Table Of Content

Phil. Trans.R.Soc. B(2012) 367, 1310–1321 doi:10.1098/rstb.2011.0416 Review Metacognition in human decision-making: confidence and error monitoring Nick Yeung* and Christopher Summerfield DepartmentofExperimentalPsychology,UniversityofOxford,SouthParksRoad,OxfordOX13UD,UK People are capable of robust evaluations of their decisions: they are often aware of their mistakes even without explicit feedback, and report levels of confidence in their decisions that correlate with objective performance. These metacognitive abilities help people to avoid making the same mistakestwice,andtoavoidovercommittingtimeorresourcestodecisionsthatarebasedonunreli- ableevidence.Inthisreview,weconsiderprogressincharacterizingtheneuralandmechanisticbasis of these related aspects of metacognition—confidence judgements and error monitoring—and identify crucial points of convergence between methods and theories in the two fields. This con- vergence suggests that common principles govern metacognitive judgements of confidence and accuracy; in particular, a shared reliance on post-decisional processing within the systems res- ponsible for the initial decision. However, research in both fields has focused rather narrowly on simple, discrete decisions—reflecting the correspondingly restricted focus of current models of the decision process itself—raising doubts about the degree to which discovered principles will scale up to explain metacognitive evaluation of real-world decisions and actions that are fluid, temporally extended, and embedded in the broader context of evolving behavioural goals. Keywords: decision-making; confidence; metacognition; error monitoring 1. INTRODUCTION andtheroadbecomesslipperyandwet.Theapproach (a) Metacognition in perceptual choice toeachbendstillrequiresadeftanglingofhandlebars, Imagine yourself cycling along a narrow, winding but now you are less certain about whether each countrylaneonasummer’sday.Successfullynegotiat- chosen action is optimized to steer the bicycle in a ing each twist and turn in the road requires the fashion that appropriately meets the gradient of the interpretation of a variety of subtle sensory cues and curve and the camber of the road. Action selection their conversion into appropriate motor commands. proceeds as before, but you are unsure whether the For example, incoming visual signals that reveal the actions chosen are the most appropriate ones. In curvature of the road, or motion parallax signals aris- fact, you might even experience an immediate sense ing from the trees beyond, allows you to gently that you have just made a poor choice. In short, you adjust the handlebars towards the left or right to are less confident about your decisions. steer the bicycle smoothly round each bend. Curiously, despite universal agreement that an Over past decades, psychologists and neuroscien- accompanying sense of confidence is a subjectively tists have devoted substantial effort to understand salient property of almost all our decisions, there is the neural and computational mechanisms by which currently little consensus about how we might incor- actionsareselected on the basis of a streamof incom- porate decision confidence into formal models of ing sensory information [1]. Because this sort of choice behaviour or explore its biological substrates. sensorimotor control often requires an observer to Fundamental questions remain unanswered. For commit to one discrete action from among several example, is the information that gives rise to the possible candidates, this literature has been informed ‘second-order’estimateofconfidenceinachoiceiden- by computational models that offer a formal account tical to that determining the ‘first-order’ choice itself? of how categorical decisions are made. These Why are we generally more sure that we are correct models,describedinmoredetailbelow,haveprovided than that we have made an error, even for difficult an excellent account of the decisions and decision choices? Why do we sometimes appear to ‘change latencies exhibited by observers selecting actions on ourmind’afteramotorprogrammehasbeeninitiated? the basis of ambiguous sensory information [2–5]. Arethesechangesofmindnecessarilyaccompaniedby However, now imagine that as you are cycling, the awareness that the initial choice was incorrect? skydarkensanditbeginstorain—visibilityisreduced, In what follows, we review the literature that has posed these and related questions about decision con- fidence in perceptual choice tasks. Subsequently, we *Authorforcorrespondence([email protected]). highlight links between this work and a literature that has considered how people monitor whether they One contribution of 13 to a Theme Issue ‘Metacognition: computation,neurobiologyandfunction’. have made an error under conditions of uncertainty 1310 Thisjournalisq2012TheRoyalSociety Review. Confidence and error monitoring N. Yeung & C. Summerfield 1311 time This simple model has much to recommend it both as a normative and descriptive account of categorical choice. The DDM has been widely and very suc- q cessfully applied to decision-making in a range of cognitive domains—from low-levelperceptualdecisions a to retrieval of facts from semantic memory, to econo- e response time for trial 1 mic decision-making under uncertainty [6–8]—and c n accounts neatly for several empirical characteristics of e d vi reaction times (RTs) observed for binary choices in e b response time for trial 2 thesetasks.Firstly,thenaturalgeometryoftheaccumu- lation-to-boundprocesspredictstheobservedrightward –q skew in RT distributions. Secondly, varying the bound offers an elegant account of the economy of speed and accuracy that characterizes mental chronometry tasks. Thirdly,byallowingdriftrateandoriginofaccumulation Figure1.Thedrift-diffusionmodel.Accumulatingevidence tovaryacrosstrials, the model accounts for the relative (thedecisionvariable,y-axis)overtime(x-axis)isshownfor RTof correct and error trials under conditions where twoillustrativetrials(markedaandb,greyandblacklines), speed or accuracy are emphasized [8]. Moreover, the one on which the choiceuis made and the other on which model implements a sequential probability ratio pro- thechoice –uismade.Adecisionistriggeredwhenevidence cedure that optimizes speed for a given error rate, and reachesuor –u.Grey line, trial 1;blackline, trial 2. thus takes its place among a family of statistically optimaldescriptionsofthechoiceprocess[2]. orconflict.Finally,weproposesomepotentiallyfruitful avenues for research that draw upon common themes 2. DECISION CONFIDENCE in these two literatures, building on their shared However, as illustrated above, we not only make strengthsandaddressingtheirsharedlimitations. decisions,butalsoconcurrentlyevaluatethelikelihood that those decisions will result in favourable or unfavourable outcomes. How, then, can we incorpor- (b) Formal models of perceptual choice ate decision confidence into the formal framework Even under good viewing conditions, visual infor- offered by the DDM and other quantitative models mation is corrupted by multiple sources of noise and of perceptual choice? In what follows, we review uncertainty, arising both in the external world and in recent empirical and modelling work that has the dynamics of neural processing. One sensible way attemptedtoextendthisformalframeworktoaccount to increase the signal-to-noise ratio of visual infor- formetacognitivejudgements.Muchofthisdebatehas