DECODING BASE-RATE NEGLECT 1 Expecting the Unexpected : Base-rate Neglect is Driven by Neural Activation of Common Cues

The inverse base-rate effect (IBRE) describes an apparent irrationality in human decision making whereby people tend to ignore category base rates and choose rarer options when classifying ambiguous stimuli. According to some theories, people choose rare categories for ambiguous stimuli because rare cues draw more attention. Other theories predict that people choose rare categories because the ambiguous stimuli contrast with the more well-established patterns in common categories. In this experiment, we used multi-voxel pattern analysis to decode which features of ambiguous stimuli participants are activating during an fMRI version of the IBRE task. We show that individuals engage qualitatively distinct neural processes when making rare versus common responses: choosing a rarer category involved activation of cues associated with the common category. Consistent with inferential theories of base-rate neglect, our findings suggest that this surprising behavior involves a deliberative mechanism not explained by purely associative models.


INTRODUCTION
Categorization is a critical human capacity, enabling us to organize what we have learned in a way that generalizes to new experiences.Knowledge about the relative frequency of different categories in the environment can provide a valuable source of evidence when predicting category membership.Consider a doctor tasked with diagnosing a patient who presents with two symptoms -one associated with a common disease, and the other associated with a disease that occurs far less frequently.When faced with this ambiguous case that has properties of multiple categories (diseases), it is sensible to assume that the diagnosis would be biased in favor of the most common possible outcome.Yet, research shows that base-rate information is applied inconsistently in both medical decision making (Casscells, Schoenberger, and Graboy, 1978;Bravata, 2000) and problem solving contexts (Tversky and Kahneman, 1974).A notable case of this inconsistency is the inverse base-rate effect (IBRE), an influential finding whereby ambiguous stimuli that contain features of multiple categories are more often classified as the rarer of the two options (Medin and Edelson, 1988).
A recurring debate surrounding the cognitive basis of the IBRE is whether it is the byproduct of high-level reasoning processes (Winman, Wennerholm, and Juslin, 2003;Winman et al., 2005) or a fundamental associative mechanism (Kruschke, 2001a;Kruschke, 2003;Johansen, Fouquet, and Shanks, 2007;Lamberts and Kent, 2007).In the present study, we sought to clarify the cognitive and neural bases of the IBRE using fMRI and an adapted version of a category learning paradigm known to produce the effect (Medin and Edelson, 1988;Kruschke, 1996).Participants were trained to distinguish between four possible diseases based on combinations of visual cues.On each trial, predictions are made based on two co-present cues: one an imperfect predictor I, and the other either a common perfect predictor PC or rare perfect predictor PR.Each perfect predictor has a deterministic relationship with a disease, whereas the imperfect predictor is associated with multiple diseases.The only distinction between PC and PR is the disease base rates, with common diseases appearing three times more often than the rare ones throughout training.After learning, the key IBRE finding is that participants tend to ignore the base rates and choose the rare category when confronted with an ambiguous stimulus that contains both of the previously perfect common and rare predictors (PC.PR).
Most efforts to model the IBRE have focused on the role of attention in learning, and how fundamental attentional processes can produce this seemingly irrational behavior.Thus far, the most successful attentional models for accommodating the IBRE were formulated by Kruschke (1996;2001b).The attentional mechanism proposed to explain the IBRE is grounded in asymmetrical feature representations that form early during learning.Since I.PC occurs more frequently, participants first learn that I + PC à Common.Consequently, when encountering the less-frequent I.PR trials, individuals may shift their attention to PR and away from I, because I has already been associated with the common disease.The stronger tendency to choose the rare category when confronted with PC.PR is then explained by the stronger association that forms between PR and the rare outcome due to learned selective attention (Medin and Bettger, 1991;Kruschke, 1996).
Despite the consistency between the results of cue learning tasks and predictions from attention learning models, disagreement remains as to whether the IBRE is a basic attentional phenomenon, or whether higher-order inferential processes contribute to the surprising pattern of results (Juslin, Wennerholm, and Winman, 2001;Winman et al., 2005).Like the attention learning accounts, an inferential account may predict that people will learn the more common disease earlier in the task and the most strongly.However, these accounts differ in terms of why people chose the rare stimulus on the ambiguous PC.PR trials.From an inferential account, because I + PC à Common is so well learned, PC.PR is more unexpected for the common category, which leads participants to infer that it must come from the rare category (Juslin et al., 2001).Specifically, according to inferential theories, participants choose the rare category because the presence of the rare cue creates a stronger sense of mismatch between the presented stimuli and their memories of the common category.
Although attentional and inferential theories posit very different mechanisms as the basis of the IBRE, they have remained unresolvable at the level of behavior thus far.One possibility to resolve this debate is to begin to look to the brain to provide additional data on the underlying processes occurring during IBRE tasks that remain hidden to behavior.As a first pass, the goal of distinguishing between the neural regions associated with high-level inferential processes, such as hypothesis testing, rule switching, and problem solving and lower-level associative learning have been one of the main foci of neurobiological category learning research (Ashby and Maddox, 2005;Smith and Grossman, 2008;Seger and Miller, 2010).For example, rule-based processes, akin to those underlying inferential theories, are thought to depend upon cortico-striatal loops involving the PFC and head of the caudate nucleus, whereas simple visual associations may be learned using visual and procedural loops connecting the tail of the caudate with visual regions and motor outputs.
