Computational and Neuronal Basis of Visual Confidence

The paper addresses how the brain evaluates the reliability of its own perceptual decisions—specifically, the computational and neuronal basis of visual decision confidence. Perception operates under uncertainty due to coarse, noisy, and indirect sensory inputs, making errors inevitable. Beyond selecting the most plausible world state, the brain must track the certainty of its interpretations, giving rise to decision confidence. The review adopts the view that confidence aligns more closely with subjective decision reliability (likelihood of making the same choice again) rather than the posterior probability of correctness. It highlights that while humans uniquely verbalize confidence, nonhuman animals (e.g., monkeys and rats) display confidence-mediated behavior, enabling mechanistic study. Recent advances—explicit confidence reports in macaques, principled incentivization schemes, and signal processing methods for human noninvasive recordings—allow tighter links between computational models and neural substrates. The purpose is to synthesize behavioral paradigms, process models (static and dynamic), and neural evidence to outline how sensory population activity is transformed into metacognitive assessments, and to situate confidence within broader performance-monitoring functions.

The review synthesizes several strands of literature: (1) Behavioral paradigms for probing confidence in binary perceptual tasks, spanning explicit reports (direct ratings or comparative judgments) and implicit measures (waiting-time and opt-out tasks), with emphasis on incentivization to ground reports in reward-based optimal criteria. It distinguishes manipulations of stimulus strength (distance from category boundary) from stimulus reliability (factors like contrast, size, duration, eccentricity) as orthogonal contributors to task difficulty, both of which confidence must track. (2) Static signal detection theory (SDT)-based models, especially decision reliability frameworks where confidence derives from the distance-to-criterion normalized by decision-variable (DV) uncertainty. The CASANDRE model formalizes confidence as a noisy estimate of decision reliability and identifies meta-uncertainty (uncertainty about DV uncertainty) as the key determinant of metacognitive quality. Comparative analyses across studies show meta-uncertainty increases with the number of stimulus reliability levels, and model-based metrics outperform agnostic measures (e.g., meta-d′, meta-I) which conflate metacognition with sensitivity and bias. (3) Dynamic sequential sampling (SS) models (evidence accumulation, drift-diffusion, race models) that relate confidence to the balance of evidence at decision time, accounting for relationships among accuracy, reaction time (RT), and confidence. Race models map the losing accumulator’s state to log posterior odds of being correct, capturing human and macaque behavior in RT tasks with explicit confidence. The literature debates whether confidence is computed online during deliberation versus postdecisional accumulation; evidence supports both, suggesting provisional estimates can coexist with postdecision refinement. (4) Neural correlates across species: macaque V1 population activity can predict both choice and confidence, with nonlinear transformations better decoding confidence than linear ones, consistent with distinct downstream computations. In macaque LIP, neural dynamics support concurrent evolution of choice and confidence signals, with population decoding revealing simultaneous ramping for both. (5) Human noninvasive studies identify EEG correlates of evidence accumulation that predict subjective confidence independently of accuracy, RT, and evidence strength, and support architectures with separate (but interacting) accumulators for choice and confidence. (6) Links to performance monitoring: overlapping medial frontal cortex (MFC) substrates for error detection, conflict monitoring, and confidence suggest shared metacognitive mechanisms, while emphasizing differences between monitoring relative to internal goals versus external ground-truth states.

As a narrative review, the paper collates and evaluates methods from behavioral, computational, and neurophysiological studies. Key methodological elements include: - Behavioral task designs: binary perceptual judgments (e.g., orientation or motion discrimination, dot color ratio) with explicit confidence reports (direct scales or pairwise comparisons) or implicit measures (waiting time, opt out). Tasks often manipulate stimulus strength (distance from category boundary) and orthogonal stimulus reliability factors (contrast, duration, size, eccentricity), sometimes under incentivization schemes where high confidence is high risk/high reward to yield reward-rational criteria. - Model-based analysis: static SDT frameworks compute primary choices by comparing a noisy DV to a decision criterion; confidence arises from a normalized distance to criterion reflecting decision reliability. The CASANDRE model introduces meta-uncertainty to capture noise in reliability estimation and fits human and monkey choice-confidence data, enabling principled metacognitive metrics. Dynamic SS frameworks (drift-diffusion/race models) track concurrent accumulations for alternatives; confidence is read from the balance of evidence (e.g., losing accumulator), providing time-dependent predictions of accuracy-RT-confidence relationships. - Neurophysiology in nonhuman primates: population recordings in V1 during fine orientation discrimination with explicit confidence reports; linear vs nonlinear decoders trained to predict choices and confidence on single trials, assessing whether distinct transformations of sensory activity underlie decision content vs confidence. LIP recordings during RT motion discrimination with simultaneous confidence report; analysis of ramping (accumulator-like) dynamics, autocorrelation signatures of accumulation, and time-resolved logistic decoding for choice and confidence to assess parallel emergence. - Noninvasive human electrophysiology: EEG decoding of accumulation-like signals during tasks with opt-out or explicit confidence; linking slopes/amplitudes of central parietal positivity and motor signals to subjective confidence, accuracy, RT, and evidence strength. Multimodal (EEG-fMRI) localization (e.g., VMPFC) of early confidence signals. - Comparative and perturbation approaches: microstimulation/optogenetic modulation in MT during opt-out tasks to test causal effects on both choice and confidence; pulvinar inactivation to dissociate confidence signals from choice sensitivity. - Quantitative evaluation: fitting model predictions to behavioral psychometric/chronometric and confidence functions; relating decoded neural decision variables to confidence on a trial-by-trial basis (including within-condition analyses to test U-shaped DV-confidence relations).