mation is to sample the external world repeatedly hinged on a simple question: can confidence be read and integrate this information over time, making a outdirectlyfromtheDVatthetimeofchoice(decisio- decision only when the information is considered to nal locus model), or do confidence judgements be of sufficient quality [5]. This idea forms the basis depend on new information arriving beyond the of a class of model in which binary choices are decision point (post-decisional locus models)? [9]. described via an accumulation-to-bound mechanism, As an aside, we note that this question only arises with successive samples of information totted up if one is concerned with modelling the dynamics until they reach a criterial level or ‘bound’, upon of the decision process, such as with the DDM or which a commitment to action is made (figure 1). relatedevidence-accumulationmodels.Otherdecision- Different versions of this model make divergent theoretic models, such as signal detection theory assumptions about exactly what quantity is integrated (SDT),havegivendetailedtreatmentofhowtomeasure en route to a decision—i.e. about precisely how the observers’ metacognitive sensitivity, specifically, their ‘decisionvariable’(DV)iscomposed.Tosimplifymat- abilitytodistinguishtheirowncorrectandincorrectjud- ters, in this review, we focus on one variant of this gements [10–12]. However, because SDT is a static model—the ‘drift-diffusion’ model, or DDM—in model of the decision process in which the temporal which the DV corresponds to the relative likelihood dynamicsofevidenceaccumulationareignored,primary of the two choices being correct, given the stimulus (‘type I’) and metacognitive (‘type II’) judgements [3]. In the DDM, the DVon sample v is updated on t mustnecessarilybe basedonthesameevidence(albeit each sample t with an increment composed of two potentially corrupted by different sources of noise quantities: d, a linear drift term that encodes the rate [12]), and thusthese models are by definition decision of evidence accumulation, and cW, Gaussian noise locusmodels. with a mean of zero and a variance of c2: n ¼n þdþcW: ð1:1Þ (a) Decisional locus models t t(cid:2)1 The very earliest investigations of decision confidence Decisions are made when the DV exceeds a fixed revealed that, perhaps unsurprisingly, we are more deviation from zero, u, such that during evidence certain about our perceptual choices when sensory accumulation: inputs are stronger [13], and when we are given longer tosamplesensoryinformation[14].Itthusfol- (cid:2)u.v.u: ð1:2Þ lows that confidence should reflect both the quality of Phil.Trans.R.Soc.B(2012) 1312 N. Yeung & C. Summerfield Review. Confidence and error monitoring the evidence (represented by the drift rate,d) and the onlow-confidencetrialsweresubstantiallyattenuated, quantity of evidence at the choice point (represented falling equidistant between those for the chosen and bytheabsolutevalueofthedecisionbound,juj).How- unchosen targets. The authors interpret these data as ever, it is also clear that neither of these quantities showing that confidence is a simple product of quan- aloneissufficient todescribeanobserver’sconfidence tity and quality of the DV, as proposed by early intheirchoices.Anydiffusion-to-boundmodelassum- accounts of decision confidence, obviating the need ing that confidence is directly indexed by the state of for a separate metacognitive monitoring process [20]. evidence at the time of choice inevitably predicts that Kepecs et al. [19]employed an odourdiscrimination all choiceswill be madewith preciselythe same confi- paradigm in conjunction with single-cell recordings in dence(correspondingtotheevidencelevelrequiredto the rat orbitofrontal cortex (OFC; see also Kepecs & reachthebound).Similarly,anymodelproposingthat Mainen [25]). They report an intriguing finding: that confidence reflects evidence quality implicitly assumes OFC firing rates discriminated between correct judge- that observers have direct access to this quantity— mentsanderrorsafterthechoicebutbeforetheoutcome which,iftheyhad,wouldobviatetheneedforasequen- had been revealed, even when objective difficultyof the tialsamplingapproachinthefirstplace(seePleskac& choiceparadigmwascontrolledfor.Becausevariancein Busemeyer[15]foranexcellentrecentreview). theseresponsescouldnotbeexplainedbyotherfactors, Various accounts dating back over 100 years such as the reinforcement history over past trials, the have therefore proposed that confidence reflects authors suggest that these neurons encode confidence some interaction between the quantity and quality of estimates associated with the current decision. They evidence—for example, that under the DDM, confi- showthattheactivityoftheseneuronscanbeexplained dence scales with the product of d and u [16,17]. byaclassofmodelinwhichtwoevidencetalliesraceto This account has attractive properties—for example, thedecision bound (rather than a single quantityrepre- it predicts that when the observer has the option of senting their difference, as in the DDM). Specifically, terminating evidence accumulation with a choice at neuronal responses reflected the difference between any point, RTs are faster for high-confidence trials thetwototalsatthetimeofthedecision—makingthisa [18]. Moreover, this view enjoys support from two decisional locus model, even thoughthe relevant neural prominent neurophysiological studies detailed in §2b. activity was sustained through the post-response period—andthisactivityinturnpredictedthelikelihood thatthechoicewascorrect[19]. (b) Neural substrates of decision confidence At first glance, these two studies offer convincing Two recent single-cell recording studies, one involv- evidence thatneuronsintheparietalandorbitofrontal ing rats [19] and the other monkeys [20], claim to cortices encode the subjective confidence associated have identified neurons encoding subjective decision with achoice: agraded quantitythatcan be estimated confidence. Kiani et al. [20] recorded from neurons directly fromthe evidence on which the original, first- inthelateralintraparietalcortex(LIP)ofmacaquesdis- orderchoice was based. In both cases, the authors are criminating the direction of motion in ambiguous assiduous in attempting to rule out alternative expla- moving-dot arrays [21] with a saccade to one of two nations of their findings, such as attention or learned targets placed either side of the motion stimulus. This reward contingencies. However, to return momenta- experimental approach has previously been used to rily to our discussion of models of categorical choice, demonstrate that LIP neurons whose receptive field two well-replicated behavioural phenomena cast overlaps with the chosen target display a characteristic doubt on any model in which subjective confidence acceleration of spiking activity while the monkey views directly reflects the evidence accumulated up to the the moving-dot stimulus, whereas those overlapping choice point. First, it has consistently been shown with the non-chosen target show a relative dampening thatconfidenceincorrectchoicesisstrongerthancon- of activity [22,23]. This build-up scales with signal fidence in incorrect choices, even when choice strength, and terminates when an action (saccadic eye difficulty is controlled. As previously pointed out movement) is selected, prompting the view that LIP [15], this finding cannot be explained by decisional firing rates