Despite well-established differences in the neural systems engaged for rulebased and associative tasks (Nomura et al., 2007;Schnyer et al., 2009), relying on such differences to infer whether rule-based or associative mechanisms are engaged in a particular task is much more tenuous.For one, to do so commits an affirming the consequent or "reverse inference" fallacy (Poldrack, 2006); many brain regions activate for a variety of cognitive functions, and thus activation alone does not permit inferences about the underlying process.Indeed, in a recent study examining a patterning task that is largely agreed upon to differentially tap associative and inferential mechanisms, it was not possible to cleanly infer that separable systems were involved based on the activation patterns alone (Milton et al., 2016).
Although classic localization-based fMRI would not be able to firmly establish the mechanisms underlying the IBRE and resolve the debate between inferential and associative theories, recent advances in representational multivoxel pattern analysis (MVPA) may (Norman et al., 2006;Kriegeskorte, Mur, and Bandettini, 2008;Davis and Poldrack, 2013).Instead of focusing on localization of cognitive function, MVPA allows researchers to map information contained in fMRI patterns onto experimental and psychological states.In the simplest form, MVPA can allow researchers to perform "mindreading" and decode the perceptual objects participants are looking at or retrieving from memory (Rissman and Wagner, 2012;Haxby, Connolly, and Guntupalli, 2014).
Studies employing MVPA to test mechanistic theories of cognition often focus on how cognitive processes warp the psychological stimulus space relative to the physical space (Davis and Poldrack, 2013;Davis and Poldrack, 2014).Predictions for this warping are tested by examining how and whether such predicted changes in the psychological stimulus space are reflected in changes in the neural similarities between elicited for different stimuli.For example, one recent study found, consistent with attentional learning models, that features that are attended during learning to become more prominent in neural similarity spaces in the lateral occipital cortex as a result of learning (Mack et al., 2013).Similarly, memory-based category representations stored in the hippocampus reflect these attention-based changes (Mack et al., 2016).These results suggest that MVPA could be used to test, which theories, attentional or inferential, provide the best account of the similarity relations between neural representations during an IBRE experiment by examining predictions for how such mechanisms may warp the underlying neural representational spaces.
Because associative and inferential theories differ in terms of whether participants are mainly considering the rare cues (associative theories) or common cues (inferential theories), it is possible to use MVPA to distinguish between the two theories by examining the activation of rare and common cues during ambiguous test trials.Specifically, if the IBRE is attributable to the associative mechanism of learned selective attention, ambiguous test trials should be accompanied by stronger activation of patterns associated with the rare cue (PR).On the other hand, if a form of inference is responsible for the IBRE, one would expect greater activation of patterns associated with the common cue (PC) as the well-known outcome is consciously considered and subsequently ruled out as an option.Unlike attentional effects that are thought to apply consistently across cases, the use of inference is thought to occur only on ambiguous trials receiving rare responses (Juslin et al., 2001).
Although MVPA-based category learning studies have proven able to distinguish between novel stimuli that vary in subtle features (Mack et al., 2013;Davis and Poldrack, 2014;Davis et al., 2014b;Mack et al., 2016), the current used real-world object categories (faces, objects, and scenes) as cues.These cues have a well-defined representational topography across the cortex, and are readily decodable even with very basic similarity-based analysis (Haxby et al., 2001).Thus we expected these realworld categories would allow maximal power for discriminating subtle changes in attentional weighting in the present task.Further, by localizing regions sensitive to these cues prior to engaging participants in the experimental task, we were able to create an independent neural similarity space that is unbiased with respect to any learningdependent salience effects, and thus allows stronger representational conclusions (Davis et al., 2014a).Specifically, we can examine how expression of face, scene, and object activation patterns during the key ambiguous test trials relates to those predicted by associative and inferential accounts.
In addition to our MVPA predictions for how pattern similarity will differ according to associative and inferential accounts of the IBRE, we expect univariate differences in brain regions involved with computing uncertainty and decision evidence signals between novel and well-learned stimuli.During the test phase, participants are exposed to cue pairings that have been previously learned or are presented in novel combinations.Novel stimuli should be associated with higher uncertainty, whereas well-learned stimuli should be associated with more certainty.Recent neuroimaging studies suggest that the lateral and dorsal PFC track categorization decisions with higher uncertainty (DeGutis and D'Esposito, 2007;Seger et al., 2015;Davis, Goldwater, and Giron, 2016) whereas the ventromedial prefrontal cortex (vmPFC) tracks decisions with high decision evidence (Lebreton et al., 2015;Davis et al., 2016).To foreshadow our results, consistent with inferential accounts of the IBRE, our analysis revealed stronger activation of neural patterns associated with common cues on conflicting PC.PR trials that were categorized as a rare disease.No differences in cue activation were observed when responses were consistent with category baserates on these ambiguous trials.Further, we show that the representational strength associated with common predictors for each participant during the learning phase is positively associated with individual differences in the tendency to ignore category base-rates at test.These results thus not only help to clarify the mechanisms underlying base-rate neglect, but also highlight the power of fMRI-based decoding for elucidating cognitive processes that have remained unresolvable at the level of behavior for decades.

