicted direction in every subject. Therefore, the prior knowledge cues learned during training biased participants' decisions at the individual level as well as at the group level. The individual criterion measurements also reflected substantial variance in the degree to which participants were biased by the cues; participants' performance ranged from near optimal to very suboptimal (Figure 2B). Note that the participants with smaller shifts did not uniformly treat the 80/20 cue as 50/50, which would have resulted in a cluster of criterion measurements near the optimal 50/50 level in both cue conditions. Instead, the participants with smaller shifts produced widely varying criterion levels.

Imaging Data: Report Change Correlations and ROI Identification

To quantify each subject's behavior, we used the magnitude of the bias shift between the 80/20 and 50/50 conditions, exploiting the substantial across-subject variance observed in the behavioral data (Figure 2B). To obtain a metric of fMRI signal change, we calculated for each individual participant the difference between the within-ROI fMRI signal level during the 80/20 versus 50/50 trials that drove the shift in that subject's behavioral curves. These trials, which we called report change trials, were those in which the participant reported that a given target was of the indicated category during an 80/20 run and that the identical target was of the contraindicated category during a 50/50 run.

To identify brain regions mediating the use of prior knowledge to bias decisions, we searched for clusters of brain activity changes that correlated with the extent of the bias shift across our individual participants. Clusters passing a cutoff size for significance were identified in dorsolateral pFC (middle frontal gyrus), inferior frontal junction (IFJ), anterior insula, inferior parietal lobule (IPL), intraparietal sulcus (IPS), head of the caudate, posterior cingulate gyrus, and a focal region of FuG. The cluster locations are shown in Figure 3, and Talairach coordinates are provided in Table 1.

Figure 3. 

Brain regions where activity changes during report change trials correlated with size of bias shift. Axial sections are ordered from ventral (z = −16) to dorsal (z = 44). The clusters in red survived a size cutoff corresponding to p < .05 corrected for multiple comparisons. For Talairach coordinates, see Table 1. The bar chart shows that RTs for report change trials and matched expectation trials were the same (ns at both 80/20 and 50/50, paired t test), ruling out a time-to-task effect. Error bars are ±1 SE; dashed line is the mean.

Figure 3. 

Brain regions where activity changes during report change trials correlated with size of bias shift. Axial sections are ordered from ventral (z = −16) to dorsal (z = 44). The clusters in red survived a size cutoff corresponding to p < .05 corrected for multiple comparisons. For Talairach coordinates, see Table 1. The bar chart shows that RTs for report change trials and matched expectation trials were the same (ns at both 80/20 and 50/50, paired t test), ruling out a time-to-task effect. Error bars are ±1 SE; dashed line is the mean.

Table 1. 

Cluster Locations: Regions Where Activity Changes during Report Change Trials Correlated with Size of Bias Shift

Cluster
x
y
z
Volume, mm3
Mid-FuG (R) 41 35 20 3585 
IFJ (R) 45 32 958 
IFJ (L) −46 27 2591 
AntIns (L) −29 15 15 1242 
IPL (R) 49 −36 40 1029 
IPS (R) 30 −57 39 2875 
IPL/IPS (L) −42 −45 43 4366 
CaudHead (R) 15 15 781 
PostCing (bi) −7 −29 24 923 
FuG (R) 44 −54 −10 1455 
FuG (L) −40 −59 −10 3549 
Cluster
x
y
z
Volume, mm3
Mid-FuG (R) 41 35 20 3585 
IFJ (R) 45 32 958 
IFJ (L) −46 27 2591 
AntIns (L) −29 15 15 1242 
IPL (R) 49 −36 40 1029 
IPS (R) 30 −57 39 2875 
IPL/IPS (L) −42 −45 43 4366 
CaudHead (R) 15 15 781 
PostCing (bi) −7 −29 24 923 
FuG (R) 44 −54 −10 1455 
FuG (L) −40 −59 −10 3549 

R = right; L = left; PostCing (bi) = bilateral posterior cingulate; CaudHead = head of the caudate.

Scatter plots of bias shift versus signal change in all identified ROIs showed a positive slope and near-zero intercept (Figure 4). The slope and intercept information indicates that the correlations were driven by greater activity change in the participants making more use of the cues versus little activity change in the participants making less use of the cues. These findings confirm that the brain regions revealed by our ROI identification approach were involved in the use of prior knowledge to bias decisions adaptively. The cluster selection process (Methods) ensured that surviving regions would have significant mean within-region correlation values (for these, see Table 3).

Figure 4. 

Behavior versus fMRI activity. X axes show bias shift, 80/20 minus 50/50, normalized by the ideal observer's bias shift. Y axes show change in mean fMRI signal, 80/20 minus 50/50, during report change trials. Each point represents data from one subject. The intercepts are near zero and the slopes are positive. The cluster selection process ensured that surviving regions would have significant mean within-region correlation values (for these, see Table 3).

Figure 4. 

Behavior versus fMRI activity. X axes show bias shift, 80/20 minus 50/50, normalized by the ideal observer's bias shift. Y axes show change in mean fMRI signal, 80/20 minus 50/50, during report change trials. Each point represents data from one subject. The intercepts are near zero and the slopes are positive. The cluster selection process ensured that surviving regions would have significant mean within-region correlation values (for these, see Table 3).

Imaging Data: Tuning Curves

For an in-depth look at the brain data, we plotted the within-ROI data from the 80/20 condition along the dimension of target curvature (Figure 5). To produce the tuning curves, the targets were first sorted by curvature level into nine bins. Activations were then calculated for each individual participant via a multiple regression analysis, with nine regressors, each representing all of the targets in one bin. These results were averaged across participants to obtain within-ROI grand means and standard errors. The plots of grand means and standard errors represent within-ROI tuning curves along the dimension of smooth to bumpy target curvature. The data from participants for whom the 80/20 cue meant 80% A, 20% B are shown as orange tuning curves in Figure 5, and the data from participants for whom the 80/20 cue meant 80% B, 20% A are shown in red. (We refer to the former group of participants as Group A and to the latter as Group B.)

Figure 5. 