encode the level of evidence available for a locus models in which confidence is a mere function choice—inotherwords,their firing rates form aneural of diffusion model parameters, because these para- representationoftheDVproposedbytheDDM[1]. meters are invariant across correct and incorrect To investigate decision confidence, Kiani et al. trials (but see Galvin et al. [10] for relevant analysis added a new twist to this protocol. Typically, the usingSDT). Second, andcrucially,subjects occasion- monkey is rewarded for correct but not incorrect ally change their mind between the first-order choice choices, but here, on a fraction of trials, the monkey and the second-order estimation of decision confi- was offered a ‘sure bet’ option such that a certain dence. At least on these trials, the decisions must be but lower-valued reward could be obtained via a sac- influenced by processing that occurs after the first- cade to a third response option. In ethology, it is order choice point. In §2c, we consider models that commonly assumed that an animal choosing the safe propose such a mechanism. option when evidence is scarce must have access to metacognitive information about the likelihood that it will make an error on the main task [24]. The (c) Post-decisional locus models and changes authors found that not only did their monkeys use of mind thisoptionjudiciously—respondingwhenthestimulus Resulaj et al. [26] report a behavioural experiment in wasweakorambiguous—butalsothatLIPfiring rates which human subjects indicated the direction of a Phil.Trans.R.Soc.B(2012) Review. Confidence and error monitoring N. Yeung & C. Summerfield 1313 random dot motion stimulus by moving a handle to a predicting decision confidence might not be confined leftwards or rightwards target some 20cm away. This tothestimulationperiodalone.However,onepotential design allowed the researchers to isolate trials on caveattothisviewisthattheactivityofLIPneuronsis which subjects began to move towards one target but known to drop off sharply once the eye movement is then changed their mind and veered-off towards the made [22,23]. Thus, it remains to be shown whether other. Careful behavioural analyses demonstrated a variation in the post-decisional LIP signal could con- number of interesting phenomena. Firstly, changes of tribute to a later confidence judgement, or whether a mind were not symmetric: subjects switched more separaterepresentationoftheevolvingDVissupporting often from an incorrect to a correct choice. Secondly, theobservedchangesofmind.Thiscaveataside,these change-of-mind trials tended to occur when, owing emerging findings suggest that post-decisional proces- to stochasticity in the stimulus display, motion sing plays a crucial role in metacognitive judgements, energy began by favouring the initial choice, but sub- which can lead to changes of mind or support ratings sequently came to favour the alternative, switched-to ofconfidenceinaninitialdecision. option.Becausethemotionstimulusoffsetoncemove- ment began, subjects must have capitalized on the balance of information in the immediate pre-decision 3. ERROR MONITORING period when deciding to change their mind. Notably, People are often aware of their own mistakes; for although this motion information was available prior example, in choice RT tasks when time pressure is to the decision, the switch occurred only once move- applied to induce errors in simple judgements. Error ment initiation began, suggesting that evidence monitoringisthemetacognitiveprocessbywhichweare accumulation continued beyond the point at which abletodetectandsignalourerrorsassoonasaresponse the initial choice was made. has been made. This process plays a crucial role in To account for these and other data, a number of adaptive human behaviour, allowing our actions to be researchers have proposed models in which, in con- shaped by their outcomes both in the short term, for trast to the classical DDM, evidence accumulation example,byrespondingmorecautiouslytoavoidfurther continues even beyond the choice point, with this errors,andinthelongerterm,throughgraduallearning extra variability in the DV also contributing to esti- ofappropriatestimulus–responsecontingencies. mates of subjective confidence when probed at a Whereas the studies described above have typically later time point [9,15]. Resulaj et al. propose that asked subjects to report their confidence that they their data can be explained by just this type of made the correct choice, studies examining error model, with changes of mind occurring when latent monitoring have tended to ask subjects the converse information in the processing pipeline drives the DV question—i.e. to report the likelihood that they made across a second, ‘change-of-mind’ bound. A related an error. Although these judgements seem like two account, the two-stage dynamic signal detection sides of the same coin, methods and assumptions in (2DSD) model [15], likewise proposes that the diffu- the two literatures have often been quite different. For sion process continues beyond initial choice, with example,decisionconfidencehastypicallybeenstudied decision confidence reflecting the absolute value of usingtasksthatremainchallengingevenwhenextended the DVat the post-decision point at which a second- processing times are permitted, such as challenging orderdecisionisrequired.Thus,incontrasttodiscrete psychophysical discriminations. Under these circum- changesofmindthatareexpressedasovertcorrections stances, subjects are sometimes sure they responded ofaninitialresponse,the2DSDmodelallowsforcon- correctlyandsometimesunsurewhether theyare right tinuously varying levels of confidence that provide an orwrong,butrarelycertainthatthattheymadeamis- explicit judgement about an earlier response. take [15]. In contrast, research on error monitoring Behaviourally,thesepost-decisionprocessingmodels has mostly been studied using simple but time- are able to account for a broad range of findings con- pressured tasks in which subjects are usually aware of cerning decision confidence. Firstly, they correctly their errors and very rarely unsure whether their predict that observers will change their mind more decision was right or wrong. Framed in terms of the often from incorrect to correct responses than vice DDM,thedistinctionconcernswhethererrorsandsen- versa, because beyond the bound the DV on error sitivitytoprocessingnoisearisebecauseoflowdriftrate, trials will tend to regress towards the mean, whereas d,inthecaseofperceptualambiguity,oradoptionofa after correct responses it will continue to grow, driven lowthreshold,u,toengenderfastresponding.Thereis bythetrue underlying drift rate.This observation also nonetheless obvious similarity between metacognitive naturally explains another conundrum associated with judgements of confidence and error likelihood, and it decision confidence: that second-order confidence is isthereforeunsurprisingthat models oferror monitor- generally higher for correct trials than incorrect trials. ing turn out to complement those more recently Indeed, reconsidering for a moment the data from developedinperceptualdecision-makingresearch. Kiani et al., we note that these authors report that neuralactivityinthedelaybetweenstimulusoffsetand (a)Post-decisionprocessinginerror monitoring movement execution exerted a separate, independent Rabbittandco-workers’pioneeringworkbeginning in influence over the decision to choose the ‘sure bet’ the 1960s established that error monitoring relies on option [20]. Thus, although the authors argue that a post-decision processing. Their experiments showed mechanistic description of decision confidence does that people can very reliably detect and correct their notrequireustoinvokeadistinctmetacognitiveprocess own errors without requiring explicit feedback [27], separate from evidence accumulation, the evidence butthatthisabilityisimpairedwhenstimulusduration Phil.Trans.R.Soc.B(2012) 1314 N. Yeung & C. Summerfield Review. Confidence and error monitoring time uncertainty—or conflict—in the decision process after an initial response [34], or as inconsistency between the outcomesofparalleldecisionprocessesat different (a) processing stages (e.g. perceptual categorization and q response selection) [35]. While varying in their details and precise predictions [32,34], common to all pro- ce (b) qcom posals is the claim that metacognitive accuracy en judgementsdependonpost-decisionprocessing. vid (c) In effect, applying the framework provided by the e DDM and its variants, one is obliged to assume that –q error detection reduces to error correction: that errors are detected whenevercorrective activity reaches a cri- (d) teria level—the change-of-mind bound. However, this conclusion is difficult to reconcile with evidence that errorcorrectionanddetectionareatleastpartlydissoci- Figure2.TheoriesoferrordetectionwithintheDDMframe- able, a finding that was again presciently reported in work. The drift-diffusion process is illustrated schematically Rabbitt’sseminalwork.Rabbitt’sstudiesdemonstrated fortwotrials,oneinwhichdecisionuisthecorrectresponse that error corrections are a relatively automatic and andonetrialinwhichthisdecisionisincorrect.Bothdecisions unreflective consequence of post-decision processing. occur at the same time point (a). Following the correct response (grey line), post-decision processing continues to Thus, they can be extremely fast, occurring within accumulate in favour of the decision just made. Following 10–20msoftheinitialerror[36],andmaybeproduced errors (black line), the drift rate regresses to its true mean, evenwhensubjectsareinstructedtoavoiddoingso[37]. causing the DV to re-cross the decision bound (b), sub- Incontrast,explicitdetectionandsignallingoferrorsis sequently cross a change-of-mind bound (c), and finally muchslower,voluntary,more pronetointerference by crosstheoriginallycorrectdecisionbound, –u(d).Thegrey distractingtasks,andmoresensitivetocognitivedecline shaded area indicates a period of uncertainty, or conflict, in normal ageing [29]. Indeed, people sometimes betweenthere-crossingoftheubound(b)andlatercrossing remainunawareoferrorsthattheyneverthelesscorrect ofthe –ubound(d). [38].Collectively,thesedifferencessuggestthatexplicit error detection cannot be a mere consequence of is reduced [28], suggesting its dependence on contin- post-decision correction: further processing or evalu- ued processing of the stimulus after an initial error ation must intervene between initial correction and (which is curtailed when stimuli are very briefly pre- later explicit awareness that an error has occurred. sented). Error monitoring is also impaired when Consistent with this analysis, recent investigations subsequent stimuli appear soon after the initial have identified dissociable neural correlates of error response [29], and responses to those later stimuli correctionanddetection. are themselves postponed following errors [30], con- sistent with the notion that this monitoring involves (b) Neural substrates of error monitoring the same mechanisms as the initial decision. Researchinterestinerrormonitoringincreasedsubstan- Summarizing these findings, Rabbitt likened evi- tially following the discovery of scalp EEG potentials dence accumulation in decision-making to votes in a that reliably occur time-locked to decision errors. committee,inwhichincorrectdecisionsaresometimes Most studies have used the Eriksen flanker task, in made on the basis of incomplete information, but ‘as which subjects perform a speeded categorization of a subsequent votes come in, a more accurate consensus centraltarget(e.g.HorS),whileignoringflankingdis- will accumulate and the earlier mistake will become tractorsthataresometimescompatible(e.g.HHH)and apparent’[28].Thus,errorsarecharacterizedbybipha- sometimes incompatible (e.g. SHS) with that target. sicevidenceaccumulation,withinitialaccumulationin Withmodestspeed pressure,error rates onincompati- favour of the incorrect response followed by later drift ble trials can exceed 20 per cent. Following such towards the correct decision (as the trial-average drift errors,anegativeevent-relatedpotentialoverfrontocen- rateregressestothetruemean).Bycontrast,continued tral sites is observed within 100ms of the incorrect evaluation following correct responses tends simply response, followed by a later positive wave peaking to reinforce the original decision. This model is an 200–400ms later over parietal cortex [31]. Labelled obvious precursor to more recent accounts of decision theerror-relatednegativity(ERN/Ne)anderrorpositiv- confidence[15]andchangesofmind[26]. ity(Pe),respectively,theseEEGcomponentshavebeen All subsequent models of error detection have widelystudiedtoprovideinsightintoerror monitoring adoptedRabbitt’sbroadframework,withdebatefocus- inhealthyandclinicalpopulations. ing instead on the precise mechanism by which post- Converging evidence has identified anterior cingu- decision processing leads to error detection. Figure 2 late cortex (ACC) as the source of the ERN. For illustrates some key model variants. Within a standard example, in simultaneous EEG–fMRI recordings, DDMframework,errorscouldbedetectedassuccessive single-trial ERN amplitude correlates reliably only crossingsofdecisionboundariesforthetwocompeting with activity in a focused ACC source [39]. The responses [31,32] or as ‘double crossings’ of a single sourceofthePeislesswellcharacterized,butevidence decision bound [33]—both close relatives of Resulaj that it is a variant of the well-studied P3 component et al.’s notion of a change-of-mind bound. Errors can [40] would imply widely distributed neural generators also be detected in terms of the occurrence of inparietalandprefrontalcortex[41].TheP3hasbeen Phil.Trans.R.Soc.B(2012) Review. Confidence and error monitoring N. Yeung & C. Summerfield 1315 suggested to reflect the distributed action of norepi- focused on the finding that subjects usually respond nephrine released by the locus coeruleus brainstem more slowly on trials immediately following errors nucleus in response to motivationally salient events [27].Althoughthiseffectatleastpartlyreflectsthedis- during decision-making [42]. tracting occurrence of a rare event [50], as errors The functional significance of the ERN and Pe is a typically are, a consensus view holds that post-error matter of ongoing debate. Competing theories of the slowing reflects strategic adaptation to prevent further ERN propose a role in error detection, reinforcement errors [51]. EEG studies have subsequently shown learning and conflict monitoring; while theories of that the degree of observed slowing scales with the the Pe include conscious awareness of errors, affective magnitudeoferror-relatedERN/Peactivity[46].Com- responses and behavioural adjustments to avoid putational models implementing error-related control further errors [43]. These