• Confidence as decision reliability: Across SDT-based models, a normalized distance-to-criterion yields a principled confidence variable reflecting subjective decision reliability. Recent models (e.g., CASANDRE) demonstrate that confidence reports across varied stimulus strengths and reliabilities can be captured by a single confidence–consistency relationship. - Meta-uncertainty constrains metacognition: CASANDRE identifies meta-uncertainty as the primary determinant of metacognitive quality; empirical reanalyses show meta-uncertainty increases with the number of stimulus reliability levels across multiple studies, indicating task dependence of metacognition. - Training and species comparisons: In a shared incentivized task, well-trained macaques initially outperformed novice humans in metacognitive ability; humans required two additional 1,100-trial sessions to match monkeys’ performance, providing rare evidence of metacognitive learning. - Dynamic balance-of-evidence rule: Race models link confidence to the state of the losing accumulator (mapped to log posterior odds), predicting lawful relationships among accuracy, RT, and confidence observed in human and macaque RT tasks with explicit confidence reports. Confidence is typically higher for correct than incorrect choices, and higher confidence decisions are faster. - Sensory population basis with distinct transformations: In macaque V1, both choice and confidence can be decoded from sensory population activity; nonlinear decoders consistently outperform linear ones for confidence (but not for choice), indicating specialized nonlinear transformations for confidence estimation distinct from the linear readout used for categorical choice. The neurally decoded confidence variable shows a U-shaped relation with the DV, including within fixed stimulus conditions. - Parallel evolution in decision circuits: In macaque LIP, accumulator-like dynamics and autocorrelation signatures indicate concurrent evolution of signals for both choice direction and confidence level. Time-resolved decoding shows simultaneous ramp-up for choice and confidence, with the choice decoder peaking slightly earlier; trial-by-trial decoded choice strength correlates with decoded confidence for in-RF (winning) choices. - Human EEG correlates: Accumulation-like EEG signals (e.g., central parietal positivity) predict subjective confidence independently of accuracy, RT, and evidence strength; postdecision accumulation signatures also relate to metacognitive judgments. Evidence supports architectures with separate but interacting accumulators for choice (motor cortex) and confidence (parietal signals), allowing confidence to modulate decision policies. - Performance monitoring links: Overlapping MFC substrates support error and conflict monitoring and may interface with domain-general confidence judgments, though confidence about external states and error monitoring relative to internal goals remain conceptually distinct.

The synthesized evidence supports a hierarchical, parallel-processing account of visual confidence. The same sensory population activity that supports perceptual decisions is transformed through distinct computations to estimate decision reliability. Static models explain the dependence of confidence on both stimulus strength and reliability and formalize metacognitive limitations via meta-uncertainty. Dynamic models account for the tight coupling between confidence and decision time through the balance of evidence at decision termination, while accommodating both online and postdecisional contributions. Neurophysiological findings bridge models to biology: V1 population codes contain sufficient structure for downstream circuits to compute confidence via nonlinear readouts; downstream decision areas like LIP exhibit concurrent accumulator-like dynamics for both choice and confidence, consistent with parallel deliberation that can shape ongoing behavior. Human EEG results converge on similar principles, revealing early-arising confidence signals and possible dual-accumulator architectures wherein confidence can regulate decision leakiness or thresholds. Together, these findings address the central question of how the brain computes and uses confidence: confidence is not a detached introspective judgment but an integral computation derived from the same evidence stream, enabling adaptive control of decisions, learning, and strategy selection. The relation to performance monitoring suggests a broader metacognitive framework in which internal error/conflict signals and perceptual reliability estimates inform behavior, potentially sharing representational geometries across domains.

This review consolidates computational and neural evidence that perceptual decision confidence arises from hierarchical transformations of sensory population activity, producing an estimate of decision reliability that evolves in parallel with decision formation. Key contributions include: formal process models (static and dynamic) that jointly capture confidence–accuracy–RT relationships; principled, model-based metrics of metacognitive ability (e.g., meta-uncertainty); and neural demonstrations that distinct transformations of sensory codes and concurrent accumulation dynamics underpin confidence in primates. Future directions include unifying static and dynamic accounts into a comprehensive process model; generalizing mechanisms across modalities, tasks, and confidence reporting behaviors; elucidating how confidence guides evidence sampling and policy adjustments; establishing mechanistic links between probabilistic sensory codes and subjective confidence; mapping how confidence signals influence hierarchical decision-making and learning; and clarifying circuit-level relations between perceptual metacognition and action selection conflict monitoring.

• Scope of tasks and regions: Current neural evidence is derived from a limited set of tasks (e.g., motion and orientation discrimination) and brain areas (V1, MT, LIP, pulvinar, MFC), limiting generalizability across modalities and contexts. - Measurement constraints: Noninvasive human methods (EEG, fMRI) provide coarse spatial or temporal resolution; mapping model accumulators to specific neural populations remains indirect. - Incentivization and reporting: Human studies often lack explicit incentivization, potentially inducing idiosyncratic or suboptimal strategies; cross-study variations in confidence scales and report formats complicate comparisons. - Model identifiability: Alternative sources of confidence noise (e.g., fluctuating criteria) can mimic meta-uncertainty; distinguishing online vs postdecisional contributions can be task-dependent and difficult to dissociate. - Causal mapping uncertainties: The correspondence between neural populations (e.g., LIP signals) and model constructs (winning vs losing accumulators) is not one-to-one; causal roles of identified regions in confidence computation require further perturbation and circuit-level studies.