Tuning curves. Curvature levels were binned into nine bins; the curves show mean within-ROI fMRI signal level versus baseline at each bin. Shading indicates ±1 SE across subjects. Gray shading indicates bins in which the targets were not ambiguous. The short horizontal lines above the tuning curves show mean curvature levels for Group A participants (orange) and Group B participants (red), ±1 SE across individuals, of the maximum fMRI response within the ambiguous portion of the stimulus set. Stars indicate the significance of the difference between mean curvature levels, calculated via one-tailed t test: *p < .05, **p < .01, ***p < .001. The estimated p values for the unstarred ROIs were as follows: LH IFJ, p > .056; RH head of the caudate (CaudHead), p > .29; bilateral posterior cingulate (Bi PostCing), p > .060; RH FuG, p > .25; LH FuG, p > .20.

Figure 5. 

Tuning curves. Curvature levels were binned into nine bins; the curves show mean within-ROI fMRI signal level versus baseline at each bin. Shading indicates ±1 SE across subjects. Gray shading indicates bins in which the targets were not ambiguous. The short horizontal lines above the tuning curves show mean curvature levels for Group A participants (orange) and Group B participants (red), ±1 SE across individuals, of the maximum fMRI response within the ambiguous portion of the stimulus set. Stars indicate the significance of the difference between mean curvature levels, calculated via one-tailed t test: *p < .05, **p < .01, ***p < .001. The estimated p values for the unstarred ROIs were as follows: LH IFJ, p > .056; RH head of the caudate (CaudHead), p > .29; bilateral posterior cingulate (Bi PostCing), p > .060; RH FuG, p > .25; LH FuG, p > .20.

As shown in Figure 5, the tuning curves peaked at contraindicated stimulus levels in mid-FuG, IFJ, IPL, IPS, and anterior insula—as indicated by a separation of the curves at one or both ends of the curvature dimension—but were essentially flat in the head of the caudate and FuG, with only a hint of the trend in posterior cingulate. To further quantify this observation, we identified the peak ambiguous bin in each individual subject. The across-subject means of these values (±1 SE) are plotted as horizontal lines above the tuning curves. In mid-FuG, right hemisphere (RH) IFJ, IPL, IPS, and anterior insula, the difference between the means between Group A and Group B participants was highly significant (p values derived from a one-tailed t test ranging from p < .001 to p < .01; see Figure 5), but in head of the caudate and bilateral FuG, the means were the same in both participant groups. In left hemisphere (LH) IFJ and bilateral posterior cingulate, we observed a trend toward a difference, but this did not reach significance.

Imaging Data: Matched Expectation Control

We next sought to determine whether the above observations reflected a direct effect of prior knowledge on decision behavior, as opposed to an effect of stimulus expectedness or unexpectedness induced by the prior knowledge cues. To make this distinction, we exploited another aspect of the behavioral data. The target in each report change trial had a given curvature level; accordingly, we searched each subject's behavioral data to identify trials in which targets had the same curvature level, but the subject's reports did not contribute to the observed bias shift. Because stimulus expectedness or unexpectedness levels in these trials and the report change trials were identical, we called these trials matched expectation trials. If stimulus expectedness or unexpectedness were the source of the correlations obtained with the report change trials, then the matched expectation trials should produce similar correlations. In fact, however, the correlations dropped to near zero when the same analysis was performed using the matched expectation trials and did not reach significance in any ROI (Table 3). There was no RT difference between the matched expectation trials and the report change trials (Figure 3), so the null effect in the matched expectation analysis cannot be attributed to RT differences. The matched expectation analysis indicated that the correlations were not due to stimulus expectedness or unexpectedness induced by the prior knowledge cues, but instead reflected the magnitude of the effect of these cues on decision behavior.

Imaging Data: Relative versus Absolute Bias Optimality

Because the report change correlation analysis was based on a shift in behavioral decision bias, its results represented a relative effect. That is, it showed that within-ROI activity changes correlated with bias optimality at 80/20 relative to 50/50. We were also interested to test whether within-ROI activity levels correlated with the optimality of the absolute bias adopted within either the 80/20 or the 50/50 condition. In this case, the hypothesis was that participants whose ability to account for prior knowledge was closer to optimal, as measured by absolute bias levels at either 80/20 or 50/50, would experience greater within-ROI activity levels during report change trials than participants performing less optimally.

This hypothesis predicted that scatter plots of deviation from optimal bias versus activity level should have positive intercepts and negative slopes. However, the predicted pattern was not observed in the scatter plots and did not reach significance in any of the ROIs (Table 3) in either the 80/20 or 50/50 condition. We did observe three exceptions to this observation, namely significant correlations at 50/50 in RH IPS, bilateral posterior cingulate, and LH FuG (Table 3), but these were driven by data from a single participant whose 50/50 absolute bias measurement was identified as an outlier by Grubbs' test (Grubbs, 1969). When this participant was removed, correlations in all three regions fell to well below significance (Table 3 caption). For completeness, we applied Grubbs' test to the bias shift magnitudes and the 80/20 absolute bias measurements as well. No outliers were found in the set of bias shift magnitudes. In the set of 80/20 absolute bias measurements, one outlier was found, but the absolute 80/20 correlations were insignificant in all ROIs regardless of whether that outlier was included or excluded. These findings indicate that within-ROI activity accounted for interparticipant behavioral differences in relative but not absolute bias placement.

Imaging Data: Visual Cortex

The results presented thus far have not ruled out the possibility that the report change correlations used in ROI identification could reflect the degree to which the overall level of attention to the stimuli varied across participants or fluctuated across trials (Cohen & Maunsell, 2010, 2011). Our study was designed to investigate how participants used the prior knowledge information to help them make their decisions, not how participants attended to the stimuli. Therefore, if some participants paid more attention to the stimuli than others—or if attention across trials fluctuated more in some participants than in others—that discrepancy could have affected the results. To address this potential problem, we reasoned that increases in attention should increase fMRI activity in early visual areas. We identified ROIs by searching for visually responsive regions, as defined by regions in occipital cortex with highly significant (p < .000001, uncorrected) positive activations across all presentations of the targets. The resulting visual cortex (VC) ROIs (Figure 6; Table 2) were near the junction of the middle occipital and inferior occipital gyri bilaterally, with Talairach coordinates 28, −82, −8 and −33, −77, −11. These coordinates were nearly identical to the coordinates of the central 2° of the visual field in areas V1, V2, V3, and V4 (i.e., 25, −80, −9 and −29, −78, −11 as identified in a detailed retinotopic study by Dougherty et al., 2003). Because our targets were presented within the central 2°, we conclude that the VC ROIs represent the central field representation of the early visual areas. No across-subject correlations between behavioral responses and fMRI data from the VC ROIs were significant (Table 3), implying that the positive report change correlations seen in the fronto-parietal ROIs were not due to across-subject variance in overall attention.