Participants
Twenty-four healthy right-handed volunteers (age range 18 -58, mean age ± SEM = 25.08 ± 1.62, 13 women) were recruited through online newsletters and flyers posted at Texas Tech University.Two participants were excluded from the final analysis: one due to excessive head motion and one for falling asleep in the scanner.All subjects provided informed consent prior to participation, and were compensated $35 for a 1.5hour session.The study protocol was approved by the Texas Tech University Human Research Protection Program. .

Experimental Paradigm
Participants completed a category learning task based on a structure employed by several previous studies (Medin and Edelson, 1988;Kruschke, 1996) with a premise of predicting disease outcomes in hypothetical patients.The category structure (Figure 1) included four diseases overall, with two common diseases that were encountered at a rate of 3:1 compared with the rare diseases.Faces, objects, and scenes were used as cues to maximize the ability to detect feature-based selective attention via MVPA.Accordingly, participants were instructed that they would be learning to diagnose different patients based on people they have interacted with, places they have been, and objects that they recently used.Faces were always imperfect predictors, in that each face was associated with two different disease outcomes.Objects and scenes had a deterministic relationship with the disease outcomes, in that each distinct image was associated with only one disease.The face, object, and scene images used were blackand-white squares presented on a white background with black text.Learning occurred over three scanning runs, with each run presenting four blocks of the stimulus set, resulting in 32 trials per run and 96 learning trials over the whole training phase.Individual trials involved presenting an image pair side-by-side on the projection screen and prompting participants to respond to the question, "Disease 1, 2, 3, or 4?" with a button press within 3 s.Feedback was then provided for 1.75 s, indicating a correct or incorrect response accompanied by the correct disease label.The image pair remained on the screen during feedback.Variable fixation periods drawn from truncated exponentials (mean = 3 s) separated stimulus presentation from feedback, and feedback from the next trial (mean = 2 s).Within-pair stimulus position (left or right) was randomized on each trial, and presentation order of cue pairs was randomized within each block for every participant.Figure 1 depicts the progression of a trial during learning.Participants were randomly assigned to one of two conditions to balance which images were presented together during training and test, and disease labels were randomized across participants.
Following the learning phase, participants engaged in three runs of a test phase where they were asked to classify novel combinations of the previously trained cues as predicting disease 1, 2, 3, or 4. As in learning, participants were given a maximum of 3 s to make a response on each trial.Variable fixation (mean = 3 s) separated each trial.
Feedback was not provided following responses during test.The test set included the cue pairs from the original training set occurring at the previously-encountered 3:1 ratio, each cue individually, two presentations of each possible PC.PR pair, and the perfect predictors paired with a previously unpaired imperfect predictor (see Table 1, Results).In all, there were 18 unique test items and 26 total items in the test set.Presentation order of the test items was randomized for each of the three runs, with participants rating two test sets per run, for a total of 156 items throughout the test phase.
Prior to beginning the disease prediction task, participants completed two functional localizer scans.The fMRI data collected in these preliminary scans was used to both identify the location of peak BOLD responses to each visual stimulus categories (faces, objects, and scenes) for each participant and estimate distributed patterns of neural activation that were characteristic of the three stimulus categories.Each trial of the localizer task involved presenting a real-world face, object, or scene individually on the screen and asking subjects to make a button press corresponding to the appropri-ate item category within 2.5 s.Variable fixation lengths drawn from a truncated exponential (mean = 3 s) separated each trial.Over the duration of the localizer phase, subjects categorized 38 examples of each stimulus type.The black-and-white images used during the localizer runs were presented in a random order and did not include any of the stimuli used for the experimental task.

Image Acquisition
Imaging data were acquired on a 3.0 T Siemens Skyra MRI scanner at the Texas Tech Neuroimaging Institute.Structural images were acquired in the sagittal plane using

fMRI Analysis and Preprocessing
Functional data were preprocessed and analyzed using FSL (www.fmrib.ox.ac.uk/fsl).
For univariate analysis, functional data were spatially smoothed using a 6 mm FWHM Gaussian kernel.No smoothing was performed on functional data used for the representational similarity analysis (RSA; Kriegeskorte et al., 2008).First-level statistical maps were registered to the Montreal Neurological Institute (MNI)-152 template using boundary-based registration to align the functional image to the structural image, and 12 df to align the structural image to the MNI-152 template.
Three-level statistical analysis of the functional data was carried out using FSL's FEAT.At level one, within-run associations between task regressors and functional time series were examined, convolving each regressor with a double-gamma hemodynamic response function.Nuisance regressors were included motion parameters and their temporal derivatives.Additional regressors were included to censor (scrub) volumes that exceeded a framewise displacement of 0.9 mm (Siegel et al., 2014).Additionally, prewhitening was performed at level 1 to control for temporal autocorrelation in the hemodynamic response.At level 2, within-run parameter estimates for task variables were averaged for each subject using a fixed effects model.Group-level statistical maps were estimated using non-parametric tests in FSL's Randomise, minimizing the possibility of inflated family-wise error rates that may occur when using cluster correction that relies on Gaussian Random Field Theory (Eklund, Nichols, and Knutsson, 2016).Mass-based cluster thresholding was used to correct for multiple comparisons at p < .05,while a cluster-forming t-critical value of 2.52 and 6 mm variance smoothing were also applied to final statistical maps.