debates notwithstanding, over distance-to-bound, juj, account for detailed prop- it is now clear that the ERN and Pe dissociably map erties of empirically observed post-error slowing. In onto processes related to error correction and error one class of model, detection of response uncertainty detection, respectively. Thus, ERN onset coincides (conflict) immediately following error commission with the onset of error-correcting activity as revealed leads to an increase in the bound—and, hence, more through EMG recordings [44], typically around the cautious responding—on subsequent trials [52]. time of error commission [31], and its amplitude Recent extensions of this idea suggest that conflict varies with both the speed [45] and probability [46] detection may also be used to adjust the bound oferrorcorrection. In contrast, Pe amplitude isinsen- dynamicallyevenasadecisionisbeingmade[51,53]. sitive to the strength of correcting activity when error Error signals not only support subtle adjustments detection rates are controlled [47]. Conversely, that optimize online decision-making, they also play a although both the ERN and Pe are found to covary key role in longer term adjustments during learning. with subjective ratings of response accuracy [46,48], Holroyd & Coles [35], for example, suggest that the correlations involving the ERN disappear when the ERN reflects reinforcement learning of action values. two components are carefully dissociated. In antisac- They showed that the ERN migrates in time as new cade tasks in which subjects correct all of their errors stimulus–responsemappingsarelearned,frominitially butdetectonlyhalfofthem,ERNamplitudeisequiv- being triggered by environmental feedback to later alent for aware and unaware errors, whereas the Pe is being driven by internal representations of the learned robustly observed only when subjects detect their mappings, a pattern that mimics the migration of errors [38]. Taken together, these findings suggest dopaminergic reward prediction error signals from that whereas the ERN directly indexes automatic unconditionedrewardstopredictivestimuliduringcon- post-decision processes leading to rapid error correc- ditioning [54]. Meanwhile, fMRI activity in ACC and tion, the later Pe is selectively associated with explicit neighbouringcortexatthetimeofanincorrectresponse detection and signalling of errors. These results thus have been shown to predict response accuracyon later provide converging evidence for the view that error presentationsoftherelevantstimulus[55,56]. correction and detection reflect distinct processes. Moststudiesofpost-erroradjustmentshavefocused Arecentstudyhasshedlightonspecificallyhowthe on the ERN and ACC, reflecting wide interest in the Perelatestoerrordetection[49].Subjectsperformeda role of ACC in reinforcement learning [35,57], difficultbrightnessdiscriminationunderspeedpressure rather than on the later Pe component. However, the toinduceamixtureoferrorsresultingfromperceptual ERN and Pe typically covary across conditions and, ambiguity and decision urgency. After each response, when the two components are dissociated, post-error they made a binary correct/error judgement, with adjustments are only observed following detected monetary incentives varied across blocks to encourage errors on which a Pe component is present [38], either liberal or conservative error signalling. Signal suggesting that the latter may be a more direct corre- detectionanalysisindicatedthatsubjects’accuracyjud- late of the learning mechanisms by which future gements could be well fit by assuming that these behaviour is adapted following an error. judgementsreflectedacontinuumofconfidence(from sure correct to sure incorrect), with subjects applying a criterion—a metacognitive u—that varied according 4. INTEGRATIVE MODELS OF DECISION to the incentive regime. Critically, error signallingper- CONFIDENCE AND ERROR MONITORING formancewascloselyrelatedtobetween-conditionand The discussion above highlights convergent evolution trial-by-trial variation in Pe amplitude (but not the in work on decision confidence and error monitoring, ERN). That is, Pe amplitude appeared to provide a suggesting that common principles govern metacogni- direct neural index of continuously varying decision tive judgements regardless of whether errors arise confidenceonwhichsubjectsbasedtheirmetacognitive because of intrinsic task difficulty (low drift rate, d) judgements,withcategoricalsignallingoferrorsoccur- or because of externally imposed decision urgency ring when confidence that the response was wrong (low threshold, u). In what follows, we consider the exceededsomecriteriallevel. implications of this convergence for future research, both positive (in terms of mutually informative les- (c) Impact of error monitoring on behaviour sons) and negative (in terms of shared limitations). Theresearchdescribedabovedocumentsthemechan- istic and neural basis of error monitoring. A parallel (a) Converging themes lineofresearchhasconsideredtheimpactoferrormoni- Theories of decision confidence and error monitoring toring on future behaviour. Much of this work has each emphasize the importance of post-decision Phil.Trans.R.Soc.B(2012) 1316 N. Yeung & C. Summerfield Review. Confidence and error monitoring processing, and likewise point to a key dissociation behaviourasaseriesofdiscretedecisions,eachsubject between, on the one hand, continued processing of toindependentmetacognitivescrutiny,isausefulcon- stimulus information within the decision-making venience when developing experimental tests and system (leading to changes of mind and error correc- formal models. However, it remains unclear whether tions) and, on the other, the formulation and the findings will scale up to explain real-world expression of explicit judgements of confidence and decisions and actions that are fluid, temporally accuracy. The literatures have developed these ideas extended and embedded in the broader context of in complementary directions, such that each literature evolving behavioural goals. offers valuable lessons for the other. First, whereas Letusrevisitforamomentouropeningexampleof confidence is typically characterized as varying along cycling along on a winding country lane. Obviously, a continuum, and formalized as such in accounts we do make occasional discrete categorical such as the 2DSD model [15], error detection is decisions—for example, when choosing between two often characterized as an all-or-none process [32,33]. available paths at a fork in the road. However, most Thus, according to many current theories of error of our decisions and actions unfold gradually, shaped monitoring, binary yes/no error judgements are an by our interactions with the environment and an intrinsic feature of the monitoring system rather than ever-changing stream of incoming sensory infor- a reflection of the arbitrary metacognitive decision mation, as exemplified by the continuous, subtle thatsubjectsareaskedtomake.Assuch,thesetheories adjustmentsofhandlebarsandbrakesneededtomain- cannotexplainhowsubjectsareabletoexpressgraded tain direction and balance when cycling. Current confidence in their accuracy judgements [48,49]. theories of metacognition require a clear division in Conversely, the preceding review of work on error time between cognitive decision-making and meta- monitoring identifies important limitations in current cognitive evaluation, with the latter