Figure 6. 

Visual cortex. Left: ROI locations. For Talairach coordinates, see Table 2. Middle: X axes show bias shift, 80/20 minus 50/50, normalized by the ideal observer's bias shift. Y axes show change in mean fMRI signal, 80/20 minus 50/50, during report change trials. Each point represents data from one subject. The slopes are flat (for statistics, see Table 3). Right: Tuning curves. Curvature levels were binned into nine bins; the curves show mean within-ROI fMRI signal level versus baseline at each bin. Shading indicates ±1 SE across subjects. Gray shading indicates bins in which the targets were not ambiguous. The short horizontal lines above the tuning curves show mean curvature levels for Group A participants (orange) and Group B participants (red), ±1 SE across individuals, of the maximum fMRI response within the ambiguous portion of the stimulus set. The means were not significantly different (RH VC, p > .69; LH VC, p > .59 via one-tailed t test).

Figure 6. 

Visual cortex. Left: ROI locations. For Talairach coordinates, see Table 2. Middle: X axes show bias shift, 80/20 minus 50/50, normalized by the ideal observer's bias shift. Y axes show change in mean fMRI signal, 80/20 minus 50/50, during report change trials. Each point represents data from one subject. The slopes are flat (for statistics, see Table 3). Right: Tuning curves. Curvature levels were binned into nine bins; the curves show mean within-ROI fMRI signal level versus baseline at each bin. Shading indicates ±1 SE across subjects. Gray shading indicates bins in which the targets were not ambiguous. The short horizontal lines above the tuning curves show mean curvature levels for Group A participants (orange) and Group B participants (red), ±1 SE across individuals, of the maximum fMRI response within the ambiguous portion of the stimulus set. The means were not significantly different (RH VC, p > .69; LH VC, p > .59 via one-tailed t test).

Table 2. 

Cluster Locations: Occipital Regions with Positive Activations across All Trials

Cluster
x
y
z
Volume, mm3
VC (R) 28 −82 −8 852 
VC (L) −33 −77 −11 213 
Cluster
x
y
z
Volume, mm3
VC (R) 28 −82 −8 852 
VC (L) −33 −77 −11 213 

R = right; L = left.

Table 3. 

Across-subject Correlations of Bias Measurements versus fMRI Activity

ROI
Report Change Trials, 80/20 Minus 50/50
Matched Expectation Trials, 80/20 Minus 50/50
Report Change Trials, 80/20 Alone
Matched Expectation Trials, 80/20 Alone
Report Change Trials, 50/50 Alone
Matched Expectation Trials, 50/50 Alone
Mid-FuG (R) .002 .38 .84 .00 .32 .05 .98 .00 .67 .01 .43 .03 
IFJ (R) .002 .39 .99 .00 .91 .00 .62 .01 .41 .03 .74 .01 
IFJ (L) .0003 .49 .54 .02 .18 .09 .83 .00 .25 .07 .59 .01 
AntIns (L) .001 .41 .95 .00 .22 .07 .91 .00 .69 .01 .65 .01 
IPL (R) .007 .31 .99 .00 .40 .04 .87 .00 .74 .01 .79 .00 
IPS (R) .004 .35 .97 .00 .02 .26 .09 .13 .07 .16 .47 .03 
IPL/IPS (L) .002 .39 .67 .01 .24 .07 .59 .01 .49 .02 .80 .00 
CaudHead (R) .009 .29 .91 .00 .42 .03 .42 .03 .75 .01 .92 .00 
PostCing (bi) .002 .39 .87 .00 .03 .23 .08 .15 .13 .11 .55 .02 
FuG (R) .003 .36 .12 .12 .02 .24 .27 .06 .42 .03 .97 .00 
FuG (L) .008 .30 .11 .13 .42 .03 .63 .01 .67 .01 .35 .04 
VC (R) .38 .04 .22 .07 .74 .01 .73 .01 .71 .01 .58 .02 
VC (L) .35 .04 .10 .13 .51 .02 .70 .01 .28 .06 .95 .00 
ROI
Report Change Trials, 80/20 Minus 50/50
Matched Expectation Trials, 80/20 Minus 50/50
Report Change Trials, 80/20 Alone
Matched Expectation Trials, 80/20 Alone
Report Change Trials, 50/50 Alone
Matched Expectation Trials, 50/50 Alone
Mid-FuG (R) .002 .38 .84 .00 .32 .05 .98 .00 .67 .01 .43 .03 
IFJ (R) .002 .39 .99 .00 .91 .00 .62 .01 .41 .03 .74 .01 
IFJ (L) .0003 .49 .54 .02 .18 .09 .83 .00 .25 .07 .59 .01 
AntIns (L) .001 .41 .95 .00 .22 .07 .91 .00 .69 .01 .65 .01 
IPL (R) .007 .31 .99 .00 .40 .04 .87 .00 .74 .01 .79 .00 
IPS (R) .004 .35 .97 .00 .02 .26 .09 .13 .07 .16 .47 .03 
IPL/IPS (L) .002 .39 .67 .01 .24 .07 .59 .01 .49 .02 .80 .00 
CaudHead (R) .009 .29 .91 .00 .42 .03 .42 .03 .75 .01 .92 .00 
PostCing (bi) .002 .39 .87 .00 .03 .23 .08 .15 .13 .11 .55 .02 
FuG (R) .003 .36 .12 .12 .02 .24 .27 .06 .42 .03 .97 .00 
FuG (L) .008 .30 .11 .13 .42 .03 .63 .01 .67 .01 .35 .04 
VC (R) .38 .04 .22 .07 .74 .01 .73 .01 .71 .01 .58 .02 
VC (L) .35 .04 .10 .13 .51 .02 .70 .01 .28 .06 .95 .00 

Each entry gives a p value (first number, in bold italics) and r2 (second number). All p values were calculated via linear least squares regression. Because the fronto-parietal clusters were defined by searching for positive results with report change trials at 80/20 minus 50/50, the main observations here are negative. All of the VC analyses produced null results. All of the fronto-parietal analyses produced null results, except for the analysis that defined the ROIs and the analysis on report change trials at 50/50 alone in RH IPS, bilateral posterior cingulate, and LH FuG. Note that, unlike the bias shift magnitudes, the absolute measurements of bias at 50/50 did include one outlier as measured by Grubbs' test at a conventional significance level of p < .05. When this outlier was removed, the p (r2) values for RH IPS, bilateral posterior cingulate, and LH FuG were .56 (.02), .37 (.04), and .77 (.00), respectively.

R = right; L = left; AntIns = anterior insula; CaudHead = head of the caudate; PostCing (bi) = bilateral posterior cingulate.

Imaging Data: Additional Control Analyses

Our ROI identification strategy exploited the wide range of observed behavioral bias shifts to calculate across-subject correlations with fMRI activity dynamics, thus locating regions where the activity dynamics were consistent with a mechanism mediating the behavioral effect. It might be argued, however, that across-subject correlations could reflect trait-level differences between participants playing an important role in the differential activity levels. Conceivably, such trait-level differences might have relatively little to do with the specific process of biasing judgment. To rule out such a factor, we recalculated the report change correlations using d-prime shifts, rather than bias shifts, as the behavioral metric. In all ROIs, results were null (Table 4). The null results imply that if some underlying trait-level difference plays a role in the fMRI activity dynamics, it affects bias but not discriminability.

Table 4. 

Across-subject Correlations of d-prime Measurements versus fMRI Activity

ROI
p
r2
Mid-FuG (R) .61 .01 
IFJ (R) .27 .06 
IFJ (L) .48 .03 
AntIns (L) .67 .01 
IPL (R) .37 .04 
IPS (R) .40 .04 
IPL/IPS (L) .29 .05 
CaudHead (R) .70 .01 
PostCing (bi) .71 .01 
FuG (R) .23 .07 
FuG (L) .19 .08 
VC (R) .56 .02 
VC (L) .42 .03 
ROI
p
r2
Mid-FuG (R) .61 .01 
IFJ (R) .27 .06 
IFJ (L) .48 .03 
AntIns (L) .67 .01 
IPL (R) .37 .04 
IPS (R) .40 .04 
IPL/IPS (L) .29 .05 
CaudHead (R) .70 .01 
PostCing (bi) .71 .01 
FuG (R) .23 .07 
FuG (L) .19 .08 
VC (R) .56 .02 
VC (L) .42 .03 

Entries represent values derived from differences between report change trials in the 80/20 versus 50/50 runs. All p values were calculated via linear least squares regression. Results were null in all ROIs.

R = right; L = left; AntIns = anterior insula; CaudHead = head of the caudate; PostCing (bi) = bilateral posterior cingulate.

As discussed earlier, the participants with smaller shifts did not uniformly treat the 80/20 cue as 50/50, which would have resulted in a cluster of criterion measurements near the optimal 50/50 level in both cue conditions. Instead, the participants with smaller shifts produced widely varying criterion levels. It might be argued that the widely varying criterion levels could reflect higher levels of variability within or between participants. Because in averaged fMRI data, variability tends to produce lower averaged evoked activity, this scenario could raise a potentially significant problem for interpretation of the results. To rule out such a scenario, we divided the participant pool into halves by the size of the bias shift and calculated the averaged evoked activity in each half. The results (Figure 7) showed that activity in the participants with smaller shifts was not, in fact, lower than the activity in the participants with larger shifts; there was no difference across participant groups in any ROI.

Figure 7. 

Absolute signal change levels in participants with small versus large bias shifts. Conceivably, the bias shift correlations could have been driven by across-subject differences in absolute fMRI activity levels during the report change trials. To rule out this possibility, we divided the participant pool into halves by size of the bias shift and plotted the activity levels directly. Light gray represents the 11 participants with bias shifts less than the median of the 22 participants; dark gray represents the 11 participants with bias shifts greater than the median of the 22 participants. In no ROI was there a difference across participant groups in absolute fMRI activity levels during the report change trials.

Figure 7. 

Absolute signal change levels in participants with small versus large bias shifts. Conceivably, the bias shift correlations could have been driven by across-subject differences in absolute fMRI activity levels during the report change trials. To rule out this possibility, we divided the participant pool into halves by size of the bias shift and plotted the activity levels directly. Light gray represents the 11 participants with bias shifts less than the median of the 22 participants; dark gray represents the 11 participants with bias shifts greater than the median of the 22 participants. In no ROI was there a difference across participant groups in absolute fMRI activity levels during the report change trials.

DISCUSSION

In this study, we sought to locate and investigate the brain areas mediating the use of prior knowledge to bias decisions. The study produced four main findings: (1) We observed substantial across-subject variance in the magnitude of the difference in behavioral bias measurements between two prior knowledge cue conditions, 80/20 and 50/50. Although the decision bias shifted between prior knowledge conditions in the predicted direction in every subject, individual bias shift magnitudes ranged from near optimal to very suboptimal. (2) As a metric for the effect of prior knowledge on decision behavior, we used the magnitude of the bias shift between prior knowledge cue conditions (80/20 vs. 50/50) to search for brain regions in which activity changes correlated with the extent of the bias shift. These were dorsolateral pFC (middle frontal gyrus), IFJ, anterior insula, IPL, IPS, head of the caudate, posterior cingulate gyrus, and a focal, bilateral region within the FuG. (3) In fronto-parietal cortex and anterior insula, the greatest fMRI activity levels were seen in trials in which target curvature corresponded to the category contraindicated by the prior knowledge cue. In the posterior cingulate cortex, there was only a hint of this trend, and activity levels in the head of the caudate and the FuG were constant regardless of target curvature. (4) As an alternative metric for the effect of prior knowledge on decision behavior, we used absolute bias measurements within the 80/20 and 50/50 prior knowledge cue conditions separately. Using this metric, we found no significant correlation between optimality of prior knowledge cue use at either 80/20 or 50/50 with within-ROI brain activity levels in the same condition. This null result implies that brain activity during an explicitly cued prior knowledge task does not weight accumulated perceptual evidence by a literal representation of absolute prior probability, as would be predicted by simple signal detection models of the decision variable (Ratcliff & McKoon, 2008; Gold & Shadlen, 2007). Instead, the mechanism using current prior knowledge conditions to influence perceptual decisions appears to operate relative to past prior knowledge conditions.