Multivariate fMRI Analysis
RSA was conducted using the PyMVPA toolbox (Hanke et al., 2009) and custom Python routines.To obtain trial-by-trial estimates of the hemodynamic response, we computed a β-map (Rissman, Gazzaley, and D'Esposito, 2004) for each stimulus onset using an LS-A procedure (Mumford et al., 2012), simultaneously modeling the trials of interest as separate regressors in a GLM.The same motion regressors were included in the LS-A model as the primary univariate models described above.Because the composition of noise in the scanning environment can determine which method of estimation is optimal (Mumford, Davis, and Poldrack, 2014;Abdulrahman and Henson, 2015), independent tests of classification accuracy using a localizer task identical to that used in our experiment were conducted to reveal that that the LS-A approach was preferable to LS-S (Turner et al., 2012) for the present analysis.The estimated neural activation patterns for each onset were then registered to standard space.
The use of localizer-based multi-voxel pattern analysis enabled us to measure representational similarity to individual cues within the pairs encountered by participants during learning and test (Kriegeskorte et al., 2008).If distributed activation patterns on multi-cue trials more closely resemble those characteristic of one visual image category versus the other, it would suggest that participants are selectively attending to that information.Critically, we sought to use this technique to measure attention to rare versus common cues on the ambiguous PC.PR trials that are known to elicit nonrational rare response patterns.To obtain featural selective attention predictions, we computed similarities between the patterns for each trial in the experimental paradigm and the patterns computed for face-, object-, and scene-only trials from the independent localizer.The β-series used to compute the multi-voxel patterns was spatially localized in visual stimulus category-specific ROIs by creating 6-mm spheres around subjects' peak activation within anatomically defined regions associated with category selectivity.For the learning phase, the representational analysis was restricted to a functional localizer mask in ventral visual stream that encompassed face-, object-, and scene-selective cortical regions.For both learning and test, a Pearson correlation was used to compute 1 -r correlation distances between trials on the disease prediction task with face, object, and scene patterns within each subject.Each subject-level correlation map was transformed using Fisher's Z and aggregated over trial type for statistical comparison.Descriptive and test statistics for both the MVPA and behavioral analysis were calculated using R (https://www.r-project.org/).

Behavioral Results
Learning curves over the 12 learning blocks for common and rare disease item pairs are shown in Figure 2. All subjects reached greater than 90% accuracy over the last 4 blocks (M = 98.1%,SD = 2.4%, range = 93.5 -100%).Mean choice performance in the first block was above chance (25%) for both common (M = 63.6%) and rare (M = 43.2%)cue pairs.Consistent with attentional accounts that rely on differences in early learning, a mixed effects ANOVA revealed a significant block by cue type interaction, F (1, 21) = 9.87, p = .005:the common diseases were learned more quickly than the rare diseases, with prediction accuracy for the common and rare cue pairs becoming comparable by the fifth learning block.
Results for the test phase are summarized in Table 1.Reported values are aggregated for each trial type across the associated diseases.To determine whether subjects neglected base rates on the critical trials, choice proportions for PC.PR and PC.PRo (PRo being the other rare cue that was not paired with PC during training) Figure 2. Proportions correct for common and rare disease predictions over the 12 blocks of the training phase (mean ± SEM).Note that for both disease base-rates, performance was well above chance (25%) in the first block.
were compared to those for the imperfect predictor I (Lamberts & Kent, 2007).Consistent with previous findings, ambiguous perfect predictor pairings were categorized as rare significantly more often (M = 49.9%)than the imperfect predictors (M = 29.5%),t (21) = 4.10, p < .001,suggesting that participants largely ignored the category base rates on the trials of interest.On untrained I.PRo trials, subjects chose the direct rare disease significantly more often than all other options (M = 68.0%),t (21) = 2.36, p = .028,lending further support to the idea that rare predictors were viewed as distinctive.
The perfect predictors PC and PR produced the expected strong response patterns in favor of the disease they predicted when presented alone, as did the previously-trained I.PC and I.PR trials.In addition to response probabilities, we tested whether reaction times differed on PC.PR test trials depending on whether a rare or common response was made.On these trials of interest, it was found that RTs were considerably slower when subjects made a rare response (M = 1.47 s) than when they made a common response (M = 1.27 s), t (21) = 5.73, p < .001.The observation of slowed RTs on ambiguous trials receiving rare responses indicates that rare selections may involve more conscious deliberation relative to common selections, which may be the result of simple featurebased responding.

Multi-voxel Pattern Analysis Results: Learning
The task structure (Figure 1) demanded that participants learn to selectively attend to objects or scene cues in order to accurately predict each disease outcome.To estimate neural activity indicative of attention to each cue category, we computed correlation distances between multi-voxel activation patterns on multiple-cue learning trials and patterns elicited by faces, objects, and scenes individually from the independent localizer task.It was hypothesized that representational similarity to each visual category would be strongest for perfectly predictive cues, moderate for the present but nonpredictive cues (faces) and lowest for the non-present cues.visual cortex, and that these attention-weighted representations may be measured via MVPA to test competing theories of attention (Mack et al., 2013).

Individual Differences in Behavior Associated with Cue Activation
The presence of a relationship between cue activation during learning and participants' tendencies to engage in base-rate neglect at test may lend support to either the inferential or associative theories of IBRE according to whether activation of common or rare cues during learning predicts test behavior.Based on the predictions of associative models, the degree of attention allocated to rare predictors during learning should re-Figure 3. Time-course of multivariate fMRI pattern similarities over the three learning runs (mean ± SEM).The left panel is for trials predictive of a common disease, and the right for trials predictive of a rare disease.For both graphs, the y-axis represents z-scored dissimilarity values; the axis is flipped to aid interpretation (more negative = greater neural similarity).'I' corresponds to imperfect predictors, 'PC' corresponds to common perfect predictors, and 'PR' corresponds to rare perfect predictors.
late to test behavior, with greater representational strength for PR cues being associated with a stronger tendency to choose the rare category on ambiguous PC.PR trials.
Conversely, inferential accounts would predict that relatively stronger representations of the common predictor PC during learning would lead to an increased tendency to choose the rare category on PC.PR trials at the subject level, as more robust associations between PC and the common category should result in a greater sense of mismatch between the ambiguous stimuli and this well-established rule.
Accordingly, we computed Pearson correlations to test whether the representational strength of common or rare cues during learning was associated with the extent to which participants engaged in base-rate neglect at test.Base-rate neglect values for each participant were defined as the proportion of rare responses made on ambiguous PC.PR and PC.PRo test trials.We quantified the representational strengths for common and rare cues during learning by averaging the neural similarities to common cues PC on trials that predicted a common disease, and averaging the neural similarities to rare cues PR on trials that predicted a rare disease over all 3 scanning runs (12 blocks) of the learning phase.Consistent with the predictions of inferential theories, we found that greater activation of the common predictors during learning was associated with a higher proportion of rare choices on key test trials, r = -.549,t (20) = -2.94,p = .008.Alternatively, we found no relationship between activation of rare cues PR during learning and choice proportions on IBRE trials, r = -.156,t (20) = -.705,p = .489.The association between neural similarity to PC and test behavior is depicted in Figure 4. .

Multi-voxel Pattern Analysis Results: Test
The primary research question driving our investigation concerned the distribution of attention to common and rare cues on the ambiguous test trials that produce the IBRE.
As stated above, to measure feature-based attention, we used MVPA and an independent localizer task where subjects categorized face, object, and scene images one  The interaction between neural similarity and response did not differ according to subjects' mean choice proportions, F (1, 20) = 1.31, p = .266,suggesting that the trial-level effect was consistent across participants.In sum, our RSA results for the test phase suggest that instances of base-rate neglect are driven by the consideration of common, rather than rare cues.Moreover, the presence of a similarity-by-response interaction is consistent with the inferential prediction that rare and common choices are supported by two distinct cognitive mechanisms: a controlled process that involves reasoning about the well-known category to eliminate it as an option in the face of novel stimuli, and an automatic process that supports common responding via habitual cue-outcome mappings.
. Rare responding on ambiguous trials during test was associated with slowed reaction times compared with common responding.With this finding in mind, we sought to test whether RTs on such trials covaried with subjects' neural similarity to either cue.
To examine whether response-varying similarity to cues was predictive of RT differences, we conducted a multilevel model using ln(RT) as the outcome variable and level-1 response, similarity to the chosen cue, and similarity to the unchosen cue as predictor variables within the 22 subjects over 462 observations.Here, similarities to each cue were grouped by whether the cue was consistent or inconsistent with the given response.In the context of a common response, similarity to chosen represented similarity to the common cue, and similarity to unchosen represented similarity to the rare cue; the opposite was true on trials where the rare category was selected.Logtransformed RT was used rather than the raw values to account for the strong positive skew common to RT distributions.Linear modeling including random intercepts for subjects revealed an interaction effect between response (common/rare) and similarity to the opposite cue from the unchosen category, γ = 0.075, t = 2.09, p = .037.Similarity to the cue associated with the chosen response did not vary across responses.Marginal main effects for response type (γ = -0.085,t = -2.94,p = .003)and similarity to the unchosen category (γ = -0.070,t = -2.75,p = .006)were significant.The interpretation of the positive interaction coefficient as it relates to RT is that when individuals make rare responses, greater neural similarity to the common item is predictive of faster RTs.
This pattern is consistent with the inferential hypothesis, as it suggests that selecting the rare category is facilitated by greater consideration of the common cue.