beginning when theories of decision confidence. This review suggests the former ends (‘post-decision processing’). However, that the current debate between decisional locus and for continuous and extended actions, there is no post-decisional models of decision confidence will definable time-point at which a specific decision is very likely be resolved in favour of the latter, but that finalized and metacognitive evaluation begins. Most even these post-decisional theories will need modifi- current theories are therefore ill-suited to describing cation to accommodate evidence that metacognitive thetemporallyextendedactions,comprisinghundreds awareness (cf. error detection and the Pe) cannot be or thousands of micro-decisions, that characterize reduced simply to post-decision processing (cf. error everyday behaviour. correction and the ERN)—the two are at least partly Indeed, decisions that initially appear discrete and dissociable. Perhaps more informative still will be categorical may turn out, on closer inspection, to be exploration of the role of confidence judgements in graded and transitory. For example, in the Eriksen guiding future actions: whereas research on decision flanker task, overt responses occur tightly time- confidencehaslargelyfocusedonhowconfidenceesti- locked to the point at which lateralized activity in mates are derived, a major focus of error-monitoring motor cortex exceeds a threshold value [58], just as research has been on how this kind of metacognitive saccadiceyemovementsappeartobetriggeredaccord- information might be used to modify behaviour both ing to a threshold firing rate for LIP neurons [1], in the short- [27,52,53] and long-term [35,55,56]. suggestive of a fixed decision point after which an Borrowing these insights, we might predict that para- action is initiated. However, fine-grained analyses metric estimates of confidence could support finer- reveal graded and continuous information flow at grained control of behaviour than can be achieved every stage, even downstream of motor cortex. Thus, through binary categorization of responses as correct within a single trial, motor cortex activity may latera- or incorrect—for example, by allowing parametric lize first towards one response then towards the variation in post-error slowing or by providing a other; small EMG twitches in one finger may be fol- scalar predictionerrorsignaltosupportreinforcement lowed by full movements of another; and overt learning. Thus, confidence estimates could provide actions themselves may vary in force in a graded useful information in optimizing the rate of learning: manner, for example, with incorrect actions executed onemightpredictthatpeoplewillpaygreaterattention less forcefully than correct ones [44,58,59]. Mean- to environmental feedback following responses in while, categorical or economic judgements about which they have less confidence. visual information are often preceded by exploratory eye movements, which may themselves constitute interim decisions en route to the eventual choice [60]. (b) Shared limitations In such systems, there is no single, final decision As well as sharing complementary strengths, the the- point that could mark the beginning of metacognitive ories also share common weaknesses. In particular, evaluation. Detailed analyses of neural markers of like the models of decision-making on which they are metacognition point towards a similar conclusion: based, current theories of confidence and error moni- error-relatedEEGactivityisnotonlyobservedfollow- toring have focused almost exclusively on decisions ing overt incorrect actions, but also following ‘partial’ that are discrete and punctate in time: a decision is errors in which the incorrect muscles twitch, but the made when a boundary is reached [3]; errors are incorrect action is not produced [61] and, crucially, detected when a second change-of-mind bound is in a graded fashion as a function of the level of sub- crossed [26]; and confidence is estimated at the time threshold cortical activity favouring an incorrect of a later metacognitive probe [15]. Characterizing response [34]. Metacognition appears to be graded Phil.Trans.R.Soc.B(2012) Review. Confidence and error monitoring N. Yeung & C. Summerfield 1317 and continuous in just the same manner as the revealed that human observers are exquisitely sensi- underlying decision process. tivity to evidence reliability when sampling from Human decision-making also has a continuous multiple sources of information [67]. quality whenviewedover longer timescales, with indi- Yet, most currently popular models of perceptual vidual decisions chained into sequences that serve decisions offer no way of expressing the trust or dis- longer-term behavioural goals. Thus, actions that trust associated with the evidence accumulated; all reflect definitive choices at a lower goal level (e.g. sourcesofevidencearecombinedinthecommoncur- movements of the arm, shifts of posture) may consti- rency of the DV, which then gives the strength of tute reversible, interim choices at a higher goal level evidence as a simple scalar value (the vertical location (e.g.turningleftataforkintheroad),andevenacom- ofthediffusingparticleintheDDM,orthemagnitude mitment to an individual right or left turn might be of the evidence variable in SDT). Once factored into just a partial commitment to a yet higher level goal the DV in this way, the reliability of evidence has no (e.g. reaching a specified destination) [62]. This further impact on the decision-making process. This form of hierarchical structure is built into many feature contrasts with some accounts of economic recent theories of the computational and neural basis choices, in which sensitivity to risk (i.e. outcome var- of action selection [63], in which common principles iance) as well as value (i.e. outcome mean) has been of selection and control are held to operate at each documentedatthebehaviouralandneurophysiological level of hierarchical abstraction [64]. Recent findings levels [68,69]. indicatethatmetacognitiveprocessesaresimilarlysen- Mathematically, the reliability of perceptual evi- sitive to this hierarchical structure [65]. For example, dence is orthogonal to its strength, because the mean errors that are equally discrepant in terms of low- and the variance represent different moments of a level actions are treated very differently according to probability distribution. Bayesian models that exploit their impact on global task performance [66]. Little this point—representing evidence as a probability dis- is currently known about the mechanisms by which tribution with a given mean (evidence strength) and metacognitive judgements might be embedded in variance (evidence reliability)—have been used to cal- ongoing higher level behaviour in this way. culate ideal estimates of expected value in economic Thus, a crucial limitation of current metacognitive choice tasks [70]. Applied to perceptual decision theoriesisthattheydonotreflectthewaythatourcon- tasks, the notion would be that the diffusing parti- fidence in our actions contributes to fluid, structured cle of the DDM is more accurately conceived of as a sensorimotor behaviour. Rather, they consider that probability distribution that evolves across samples errors are processed in an all-or-nothing fashion, for (figure 3), with the central tendency of that distri- example,whenapost-decisionprocesscrossesameta- bution analogous to the vertical location of the cognitivedecisionbound.Inwhatfollows,weconsider particle. Crucially, the variance of this distribution anotherwayofthinkingaboutdecisionconfidencethat provides additional information—specifically, a rep- is not subject to these limitations. resentation of evidence reliability in terms of the precision of the mean—that is not made explicit in the standard DDM (by precision, we mean the inverse 5. FUTURE DIRECTIONS of the standard deviation). Addressing the challenges outlined above will require Recently, it has been shown that neural network the development of new models that consider not modelsinwhichLIPneuronsencodethefullposterior just discrete, binary choices, but also the kinds of probability distribution associated with a stimulus can extended, goal-oriented actions that characterize capture behaviour and neural dynamics occurring everydayhuman behaviour.We conclude by consider- duringpsychophysicaldiscriminationtasksinprimates ing one promising extension of current decision [71]. This result suggests that precision may be models, which addresses the issue of the reliability of encoded in the variance of firing rates across a neural evidence in a manner that opens intriguing new ave- population, thus signalling evidence reliability in nues for understanding how metacognitive evaluation much the sameway thatmean firing rateof the popu- might occur for continuous, extended actions. lation encodes evidence strength [72]. Relatedly, Ratcliff & Starns [73] have recently proposed a (a) Evidence reliability model of decision confidence in which the decision- Formal accounts of categorical decisions, such as the relevant evidence on a single trial is not a scalar esti- DDM, are often illustrated by analogy to court of mate, but a distribution of possible values, allowing law, in which the jury weighs up evidence favouring parametric estimates of confidence in the choice to the innocence or guilt of the defendant [1]. However, be calculated asthe integral of this distribution falling this analogy also highlights an inconsistency between within graded confidence bounds. Together, these current decision models and choices made in the real findings open promising new avenues for research world: that in the latter, we usually consider the into the neural basis of categorical choice. extent to which we trust the evidence relevant for a decision. For example, in a law court, evidence from a trustworthy source (for example, official telephone (b) Confidence revisited records)mightweighmoreheavilyinajury’sdelibera- We propose that evidence strength and reliability are tions than evidence from an unreliable source (for encoded in parallel during evidence accumulation, example, a witness with a vested interest). Corre- and that this framework provides an intriguing new spondingly, careful signal detection analyses have way of thinking about decision confidence—as the Phil.Trans.R.Soc.B(2012) 1318 N. Yeung & C. Summerfield Review. Confidence and error monitoring (a)ata) time t time t +1 Swehciochndelyv,idthenecme oqduealliitsyavbalerietso deveesncribweithsiitnuaatiosninsgilne d H | trial [71], something that standard models cannot p ( achieve. In fact, by keeping trackof the likely variabil- ity of information in the external world, Bayesian perceptual hypotheses perceptual hypotheses accounts can optimally distinguish true state changes (b) in the generative information giving rise to the senses ses low variability high variability p (H | data) fromnoise[74].Precisionestimatesarethusparticula- e h ot high rly useful in situations where the causes of perceptual p y evidence may change unpredictably over time, and as al h low such may provide a better account of the sort of fluid, u pt ongoing sensorimotor integration that characterizes e c everydayactivitiessuchasridingabicycle. er p The hypothesis also leads to clear and testable time time predictions about the sensitivity of human decision- Figure3.Schematicofamodelinwhichboththemeanand making to evidence reliability. Firstly, it predicts that variance of information in an array are estimated in a serial variabilityofevidencewithinatrialshouldbothreduce sampling framework. (a) The left panel shows the posterior accuracy and lengthen RT, because during sequential probability distribution p (H j data) over a continuous sampling, the precision of the mean increases more space of possible perceptual hypotheses (e.g. these dots are slowlywhenthesamplesaredrawnfromamorevariable moving to the right with 30% coherence; this signal is 40% distribution. This prediction has recently been con- visible; etc.) at a given time, t. This distribution reflectsthe firmed in an experiment in which observers made evidence sampled from the stimulus thus far, i.e. between discrimination judgements on the average feature (e.g. onset and time t (grey dots). The new sample received at colour) of an array of multiple elements presented timetisshowninred.Rightpanel:attimetþ1,thisdistri- simultaneously on the screen [75]. Critically, this bution (light grey) is updated in the light of the newly sampled information, giving rise to anewprobabilitydistri- multi-elementaveragingtaskallowedtheexperimenters bution.Inthismodel,confidenceisreflectedintheprecision tomanipulatethemeanandthevarianceoftherelevant oftheposteriordistribution,i.e.thereciprocalofitsstandard featureinanorthogonalfashion.Theresultsshowedthat deviation.(b)Theevolvingposteriorprobabilitydistribution observers were slower to discriminate more variable over perceptualhypotheses(y-axis)foreachsuccessivetime arrays,aresultthatispredictedbytheprecisionaccount point (x-axis; blue–red colourmap; red values indicate but not by standard accumulation models such as the higher probabilities). The posterior distribution is updated DDM. Moreover, observers in this study tended to followingthearrivalofsuccessivesampleswithlowvariance downweight outlying or otherwise untrustworthy evi- (left panel) or high variance (right panel). Precision of the dence,muchlikeastatisticianmightexcludeanoutlier probabilistic representation of evidence strength increases from a sample of data. Although this study did not morerapidlyfor thelow-variabilitysamples. assess subjects’ second-order confidence in their decision, the precision accountmakesthe clear predic- precisionofevidenceaccumulatedduringasingletrial. tion that in the multi-element task, both subjective Thisrepresentationofconfidencediffersintwocritical confidence and the rate of occurrence of changes of respects from the models described above: it is based mindwilldependonarrayvariabilitytoagreaterextent on an intrinsic feature of the decision process—var- thanonitsmean,apredictionripeforfuturetesting. iance in the evidence encountered—and it is Finally, this conception of decision confidence available continuously and instantaneously. In con- makes direct contact with broader theories of the trast, existing models suggest that metacognitive role of metacognitive evaluation in behavioural con- evaluations are derived indirectly by comparing rep- trol. In particular, because precision is closely related resentations of the DV from two (or more) discrete to the concept of decision conflict [52]—greater var- time points. Precision thus provides a more suitable iance in evidence should result in greater conflict basisfor metacognitiveevaluationinthekindsoftem- between competing response options—the theory can porally extended tasks discussed above, in which no inherit ideas from research on decision conflict about discrete decision point divides cognitive decisions how precision estimates might be used to guide both from metacognitive evaluation. current performance (e.g. through dynamic modu- This hypothesis, though speculative at this point, lation of decision bounds) [51,53] and future has several attractive features. Firstly, it is consistent behaviour (e.g. through modulation of learning rate with the evidence described above on post-decision in relation to environmental signals of success or fail- processing because instantaneous precision will tend ure) [70]. As such, the precision model not only to be highly correlated with later variability in the provides a formally specified account of decision con- DV(thebasisofmostcurrentmodelsofmetacognitive fidence,butalsoleadstoimmediatesuggestionsabout judgements): variable evidence will give rise to fre- the use of confidence judgements in the optimization quent errors, and these errors will be accompanied of behaviour. by low estimates of precision at the time of the initial response and a high probability of later changes of mind or error correction. Thus, the precision account 6. CONCLUSION agrees that existing theories fit the data well, but Formal models such as the DDM have proved extre- suggests that they may do so for the wrong reasons. mely valuable in understanding human and animal Phil.Trans.R.Soc.B(2012) Review. Confidence and error monitoring N. Yeung & C. Summerfield 1319 decision-making,bysituatingexperimentalobservations judgments. Psychon. Bull. Rev. 10, 177–183. (doi:10. ofbehaviourandneuralactivitywithinapreciselyspeci- 3758/BF03196482) fied and normatively motivated framework. Direct 10 Galvin, S. J., Podd, J. V., Drga, V. & Whitmore, J. extensions of these models have proved equally useful 2003 Type 2 tasks in the theory of signal detectability: discrimination between correct and incorrect decisions. in probing the metacognitive processes by which we Psychon. Bull. Rev. 10, 843–876. (doi:10.3758/BF0 evaluate and express our degree of confidence in our 3196546) decisions. In particular, significant convergence in 11 Evans,S.&Azzopardi,P.2007Evaluationofa‘bias-free’ methods and theories of decision confidence and error measure of awareness. Spat. Vis. 20, 61–77. (doi:10. monitoring suggest that common principles may 1163/156856807779369742) governdifferenttypesofmetacognitivejudgements. 12 Maniscalco,B.&Lau,H.2012Asignaldetectiontheor- There is nonetheless important scope for current etic approach for estimating metacognitive sensitivity modelstoconsiderdecision-makingandmetacognitive from confidence ratings. Conscious. Cogn. (doi:10.1016/ evaluationinsituationsthatencompassnotonlysimple, j.concog.2011.09.021) punctate choices, but also the kinds of extended, 13 Peirce, C. S. & Jastrow, J. 1884 On small differences of sensation. Mem.NatlAcad. Sci.3,73–83. goal-directed decisions and actions that typify human 14 Vickers,D.,Burt,J.,Smith,P.&Brown,M.1985Exper- behaviour outside the experimental laboratory. We imental paradigms emphasizing state or process have proposed one such extension: the hypothesisthat limitations: 1. Effects on speed accuracy tradeoffs. Acta peoplearesensitivenotonlytothestrengthofevidence Psychol.59,129–161. theyencounterastheymakeadecision,butalsotothe 15 Pleskac, T. J. & Busemeyer, J. R. 2010 Two-stage reliability of that evidence. This simple proposal has dynamic signal detection: a theory of choice, decision far-reaching implications: it immediately suggests a time, and confidence. Psychol. Rev. 117, 864–901. novel source of information—evidence precision—that (doi:10.1037/a0019737) could guide metacognitive evaluation. Future develop- 16 Link, S. W. 2003 Confidence and random walk theory. ments in theories of human decision-making promise In Modeling choice, decision time, and confidence 893 (ed. to have similarly profound implications for our under- B. B. E. Borg). Stockholm, Sweden: International Society for Psychophysics. standing of the way in which people evaluate their 17 Heath, R. A. 1984 Random-walk and accumulator decisions in the service of adapting and optimizing models of psychophysical discrimination—a critical those decisions in the face of an uncertain, complex evaluation.Perception13,57–65.(doi:10.1068/p130057) andever-changingenvironment. 18 Vickers,D.&Packer,J.1982Effectsofalternatingsetfor speedoraccuracyonresponse-time,accuracyandconfi- N.Y. was supported by a grant from the U.S. National dence in a unidimensional discrimination task. Acta Institutes of Health (P50-MH62196). C.S. was supported by a grant from the Wellcome Trust (WT092646AIA). Psychol. 50, 179–197. (doi:10.1016/0001-6918(82) Thoughtful commentaries from Vincent de Gardelle and 90006-3) twoanonymous reviewersaregratefullyacknowledged. 19 Kepecs,A.,Uchida,N.,Zariwala,H.A.&Mainen,Z.F. 2008 Neural correlates, computation and behavioural impact of decision confidence. Nature 455, 227–231. (doi:10.1038/nature07200) REFERENCES 20 Kiani, R. & Shadlen, M. N. 2009 Representation of 1 Gold, J. I. & Shadlen, M. N. 2007 The neural basis of confidence associated with a decision by neurons in the decision making. Annu. Rev. Neurosci. 30, 535–574. parietal cortex. Science 324, 759–764. (doi:10.1126/ (doi:10.1146/annurev.neuro.29.051605.113038) science.1169405) 2 Bogacz,R.,Brown,E.,Moehlis,J.,Holmes,P.&Cohen, 21 Britten, K. H., Shadlen, M. N., Newsome, W. T. & J. D. 2006 The physics of optimal decision making: a Movshon, J. A. 1992 The analysis of visual motion: a formalanalysisofmodelsofperformanceintwo-alterna- comparison of neuronal and psychophysical perform- tive forced-choice tasks. Psychol. Rev. 113, 700–765. ance. J. Neurosci.12,4745–4765. (doi:10.1037/0033-295X.113.4.700) 22 Roitman,J.D.&Shadlen,M.N.2002Responseofneur- 3 Ratcliff, R. & McKoon, G. 2008 The diffusion decision ons in the lateral intraparietal area during a combined model: theory and data for two-choice decision tasks. visual discrimination reaction time task. J. Neurosci. 22, Neural Comput. 20, 873–922. (doi:10.1162/neco.2008. 9475–9489. 12-06-420) 23 Shadlen, M. N. & Newsome, W. T. 2001 Neural basis 4 Smith, P. L. & Ratcliff, R. 2004 Psychology and neuro- of a perceptual decision in the parietal cortex biology of simple decisions. Trends Neurosci. 27, (area LIP) of the rhesus monkey. J. Neurophysiol. 86, 161–168.(doi:10.1016/j.tins.2004.01.006) 1916–1936. 5 Wald,A.&Wolfowitz,J.1949Bayessolutionsofsequen- 24 Smith, J. D., Couchman, J. J. & Beran, M. J. 2012 The tial decision problems. Proc. Natl Acad. Sci. USA 35, highs and lows of theoretical interpretation in animal- 99–102. (doi:10.1073/pnas.35.2.99) metacognition research. Phil. Trans. R. Soc. B 367, 6 Ratcliff, R. 1978 A theory of memory retrieval. Psychol. 1297–1309.(doi:10.1098/rstb.2011.0366) Rev. 85,59–108. (doi:10.1037/0033-295X.85.2.59) 25 Kepecs,A.&Mainen,Z.F.2012Acomputationalframe- 7 Busemeyer, J. R. & Johnson, J. G. 2004 Computational workforthestudyofconfidenceinhumansandanimals. models of decision making. In Handbook of judgment Phil.Trans.R.Soc.B367,1322–1337.(doi:10.1098/rstb. and decision making (eds. D. Koehler & N. Harvey), pp. 2012.0037) 133–154.Oxford, UK:Blackwell. 26 Resulaj,A.,Kiani,R.,Wolpert,D.M.&Shadlen,M.N. 8 Ratcliff, R. & Rouder, J. N. 1998 Modeling response 2009 Changes of mind in decision-making. Nature 461, times for two-choice decisions. Psychol. