Neural Representations of Sensory Uncertainty and Confidence Are Associated with Perceptual Curiosity

The study investigates how neural representations of events evoke states of curiosity. Prior work shows that curiosity about trivia answers engages motivation and reward circuits and is inversely related to confidence, suggesting uncertainty as a trigger for curiosity. However, the neural mechanisms linking sensory/semantic uncertainty, confidence, and curiosity remain unclear, particularly because uncertainty in higher-level representations (e.g., semantic information) is not well understood. Leveraging the literature on sensory uncertainty, the authors examine perceptual curiosity using ambiguous visual stimuli and fMRI. They hypothesize that the occipitotemporal cortex (OTC) provides a multivariate representation of certainty/uncertainty that is read out by frontal regions (vmPFC, ACC) into a univariate confidence signal, which in turn relates to curiosity. The goals are to (1) establish the behavioral relationship between confidence and perceptual curiosity, (2) develop a trial-wise neural metric of OTC multivariate certainty about visual categories without assuming specific tuning curves, and (3) test whether vmPFC and/or ACC mediate the relationship between OTC certainty and curiosity.

- Curiosity research using trivia questions shows that curiosity engages motivation/reward regions and enhances attention and memory (Kang et al., 2009; Gruber et al., 2014; Baranes et al., 2015; Murphy et al., 2021). Curiosity and confidence are inversely related, supporting theories that uncertainty drives curiosity (Berlyne, 1954; Loewenstein, 1994; Golman & Loewenstein, 2018). - Sensory uncertainty is represented across primary and associative visual cortex. V1 population codes reflect both stimulus features and uncertainty (Ma et al., 2006; van Bergen & Jehee, 2021; Geurts et al., 2022; Meyniel et al., 2015), with uncertainty conveyed by distributed multivariate patterns (Russell & Reale, 2019). - Confidence is encoded in higher-order frontal areas, especially vmPFC and ACC, with univariate BOLD scaling to confidence (Lebreton et al., 2015; Gherman & Philiastides, 2018; Bang & Fleming, 2018). Frontal cortex supports metacognition (Shimamura, 2000; Fleming et al., 2010) and curiosity-driven control (Jepma et al., 2012; Shenhav et al., 2013; Silvetti et al., 2023). Readout hypotheses propose frontal areas transform multivariate sensory codes into lower-dimensional confidence signals (Meyniel et al., 2015; Pouget et al., 2016; DiCarlo & Cox, 2007; Russo et al., 2018). Geurts et al. (2022) linked V1 certainty to confidence in orientation tasks. - Gaps: prior decoding of uncertainty relies on assumptions (e.g., cosine tuning in V1) and simple stimuli. The role of higher-level visual areas (OTC) and their uncertainty representations in curiosity has not been established. No prior quantification of OTC certainty for naturalistic categories or its link to curiosity.

Participants: 32 right-handed adults (17 female), age 18–35 (mean 27.2±4.5), education 13–25 years (mean 16.1±2.8), normal or corrected vision, recruited at Columbia University with IRB approval. Compensation: fixed $40 (non-incentivized by performance). Stimuli: 42 animal and 42 man-made object images (Konkle Lab database), normalized for contrast and luminance (SHINE toolbox). Texforms generated using texture synthesis (Deza et al., 2019): measure first- and second-order statistics across pooling regions; synthesize from white noise via stochastic gradient descent (100 iterations). Pooling factor fixed at 0.28 to control distortion. Low-level features computed: luminance (mean intensity), RMS contrast (std of luminance), spatial frequency via FFT-based power spectra slope (Eskicioglu & Fisher, 1995; Li et al., 2001; Flitcroft et al., 2020). Design and procedure: - Perceptual curiosity task (4 runs, 84 trials total): Each trial: texform presented 4 s; participants generate best guess of original image. Then rate confidence (0–100) in their guess and curiosity (0–100) to see the original, each self-paced up to 5 s using MR-compatible trackball with randomized initial slider position. Then see the undistorted original image for 2 s. Categories (animal/object) equally likely. Fixed payment ensured ratings independent of monetary incentives. - Localizer task (1 run, unannounced): 24 miniblocks (12 animal, 12 object) of clear images (not used in main task). Each miniblock: 20 images at 333 ms on/333 ms ISI. One-back task during presentation. Fixation blocks (13 s) between miniblocks. Behavioral analysis: Linear mixed-effects models (MATLAB fitlme). Models included participant-specific random intercepts and slopes. Core models: - Curiosity ~ confidence + confidence^2 + (1 + confidence + confidence^2 | participant) vs linear-only. Also tested covariates luminance, contrast, spatial frequency. MRI acquisition: 3T Siemens Prisma, 64-channel coil. Functional EPI: TR=2 s, TE=30 ms, FA=80°, multiband factor=3, voxel size=2 mm isotropic, 69 axial slices (14° transverse to coronal), posterior→anterior phase encoding. Five runs total (4 task + 1 localizer). Structural T1 MPRAGE 1.0 mm isotropic. fMRI preprocessing: FSL (FEAT, FNIRT, fslmaths). Steps: brain extraction (BET), motion correction (MCFLIRT), high-pass filtering (cutoff 100 ms as specified), spatial smoothing (3 mm FWHM Gaussian), double-gamma HRF convolution, FILM prewhitening. Nonlinear registration to 1 mm MNI152 (12 DOF). Analyses continued in MATLAB. Code available upon request. ROIs: vmPFC ROI from Mackey & Petrides (2014) excluding corpus callosum overlap; ACC and OTC ROIs from Harvard–Oxford Atlas (threshold 50); all in 1 mm MNI space. GLMs: - GLM #1 (localizer miniblocks): Each of 24 miniblocks modeled as boxcar (16.66 s) with motion regressors and derivatives. Beta maps per miniblock. Derived OTC animal and man-made category templates by averaging beta maps for corresponding miniblocks within OTC ROI. - GLM #2 (single-trial task): Each texform period modeled as 4 s boxcar in separate regressors per trial. Nuisance regressors for rating periods (confidence and curiosity; boxcar duration = RT) and clear-image presentations (2 s). Motion regressors and derivatives included. Four runs modeled separately. Extracted univariate beta (mean across voxels) for vmPFC, ACC, OTC per texform. Extracted multivoxel OTC patterns per texform. OTC certainty metric: For each texform trial, compute Pearson’s r_a (correlation with animal template) and r_mm (correlation with man-made template). Define OTC Certainty = mean(r_a, r_mm) × |r_a − r_mm|. This captures model certainty (mean term) and approximation certainty (absolute difference term). Values and predictors z-scored within participant. Neurobehavioral modeling: Mixed-effects models tested: - Confidence ~ OTC Certainty (+ quadratic term tested) with random slopes/intercepts. - Curiosity ~ OTC Certainty (+ quadratic term tested) with random slopes/intercepts. - Similar models for vmPFC/ACC activity vs confidence, curiosity, OTC Certainty. Covariate models included luminance, contrast, spatial frequency. Model comparison: Bayesian Information Criterion (BIC); ΔBIC ≥ 2 considered evidence in favor (Raftery, 1995). Mediation analysis: Baron & Kenny approach (Wager et al., 2009) with bootstrapping (1,000 iterations). Tested whether frontal ROI activity (vmPFC or ACC) mediates the relationship between OTC Certainty and curiosity. Variables z-scored within participant. Reported parameters: a (OTC Certainty→ROI), b (ROI→curiosity), c (OTC Certainty→curiosity), c′ (mediation-controlled effect), and statistical comparisons of c vs c′.

Behavioral: - Curiosity and confidence showed a negative quadratic relationship: mixed-effects model significant for both terms (β_linear = −13.46, p < 0.0001, 95% CI [−15.8, −11.0]; β_quadratic = −5.60, p < 0.0001, 95% CI [−7.21, −3.99]). Quadratic model outperformed linear (BIC_quadratic − BIC_linear = −180). Demographics (age, education, sex) showed no effects; low-level image properties did not predict curiosity or confidence; adding luminance, contrast, spatial frequency preserved the negative quadratic relation. OTC category representation and certainty: - OTC localizer patterns: within-category correlations > between-category (mean r = 0.80 vs 0.58; p = 0.008), confirming reliable category templates. - In task, texform-evoked OTC patterns correlated more with matching than non-matching category templates (mean r = 0.50 vs 0.43; paired t-test p = 0.01), validating r_a and r_mm. - OTC Certainty metric properties illustrated: independent contributions of mean and absolute-difference terms (r ≈ −0.1 between terms; p = 0.04). Links between OTC Certainty, confidence, and curiosity: - Confidence increased with OTC Certainty: β = 1.95, p = 0.0008, 95% CI [0.80, 3.09]; linear model preferred (BIC_quadratic − BIC_linear = 17). - Curiosity decreased with OTC Certainty: β = −1.21, p = 0.007, 95% CI [−2.08, −0.33]; linear model preferred (BIC_quadratic − BIC_linear = 23). - Univariate OTC activity did not predict curiosity (β = −0.23, p = 0.63, 95% CI [−1.17, 0.72]). - Low-level image properties did not account for the OTC Certainty–curiosity link; covariate model retained significant OTC Certainty effect on curiosity (β = −1.15, p = 0.011). - V1 lacked category specificity (mean r true vs alternate template: 0.57 vs 0.57; p = 0.90). V1-derived certainty did not predict confidence (β = 0.12, p = 0.8) or curiosity (β = 0.04, p = 0.9). Frontal univariate signals and mediation: - Confidence positively associated with vmPFC (β = 3.32, p < 0.0001, 95% CI [1.80, 4.85]) and ACC (β = 2.34, p < 0.0001, 95% CI [0.88, 3.79]); linear models preferred (vmPFC ΔBIC = 23; ACC ΔBIC = 26). - Curiosity negatively associated with vmPFC (β_linear = −2.43, p < 0.0001, 95% CI [−3.58, −1.28]; modest quadratic component: β_quadratic = −0.66, p = 0.005; ΔBIC = −3) and with ACC (β_linear = −1.14, p < 0.0001, 95% CI [−2.26, −0.02]; β_quadratic = −0.53, p = 0.005; ΔBIC = 0). Low-level features did not significantly predict vmPFC/ACC activity; relationships with confidence/curiosity remained significant with covariates. - OTC Certainty predicted vmPFC (β_linear = 5.55, p < 0.0001; ΔBIC = 17) and ACC activity (β_linear = 9.00, p < 0.0001; ΔBIC = 3). - Mediation (bootstrapped): Direct association c between OTC Certainty and curiosity was significant (c = −0.046, p = 0.009). With vmPFC as mediator, c′ = −0.036 (p = 0.08); c′ significantly reduced vs c (c′−c = 0.0109; two-tailed p < 0.001), indicating significant mediation by vmPFC. With ACC as mediator, c′ = −0.044 (p = 0.02); c′ not different from c (c′−c = 0.002; p = 0.30), indicating no mediation by ACC. c′ in ACC model was stronger than c′ in vmPFC model (KS test ks = 0.60; p < 0.001). Overall: Perceptual curiosity peaks at lower confidence; multivariate OTC Certainty aligns with higher confidence and lower curiosity; vmPFC and ACC univariate activity track confidence (positive) and curiosity (negative). vmPFC mediates the link between OTC Certainty and curiosity, while ACC relates to curiosity via a distinct, non-mediating mechanism.

Findings support a multistage mechanism wherein multivariate sensory certainty in higher-level visual cortex (OTC) is transformed into a univariate confidence signal in vmPFC that relates inversely to curiosity. Behaviorally, perceptual curiosity mirrors epistemic curiosity with a negative quadratic relationship to confidence. The new OTC Certainty metric—combining model and approximation certainty inspired by machine learning—captures trial-wise certainty about animate vs inanimate categories without assuming specific tuning curves. This multivariate certainty is positively related to confidence and negatively to curiosity, supporting the hypothesis that uncertainty drives curiosity. Frontal correlates show vmPFC and ACC encode univariate signals related to confidence (positively) and curiosity (negatively), consistent with metacognitive roles. Mediation analyses indicate vmPFC specifically mediates the OTC Certainty–curiosity association, whereas ACC contributes to curiosity through a distinct route (e.g., control or information-gathering mechanisms) not directly explained by OTC Certainty. Univariate OTC activity did not predict curiosity, underscoring that curiosity is tied to multivariate sensory representations rather than overall response magnitude. Control analyses rule out confounds from low-level image properties and V1 category processing, indicating that the observed relationships reflect higher-level category representations. These results extend prior work beyond trivia-based curiosity and low-level feature uncertainty (e.g., V1 orientation) to complex category-level representations in OTC. They suggest that the brain may implement representational untangling, converting high-dimensional sensory codes into low-dimensional signals suitable for action control (e.g., regulating curiosity, attention).

The study introduces a novel, trial-wise metric of multivariate certainty in OTC for visual categories and demonstrates that this measure relates to subjective confidence and curiosity. Perceptual curiosity peaks at lower confidence; higher OTC Certainty predicts higher confidence and lower curiosity. vmPFC and ACC univariate signals track confidence and curiosity, but only vmPFC mediates the link between OTC Certainty and curiosity, revealing a pathway by which sensory uncertainty can generate curiosity. Future directions include: extending multivariate certainty decoding to exemplar-level representations and non-visual/semantic domains; probing additional pathways that might link OTC certainty to curiosity (potentially linear routes); clarifying ACC’s role in curiosity under tasks requiring effortful information seeking and cognitive control; and exploring how different levels of the visual hierarchy (e.g., V1 feature uncertainty) contribute under varied task demands. These advances could generalize to probabilistic representations across domains and inform computational models of how uncertainty guides information-seeking behavior.

- Granularity of certainty: OTC Certainty was computed at the category level (animals vs man-made objects), whereas participants may have formed exemplar-level guesses; certainty at exemplar resolution was not measured. - Model form: The vmPFC showed a modest quadratic relation to curiosity, while the OTC Certainty–curiosity relation was linear; this mismatch may reflect statistical noise or additional untested pathways mediating the relationship. - Low-level visual features: Although luminance, contrast, and spatial frequency did not explain results, other unmeasured low-level factors could contribute in different contexts. - V1 role: V1 did not exhibit category specificity or predict behavior here, but V1 may encode uncertainty about local features not captured by the analysis and could influence curiosity in tasks targeting local feature judgments. - Mediation inference: Mediation analyses are correlational and do not imply direct anatomical connectivity or causality; unmeasured variables or indirect pathways may contribute. - Sample characteristics: Healthy, right-handed young adults; generalizability to broader populations remains to be established.