Neural Correlates of Novelty During the Test Phase
Different theoretical models attempting to explain the IBRE generally assume different quantitative patterns and psychological responses when participants are presented with novel cue combinations in test items.Our task provided an opportunity to contrast mean BOLD activation between pretrained stimulus pairs (I.PC, I.PR) and novel pairings that were matched for visual features (I.PCo, I.PRo).The contrast for Novel > Old item pairings revealed a pattern of activation often described as the fronto-parietal control network (Gao and Lin, 2012;Spreng et al., 2013;Liu et al., 2015), including anterior cingulate cortex, dorsolateral PFC, and superior parietal cortex (Figure 6; Table 2).
Consistent with studies on face, object, and scene representation, novel pairings of items were associated with greater mean activation in areas of the ventral occipitotemporal cortex known to represent such categories (Grill-Spector and Weiner, 2014).
Contrasting Old > Novel item pairings revealed activation in the vmPFC, a region associated with the application of learned category rules (Liu et al., 2015), and high relative response confidence during decision making (Schnyer, Nicholls, and Verfaille, 2005;Lebreton et al., 2015).The increased activation of lateral and dorsal PFC for the novel cue combinations and vmPFC for previously-learned cue combinations is consistent with recent results suggesting that these regions track decision evidence during categorization tasks (DeGutis and D'Esposito, 2007;Seger et al., 2015;Davis et al., 2016).

DISCUSSION
In this study, we tested the predictions of two competing hypotheses concerning the neural mechanisms that are responsible for the IBRE.Associative theories predict that base-rate neglect is driven by learned selective attention to rare cues.An alternative theory predicts that when presented with ambiguous stimuli, people choose rare categories because such items conflict with their well-established knowledge of the common category.To resolve these theories, we used MVPA to decode which cues participants are processing when they make category decisions in the face of conflicting evidence.Our neural similarity measure proved effective at detecting attention to different visual cues during both learning and test phases.In the primary analysis, we found greater neural similarity to common cues on trials where participants went against disease base-rates by categorizing an ambiguous stimulus as a rarer category (Figure 5).
These results are consistent with the predictions of theories which propose that the IBRE is attributable to an inferential rule-based process where individuals use their knowledge about the common category to eliminate it as a viable response option (Juslin et al., 2001;Winman et al., 2005).Individuals with stronger neural representations of the common predictors during learning engaged in base-rate neglect more frequently than those with weaker representations, further suggesting that it is attention to the regularities of the common category, as opposed to the distinctiveness of the rare category, that leads participants to make rare choices when faced with ambiguous evidence.
The IBRE exemplifies a case in cognitive neuroscience where competing models that predict essentially the same behavioral patterns make very different assumptions about the cognitive processes, and accordingly, brain states, involved in producing the behavior.Our findings from the test phase (Figure 5) represent a critical step forward in  an emerging area of research using multivariate fMRI to reveal that qualitatively distinct brain states may reflect the use of multiple response strategies in the face of identical stimuli (Mack et al., 2013).While the strength of neural activity for each cue followed a predictable pattern corresponding to their behavioral relevance during learning, our test results show that when participants neglect the category base-rates, patterns of neural activation point to enhanced processing of the common cue (PC) compared to the rare cue (PR), consistent with the inferential hypothesis that activating knowledge of rare responding.No differences in featural activation were observed on trials that received a common response, suggesting that only rare responses depend on deliberative processes: inferential accounts of the IBRE assume that when participants respond according to the base rates, their responses are guided by basic feature-based similarity (Winman et al., 2005).
In line with recent evidence, we identify representational changes indicative of a selective attention mechanism that tracks the predictive value of different stimulus features during learning (Mack et al., 2016).The data suggest that attention to these goalrelevant features enhances their representation in object-selective cortex quickly and persistently in a simple rule-based category learning task (Figure 3).However, our findings build upon these results in a crucial way by demonstrating that attention to features is not static, but rather dynamically changes based on participants' cognitive inferences and their behavior.Until now, many of the fine-grained dynamics governing how participants choose and switch between strategies are lost because they are unidentifiable in behavior.For example, it is impossible to discern based on behavior alone which hypotheses participants are testing on individual trials before they respond in a consistent manner.Our results suggest that it may be possible to use MVPA methods to synchronize neural and behavioral processes to pinpoint the inferences subjects considering at a given time.
experiment adds to the broader literature on the IBRE by demonstrating the importance of examining processing differences that are dependent on subjects' responses.In the behavioral literature it is common to focus on the average distributions of responses, such as average proportion of rare choices, as opposed to examining differences between choice options for the same stimuli.Here, we show that the way people approach ambiguous test stimuli is not consistent across trials, as individuals appear to engage qualitatively distinct neural processes when producing rare versus common responses.Neither attention learning (Kruschke 1996, Kruschke 2001b) nor existing inferential models of the IBRE (Juslin et al., 2001) can adequately explain the present fMRI results, as both rely on a single mechanism that, on average, produces choice probabilities favoring the rare category.The observed pattern of RT in this experiment reinforces the idea that qualitatively distinct mechanisms may support common versus rare choices, as we found a significant slowing of RTs on ambiguous trials when subjects chose the rare disease.This result suggests that choosing the rarer response places more cognitive demand on participants, as predicted by our inferential account.While RT differences for rare versus common responses have not been examined in previous studies of the IBRE, we expect that the RT distributions of existing data would reveal response-dependent patterns similar to those found here.Future research will benefit by incorporating such results into computational models.
Contrasting mean BOLD activation on test trials composed of previously unseen I.PCo/I.PRo stimulus combinations with previously learned pairs revealed activation in a fronto-parietal network associated with rule-based category learning (Filoteo et al., 2005;Seger andCincotta, 2006, Soto et al., 2013) and decisional uncertainty (DeGuits and D 'Esposito, 2007;Seger et al., 2015;Davis et al., 2016).Alternatively, exposure to previously learned pairs at test was accompanied by activation of regions associated with high confidence / high relative decision evidence (Schnyer, Nicholls, and Verfaille, 2005;Lebreton et al., 2015;Davis et al., 2016) and the application of familiar rules (Boettiger and D'Esposito, 2005).Thus the current mean BOLD activation results are in accordance with a number of recent studies examining decision-related processes in categorization.Moreover, they are consistent with the idea that novel pairings may involve more active deliberation whereas the previously learned pairs do not.
One critical question is whether the trade-off between lateral/dorsal PFC activation for novel stimuli and vmPFC activation for previously learned stimuli reflects wholly separate processes or opposite ends of a common on-task/off-task network.Consistent with the idea that these activation patterns may indicate distinct processes, the regions activated in response to pair novelty mirror those associated with model-based learning in the reinforcement learning literature (Gläscher et al., 2010;Niv et al., 2015), which involves forming high-level representations of the contingencies between stimuli and outcomes (Daw, Niv, and Dayan, 2005).Likewise, more habitual, model-free learning has been associated with the presence of prediction error signals in ventral striatum (Gläscher et al., 2010;Niv et al., 2015).The prefrontal regions activated for the previously learned stimuli in our study are part of a dopaminergic corticostriatal loop (Seger, 2008), and thus may reflect the maintenance of previously reinforced behaviors in the absence of feedback (Noonan et al., 2012).
The trade-off between model-based and model-free choice may also be useful for understanding the stimulus-and response-dependent neural activation patterns found in the present study at a mechanistic level.For example, recent findings in associative learning have shown that participants who tend to engage in rule-based generalization for ambiguous cue pairings in a patterning task (Shanks and Darby, 1998) tend to rely more on model-based choice (Don et al., 2016).Within individuals, modelbased inference may be favored in cases of high uncertainty, while competing for control over responding with the habitual system from trial to trial (Daw, Niv, and Dayan, 2005).A similar dynamic may be involved in producing the stimulus-and responsedependent neural activation patterns found in the present study: when viewing test stimuli, participants might be prompted to use explicit rules only when a mismatch is detected between two competing cue-response mappings.Conversely, treating the pair of cues as a unitary stimulus may allow respondents to rely on efficient, wellestablished response patterns that reliably favor the common category.A question for future research is whether models that allow for flexible system recruitment (Ashby et al., 1998;Dayan, 2007) could account for both the behavioral and neural dynamics of the IBRE.
One possible explanation for our multivariate results for the key ambiguous (PC.PR) cue pairs is that activation of the common cue is attributable to uncertaintyrelated attention.Attention learning models rely on the formalisms of Mackintosh (1975), where the role of attention is to selectively process cues with the highest predictive value.Alternatively, Pearce and Hall (1980) proposed that attention instead al-locates resources to stimuli that have been associated with the most uncertainty, promoting learning and error reduction on future trials.It has been recently discussed that both forms of attention may coexist in associative learning: automatic attention akin to that proposed by Mackintosh (1975) that converges on strong predictors, and another concerned with the controlled selection of stimuli that are unreliable (Pearce and Mackintosh, 2010;Luque et al., 2016).In an event-related potential (ERP) study of the IBRE, Wills and colleagues (2014) found evidence to suggest enhanced attention to cue PR relative to PC following training.However, this evidence came from an index of early visual processing, where the present analysis examines the stimulus presentation period as a whole.Accordingly, it is possible that a rapid form of attention focusing on predictive utility was present during our experiment, with this effect being washed out by the subsequent effects of controlled attention.If both attentional processes play a role in base-rate neglect, future studies combining ERP and fMRI might help to dissociate these separate signals.
A critical question for future research is how changes in cognitive state from dual task interference may influence the IBRE.Because our inferential account is assumed to depend on deliberative processes such as representing rule information in working memory, it follows that requiring participants to engage in a concurrent task that demands WM resources should diminish their ability to apply such rules, resulting in fewer rare responses on ambiguous PC.PR trials.Although Lamberts and Kent (2007) showed that neither the inclusion of a secondary task nor forcing speeded responses during test disrupted the IBRE, the effect requires specific, asymmetrical learning conditions that suggest its occurrence is unlikely to be a product of test strategy alone (Don and Livesey, 2016).Rather, experiments in category learning (Waldron and Ashby, 2001;Zeithamova and Maddox, 2006) and predictive learning (Wills et al., 2011) suggest that adding cognitive load during the learning phase may disrupt rule acquisition and result in more habitual, feature-based generalization patterns at test.
The present findings speak to a larger debate surrounding whether human associative learning is necessarily tied to reasoning (Mitchell et al., 2009).We do not take the apparent involvement of higher-level processes in a behavior that can be captured by associative models to mean that humans are purely propositional beings.Rather, the results at hand lend support to the general view that human learning involves both propositional reasoning and some fundamental associative learning processes, while illustrating that neural representations of the same physical stimulus can differ substantially depending on which response (or strategy) is selected at a given time.
While associative theories of base-rate neglect suggest that simple errorreduction mechanisms lead individuals to form an especially strong association between rare cues and their outcomes, here we provide neurobiological evidence to the contrary.Rather, in a novel application of MVPA techniques, we demonstrate that when selecting a rarer category in the face of ambiguous evidence, people tend to focus on the cues they have encountered the most frequently.Taken together, these findings strongly suggest that high-level processes contribute to the IBRE.Accordingly, computational models that posit a purely associative explanation for base-rate neglect will require modification to account for the pattern of neural data obtained in this experiment.Contrasting the representational similarities between co-present (or unpresent) visual cues in occipitotemporal cortex appears to be a promising avenue to test theo-

Figure 1 .
Figure 1.Abstract task design and an example trial.In the headings, I = imperfect predictor, PC = common perfect predictor, PR = rare perfect predictor.The second row refers to the specific stimuli used for each cue: F = face, S = scene, O = object.Each following row corresponds to a learning trial, with a "1" indicating the presence of the cue and "0" indicating its absence.
Nine task regressors were included in a level-1 model for each of the unique trial types during test.This included onsets for I, PC, PR, PC.PR trials receiving a common response, PC.PR trials receiving a rare response, previously trained common pairs I.PC, previously trained rare pairs I.PR, and novel pairs I.PCo and I.PRo.Nuisance regressors were included for trial onsets where participants failed to make a response within the cutoff time.Two contrasts of interest based on this model were PC.PR trials with a rare response > PC.PR with a common response, and novel combination I.PCo/ I.PRo trials > old I.PC/ I.PR trials that were pretrained, as different theories of the IBRE have disagreed over the existence of response-dependent strategy use and the role of novelty/uncertainty in the effect.No significant activations were observed for rare versus common responses on PC.PR trials, thus this contrast is not discussed further in the paper.

Figure 3
depicts the timecourse of pattern similarity for predictive, non-predictive, and non-present cues during learning for both rare and common disease pairs.Lower values indicate less correlation distance or greater similarity.As predicted, a one-way within-subjects ANOVA collapsed across trial type revealed that neural similarity to perfectly predictive cues was the strongest (M = -0.071),followed by the non-predictive but present cues (M = 0.043), with the non-present cues eliciting the lowest similarity values (M = 0.139), F (2, 42) = 41.4,p < .001.This pattern of results extends recent research suggesting that category exemplars are weighted according to their relative predictive values in ventral at a time.We then computed similarities between activation patterns elicited during the test phase to those characteristic of each image type.The feature-based attention measure accurately predicted attention to control items during test; similarity values for objects were significantly greater on object-only (combined across O1, O2, and O1+O2) trials than scene-only trials (combined across S1, S2, and S1+S2), t (21) = -2.82,p = .010,and vice versa, t (21) = -5.61,p < .001.To test whether patterns of at-

Figure 4 .
Figure 4. Association between activation of common predictors during learning and individual differences in base-rate neglect.The y-axis represents the proportion of rare responses made on ambiguous PC.PR trials during the test phase.The x-axis represents z-scored dissimilarity values for the common predictor PC over the course of learning for each subject; the axis is flipped to aid interpretation (more negative = greater neural similarity).'**' = p < 0.01.

Figure 5 .
Figure 5. Results from the representational similarity analysis for ambiguous test trials PC.PR.The pair of bars on the left reflect activation patterns on trials receiving common responses, while the pair of bars on the right reflect activation patterns on trials where the participant made a rare response.Grey bars depict neural similarity to the common cue, and black bars depict neural similarity to the rare cue.The y-axis represents z-scored dissimilarity values (mean ± SEM); the axis is flipped to aid interpretation (more negative = greater neural similarity).'**' = p < 0.01.

Figure 6 .
Figure 6.Univariate fMRI results contrasting BOLD activation during the test phase for novel stimulus pairs (I.PCo and I.PRo) with pairs that had been previously encountered during learning (I.PC and I.PR).

Table 1 .
Item combinations and choice proportions for the test phase.

Table 2 .
Activated clusters and peaks (MNI coordinates) for the fMRI results in Figure6.