The observed correlations between activity change and bias shift magnitude were significant only when computed with activity from trials when participants' decision reports were as predicted by the prior knowledge cue. Correlations computed with activity from trials with matched stimuli, but in which participants' decision reports were not as predicted by the prior knowledge cue, were near zero. These results rule out stimulus expectedness as an explanation for the report change correlations. These results also imply that the mechanism imposing the bias shift represents a dynamic interaction of the prior knowledge and the eventual decision made. This conclusion conflicts with simple signal detection models of the decision variable, in which the bias shift does not depend on the trial-by-trial decisions made by the participants. However, our conclusion is consistent with a recent report that prior probability was incorporated into the decision process as a dynamic bias signal that increases as a function of decision time (Hanks, Mazurek, Kiani, Hopp, & Shadlen, 2011).

It is worth noting that a subtraction analysis, contrasting activation during decisions in the 80/20 versus 50/50 conditions (see Methods), revealed no significant clusters. A similar contrast in a recent study of prior knowledge during decisions about moving dots (Forstmann, Brown, Dutilh, Neumann, & Wagenmakers, 2010) also produced a null result. However, subtraction analyses average together all trials at each condition, regardless of whether trials contribute to behavioral effects. Thus, a simple 80/20 minus 50/50 contrast does not account for how or even whether each participant used the prior knowledge cues. The alternative approach that we used here showed that an individual participant's brain activity is reflected in that participant's use of the prior knowledge cues.

Our design controlled for spurious stimulus-level effects by training the participants with target distributions matching the cues before scanning and presenting identical 50/50 target distributions during scanning (Methods). Therefore, it is important to consider whether the participants may have realized—although no feedback was given during scanning and the target categories overlapped—that the 80/20 cue in scanning runs no longer reflected the target distributions. Two lines of reasoning imply that this was not the case. (1) If this were the case, participants would have come to this realization after gaining experience with the categories as the experiment progressed. Therefore, the mean response curves for the 80/20 condition in early runs should be more biased than the mean response curves in later runs. However, the response curves for the early and late runs overlapped (Figure 1B). (2) After scanning, we debriefed the participants, and no participant indicated any concern about the cues being misleading.

The fronto-parietal ROIs overlap regions previously identified by other studies in human participants as responding during decision tasks using stimuli and behavioral responses of various modalities (Preuschoff, Quartz, & Bossaerts, 2008; Grinband, Hirsch, & Ferrera, 2006; Huettel, Stowe, Gordon, Warner, & Platt, 2006; Milham, Banich, & Barad, 2003). In monkeys, neurons in Area 46 (cf. our middle frontal gyrus ROI) are involved in the selection and control of action based on abstract rules or task strategy during perceptual decisions (Amemori & Sawaguchi, 2006; Muhammad, Wallis, & Miller, 2006; Wallis & Miller, 2003; Roberts & Wallis, 2000; White & Wise, 1999), and neurons in the lateral intraparietal area (cf. our IPS/IPL ROI) are involved in evidence integration (Kiani, Hanks, & Shadlen, 2008; Huk & Shadlen, 2005) and probability of reward (Yang & Shadlen, 2007) during decision-making. Relatedly, a study inducing asymmetric costs in perceptual decisions (Fleming, Whiteley, Hulme, Sahani, & Dolan, 2010) found, as here, criterion shifts but not d-prime shifts with greater frontal cortex activity in human participants who showed greater bias shifts. A different study that incorporated an explicit prior probability cue with an easy decision task—in which human participants pressed a button as quickly as possible to indicate which of two single digits represented the larger number (Scheibe, Ullsperger, Sommer, & Heekeren, 2010)—found that in dorsolateral prefrontal, inferior frontal, and inferior parietal cortex, the fMRI signal over time correlated with EEG fluctuations previously shown to reflect prior-probability-based response preparation.

The greater activity for contraindicated than indicated targets that we observed in the fronto-parietal and anterior insula ROIs suggests that these regions contribute to the decision using a mechanism similar to a no-go response. In this scenario, the prior knowledge induces a default go behavioral response in favor of the indicated (80%) alternative. The role of the fronto-parietal and anterior insula regions is to overcome that default response (no-go) when sufficient perceptual information favors the contraindicated (20%) alternative. This explanation is not only consistent with our data but is also consistent with the results of previous go/no-go studies that asked human participants to respond to different cues with either finger movement (go) or no finger movement (no-go). These studies revealed preferential activation for no-go over go trials in fronto-parietal cortex and anterior insula (Watanabe et al., 2002; Casey et al., 1997; Kawashima et al., 1996).

Similar regions are known to respond preferentially to unexpected compared with expected stimuli (Melcher & Gruber, 2006; Derrfuss, Brass, Neumann, & von Cramon, 2005; Huettel, Mack, & McCarthy, 2002; McCarthy, Luby, Gore, & Goldman-Rakic, 1997). Expectedness and prior probability are conceptually intertwined, because studies of brain responses to expected and unexpected stimuli manipulate prior probability. Our matched expectation control trials teased these concepts apart, however, by showing that the data can be accounted for by a change in behavioral criterion change but not by stimulus expectedness. In fact, this control suggests that some previous findings attributed to expectedness might have been due to unexplored internal criterion changes occurring as participants experienced the stimuli.

The across-subject correlation that we used to identify ROIs also revealed regions without preferential activity for contraindicated targets, namely head of the caudate, posterior cingulate cortex, and a focal, bilateral region within the FuG. Studies in monkey have shown that the head of the caudate receives anatomical projections from dorsolateral pFC (Kemp & Powell, 1970) and is active when environmental cues are used in the preparation of behavioral responses (Rolls, Thorpe, & Maddison, 1983). A study in monkeys showed that posterior cingulate neurons tracked decision salience—the degree to which an option differs from a standard—but not the subjective value of a decision (Heilbronner, Hayden, & Platt, 2011). An imaging study has shown that a human fusiform region corresponding to our RH fusiform region was active in participants reporting nonvisual perceptual decisions (Zhou & Chen, 2008). This region's LH counterpart, corresponding to our LH fusiform region, is known as the visual word form area and has been shown by many studies to be selective in stimulus–response contexts as diverse as reading visually presented words (Cohen et al., 2002), naming colors (Price & Friston, 1997), and priming by previously presented nonsense objects (Van Turennout, Ellmore, & Martin, 2000). Thus, these regions have been shown to participate in processes relevant for perceptual decisions. Our findings add evidence supporting the conclusion that these regions are among those actually driving the process of integrating prior knowledge into decisions.

Our fMRI results extend recent work that used a moving dot decision task and a diffusion modeling approach and obtained null results in frontal and parietal cortex (Forstmann et al., 2010). The null parietal result in that study was surprising. In both monkeys and humans, activity in parietal cortex has been reported in several studies of decisions about moving dots (Kayser, Buchsbaum, Erickson, & D'Esposito, 2009; Kiani et al., 2008; Hanks, Ditterich, & Shadlen, 2006; Heekeren, Marrett, Ruff, Bandettini, & Ungerleider, 2006; Shulman et al., 2003), and in monkeys, activity in parietal neurons has been shown to integrate stimulus information with prior probabilities about the stimuli (Yang & Shadlen, 2007). The absence of parietal activity in the human dot decision task could have been due to instant feedback presented during scanning, which the current study did not include. Participants may have experienced positive feedback as a reward; indeed, both orbito-frontal and putamen neurons have been shown to respond to a cue indicating high expectation of reward (Schultz, Tremblay, & Hollerman, 2000), and the Forstmann et al. study did report positive results in both OFC and putamen.

When evaluating the matched expectation results, one might ask what was the source of the variance in performance that produced the matched expectation trials. Recall that matched expectation trials were defined as those with the same curvature levels as the report change trials, but in which the report did not change. Conceivably, the variance could be due to random fluctuations in the participants' usage of the cue, such that the participants were considering the cue information on the report change trials but failing to consider it on the matched expectation trials. Alternatively, the variance could be due to internal noise at the level of curvature estimation. (Note that, in either case, our findings imply that the fronto-parietal activity correlates with usage of the prior knowledge cue in decisions.) Random fluctuations in cue usage would imply different psychological processes for report change and matched expectation trials, but this explanation seems unlikely because the RTs were identical for both trial types (Figure 4B). We conclude that the variance was due to internal noise at the level of curvature estimation, not maintenance of the goal of using the cue.

The current study used abstract shape targets composed of curves but did not identify effects of prior knowledge in brain regions such as lateral occipital complex or kinetic occipital cortex that have previously been implicated in selectivity for curves and abstract shapes (Amir, Biederman, & Hayworth, 2011; Gillebert, Op de Beeck, Panis, & Wagemans, 2009; Zhang, Meeson, Welchman, & Kourtzi, 2008; Li, Ostwald, Giese, & Kourtzi, 2007; Dupont et al., 1997). This was surprising, because an earlier study (Esterman & Yantis, 2010), which examined fMRI activation levels during decisions about faces and houses slowly emerging from a visual noise mask, identified effects of prior knowledge (expectation) in anterior visual brain regions selective for faces and houses (i.e., fusiform and parahippocampal gyri) but not in frontal or parietal cortex. Similarly, another study (Summerfield & Koechlin, 2008), which used a different approach to identify prior knowledge effects in the brain, defining an A versis not A discrimination task between contrast-modulated stimuli as a priors condition and an A versus B discrimination task between contrast-modulated stimuli as an unbiased condition, identified prior knowledge effects in both prefrontal and extrastriate VC. Therefore, it is important to consider why the current study did not produce prior knowledge effects in lateral occipital complex or kinetic occipital. There were many slight variations in experimental design between the current and previous studies, but we speculate that the critical factor was stimulus visibility. Our nonnoisy, high-contrast ambiguous stimuli produced fMRI signal modulations in prefrontal and parietal but not VC. This observation is consistent with the idea that the effects of prior knowledge in our study took place at the level of decision, rather than as a modulation of the sensory input preceding the decision. Future experiments will be needed to compare the effects of prior knowledge on noisy versus nonnoisy stimuli while keeping other variables constant.

Acknowledgments

We thank Chris Baker, Hauke Heekeren, Shruti Japee, Sean Marrett, Alex Martin, and Lionel Rauth for helpful comments. This work was supported by the National Institute of Mental Health Intramural Research Program.

Reprint requests should be sent to Kathleen A. Hansen, Laboratory of Brain and Cognition, Building 10 Room 4C104, National Institute of Mental Health, National Institutes of Health, Bethesda, MD 20892, or via e-mail: hansenka@mail.nih.gov.

REFERENCES

REFERENCES
Amemori
,
K.
, &
Sawaguchi
,
T.
(
2006
).
Rule-dependent shifting of sensorimotor representation in the primate prefrontal cortex.
European Journal of Neuroscience
,
23
,
1895
1909
.
Amir
,
O.
,
Biederman
,
I.
, &
Hayworth
,
K. J.
(
2011
).
The neural basis for shape preferences.
Vision Research
,
51
,
2198
2206
.
Bohil
,
C. J.
, &
Maddox
,
W. T.
(
2001
).
Category discriminability, base-rate, and payoff effects in perceptual categorization.
Perception and Psychophysics
,
63
,
361
376
.
Casey
,
B. J.
,
Trainor
,
R. J.
,
Orendi
,
J. L.
,
Schubert
,
A. B.
,
Nystrom
,
L. E.
,
Giedd
,
J. N.
,
et al
(
1997
).
A developmental functional MRI study of prefrontal activation during performance of a go-no-go task.
Journal of Cognitive Neuroscience
,
9
,
835
847
.
Cohen
,
L.
,
Lehericy
,
S.
,
Chochon
,
F.
,
Lemer
,
C.
,
Rivard
,
S.
, &
Dehaene
,
S.
(
2002
).
Language-specific tuning of visual cortex? Functional properties of the visual word form area.
Brain
,
125
,
1054
1069
.
Cohen
,
M. R.
, &
Maunsell
,
J. H.
(
2010
).
A neuronal population measure of attention predicts behavioral performance on individual trials.
Journal of Neuroscience
,
30
,
15421
15453
.
Cohen
,
M. R.
, &
Maunsell
,
J. H.
(
2011
).
When attention wanders: How uncontrolled fluctuations in attention affect performance.
Journal of Neuroscience
,
31
,
15802
15806
.
Cox
,
R. W.
(
1996
).
AFNI: Software for analysis and visualization of functional magnetic resonance neuroimages.
Computers and Biomedical Research
,
29
,
162
173
.
Cox
,
R. W.
, &
Hyde
,
J. S.
(
1997
).
Software tools for analysis and visualization of FMRI data.
NMR in Biomedicine
,
10
,
171
178
.
Derrfuss
,
J.
,
Brass
,
M.
,
Neumann
,
J.
, &
von Cramon
,
D. Y.
(
2005
).
Involvement of the inferior frontal junction in cognitive control: Meta-analyses of switching and Stroop studies.
Human Brain Mapping
,
25
,
22
34
.
Dougherty
,
R. F.
,
Koch
,
V. M.
,
Brewer
,
A. A.
,
Fischer
,
B.
,
Modersitzki
,
J.
, &
Wandell
,
B. A.
(
2003
).
Visual field representations and locations of visual areas V1/2/3 in human visual cortex.
Journal of Vision
,
3
,
586
598
.
Dupont
,
P.
,
De Bruyn
,
B.
,
Vandenberghe
,
R.
,
Rosier
,
A. M.
,
Michiels
,
J.
,
Marchal
,
G.
,
et al
(
1997
).
The kinetic occipital region in human visual cortex.
Cerebral Cortex
,
7
,
283
292
.
Esterman
,
M.
, &
Yantis
,
S.
(
2010
).
Perceptual expectation evokes category-selective cortical activity.
Cerebral Cortex
,
20
,
1245
1253
.
Fleming
,
S. M.
,
Whiteley
,
L.
,
Hulme
,
O. J.
,
Sahani
,
M.
, &
Dolan
,
R. J.
(
2010
).
Effects of category-specific costs on neural systems for perceptual decision-making.
Journal of Neurophysiology
,
103
,
3238
3247
.
Forstmann
,
B. U.
,
Brown
,
S.
,
Dutilh
,
G.
,
Neumann
,
J.
, &
Wagenmakers
,
E. J.
(
2010
).
The neural substrate of prior information in perceptual decision making: A model-based analysis.
Frontiers in Human Neuroscience
,
4
,
1
2
.
Gillebert
,
C. R.
,
Op de Beeck
,
H. P.
,
Panis
,
S.
, &
Wagemans
,
J.
(
2009
).
Subordinate categorization enhances the neural selectivity in human object-selective cortex for fine shape differences.
Journal of Cognitive Neuroscience
,
21
,
1054
1064
.
Gold
,
J. I.
, &
Shadlen
,
M. N.
(
2007
).
The neural basis of decision making.
Annual Review of Neuroscience
,
30
,
535
574
.
Green
,
D. M.
, &
Swets
,
J. A.
(
1966
).
Signal detection theory and psychophysics.
New York
:
Wiley
.
Grinband
,
J.
,
Hirsch
,
J.
, &
Ferrera
,
V. P.
(
2006
).
A neural representation of categorization uncertainty in the human brain.
Neuron
,
49
,
757
763
.
Grubbs
,
F. E.
(
1969
).
Procedures for detecting outlying observations in samples.
Technometrics
,
11
,
1
21
.
Hanks
,
T. D.
,
Ditterich
,
J.
, &
Shadlen
,
M. N.
(
2006
).
Microstimulation of macaque area LIP affects decision-making in a motion discrimination task.
Nature Neuroscience
,
9
,
682
689
.
Hanks
,
T. D.
,
Mazurek
,
M. E.
,
Kiani
,
R.
,
Hopp
,
E.
, &
Shadlen
,
M. N.
(
2011
).
Elapsed decision time affects the weighting of prior probability in a perceptual decision task.
Journal of Neuroscience
,
31
,
6339
6352
.
Hansen
,
K. A.
,
Hillenbrand
,
S. F.
, &
Ungerleider
,
L. G.
(
2011
).
Persistency of priors-induced bias in decision behavior and the FMRI signal.
Frontiers in Decision Neuroscience
,
5
,
1
9
.
Healy
,
A. F.
, &
Kubovy
,
M.
(
1978
).
The effects of payoffs and prior probabilities on indices of performance and cutoff locations in recognition memory.
Memory & Cognition
,
6
,
544
553
.
Healy
,
A. F.
, &
Kubovy
,
M.
(
1981
).
Probability matching and the formation of conservative decision rules in a numerical analog of signal detection.
Journal of Experimental Psychology: Human Learning and Memory
,
7
,
344
354
.
Heekeren
,
H. R.
,
Marrett
,
S.
,
Ruff
,
D. A.
,
Bandettini
,
P. A.
, &
Ungerleider
,
L. G.
(
2006
).
Involvement of human left dorsolateral prefrontal cortex in perceptual decision making is independent of response modality.
Proceedings of the National Academy of Sciences, U.S.A.
,
103
,
10023
10028
.
Heilbronner
,
S. R.
,
Hayden
,
B. Y.
, &
Platt
,
M. L.
(
2011
).
Decision salience signals in posterior cingulate cortex.
Frontiers in Neuroscience
,
5
,
1
9
.
Huettel
,
S. A.
,
Mack
,
P. B.
, &
McCarthy
,
G.
(
2002
).
Perceiving patterns in random series: Dynamic processing of sequence in prefrontal cortex.
Nature Neuroscience
,
5
,
485
490
.
Huettel
,
S. A.
,
Stowe
,
C. J.
,
Gordon
,
E. M.
,
Warner
,
B. T.
, &
Platt
,
M. L.
(
2006
).
Neural signatures of economic preferences for risk and ambiguity.
Neuron
,
2
,
765
775
.
Huk
,
A. C.
, &
Shadlen
,
M. N.
(
2005
).
Neural activity in macaque parietal cortex reflects temporal integration of visual motion signals during perceptual decision making.
Journal of Neuroscience
,
25
,
10420
10436
.
Kawashima
,
R.
,
Satoh
,
K.
,
Itoh
,
H.
,
Ono
,
S.
,
Furumoto
,
S.
,
Gotoh
,
R.
,
et al
(
1996
).
Functional anatomy of go/no-go discrimination and response selection: A PET study in man.
Brain Research
,
728
,
79
89
.
Kayser
,
A. S.
,
Buchsbaum
,
B. R.
,
Erickson
,
D. T.
, &
D'Esposito
,
M.
(
2009
).
The functional anatomy of a perceptual decision in the human brain.
Journal of Neurophysiology
,
103
,
1179
1194
.
Kemp
,
J. M.
, &
Powell
,
T. P.
(
1970
).
The cortico-striate projection in the monkey.
Brain
,
93
,
525
546
.
Kiani
,
R.
,
Hanks
,
T. D.
, &
Shadlen
,
M. N.
(
2008
).
Bounded integration in parietal cortex underlies decisions even when viewing duration is dictated by the environment.
Journal of Neuroscience
,
28
,
3017
3029
.
Li
,
S.
,
Ostwald
,
D.
,
Giese
,
M.
, &
Kourtzi
,
Z.
(
2007
).
Flexible coding for categorical decisions in the human brain.
Journal of Neuroscience
,
27
,
12321
12330
.
Maddox
,
W. T.
(
2002
).
Toward a unified theory of decision criterion learning in perceptual categorization.
Journal of the Experimental Analysis of Behavior
,
78
,
567
595
.
McCarthy
,
G.
,
Luby
,
M.
,
Gore
,
J.
, &
Goldman-Rakic
,
P.
(
1997
).
Infrequent events transiently activate human prefrontal and parietal cortex as measured by functional MRI.
Journal of Neurophysiology
,
77
,
1630
1634
.
Melcher
,
T.
, &
Gruber
,
O.
(
2006
).
Oddball and incongruity effects during Stroop task performance: A comparative fMRI study on selective attention.
Brain Research
,
1121
,
136
149
.
Milham
,
M. P.
,
Banich
,
M. T.
, &
Barad
,
V.
(
2003
).
Competition for priority in processing increases prefrontal cortex's involvement in top–down control: An event-related fMRI study of the Stroop task.
Cognitive Brain Research
,
17
,
212
222
.
Muhammad
,
R.
,
Wallis
,
J. D.
, &
Miller
,
E. K.
(
2006
).
A comparison of abstract rules in the prefrontal cortex, premotor cortex, inferior temporal cortex, and striatum.
Journal of Cognitive Neuroscience
,
18
,
974
989
.
Preuschoff
,
K.
,
Quartz
,
S. R.
, &
Bossaerts
,
P.
(
2008
).
Human insula activation reflects risk prediction errors as well as risk.
Journal of Neuroscience
,
28
,
2745
2752
.
Price
,
C. J.
, &
Friston
,
K. J.
(
1997
).
Cognitive conjunction.
Neuroimage
,
5
,
261
270
.
Ratcliff
,
R.
, &
McKoon
,
G.
(
2008
).
The diffusion decision model: Theory and data for two-choice decision tasks.
Neural Computation
,
20
,
873
922
.
Roberts
,
A. C.
, &
Wallis
,
J. D.
(
2000
).
Inhibitory control and affective processing in the prefrontal cortex: Neuropsychological studies in the common marmoset.
Cerebral Cortex
,
10
,
252
262
.
Rolls
,
E. T.
,
Thorpe
,
S. J.
, &
Maddison
,
S. P.
(
1983
).
Responses of striatal neurons in the behaving monkey. 1. Head of the caudate nucleus.
Behavioural Brain Research
,
7
,
179
210
.
Scheibe
,
C.
,
Ullsperger
,
M.
,
Sommer
,
W.
, &
Heekeren
,
H. R.
(
2010
).
Effects of parametrical and trial-to-trial variation in prior probability processing revealed by simultaneous electroencephalogram/functional magnetic resonance imaging.
Journal of Neuroscience
,
30
,
16709
16717
.
Schultz
,
W.
,
Tremblay
,
L.
, &
Hollerman
,
J. R.
(
2000
).
Reward processing in primate orbitofrontal cortex and basal ganglia.
Cerebral Cortex
,
10
,
272
284
.
Shulman
,
G. L.
,
McAvoy
,
M. P.
,
Cowan
,
M. C.
,
Astafiev
,
S. V.
,
Tansy
,
A. P.
,
d'Avossa
,
G.
,
et al
(
2003
).
Quantitative analysis of attention and detection signals during visual search.
Journal of Neurophysiology
,
90
,
3384
3397
.
Summerfield
,
C.
, &
Koechlin
,
E.
(
2008
).
A neural representation of prior information during perceptual inference.
Neuron
,
59
,
336
347
.
Van Turennout
,
M.
,
Ellmore
,
T.
, &
Martin
,
A.
(
2000
).
Long lasting cortical plasticity in the object naming system.
Nature Neuroscience
,
3
,
1329
1334
.
Wallis
,
J. D.
, &
Miller
,
E. K.
(
2003
).
From rule to response: Neuronal processes in the premotor and prefrontal cortex.
Journal of Neurophysiology
,
90
,
1790
1806
.
Watanabe
,
J.
,
Sugiura
,
M.
,
Sato
,
K.
,
Sato
,
Y.
,
Maeda
,
Y.
,
Matsue
,
Y.
,
et al
(
2002
).
The human prefrontal and parietal association cortices are involved in NO-GO performances: An event related fMRI study.
Neuroimage
,
17
,
1207
1216
.
White
,
I. M.
, &
Wise
,
S. P.
(
1999
).
Rule-dependent neuronal activity in the prefrontal cortex.
Experimental Brain Research
,
126
,
315
335
.
Wickens
,
T. D.
(
2002
).
Elementary signal detection theory.
New York
:
Oxford University Press
.
Wilkinson
,
F.
,
Wilson
,
H. R.
, &
Habak
,
C.
(
1998
).
Detection and recognition of radial frequency patterns.
Vision Research
,
38
,
3555
3568
.
Yang
,
T.
, &
Shadlen
,
M. N.
(
2007
).
Probabilistic reasoning by neurons.
Nature
,
447
,
1075
1080
.
Zhang
,
J.
,
Meeson
,
A.
,
Welchman
,
A. E.
, &
Kourtzi
,
Z.
(
2008
).
Learning alters the tuning of functional magnetic resonance imaging patterns for visual forms.
Journal of Neuroscience
,
30
,
14127
14133
.
Zhou
,
W.
, &
Chen
,
D.
(
2008
).
Encoding human sexual chemosensory cues in the orbitofrontal and fusiform cortices.
Journal of Neuroscience
,
28
,
14416
14421
.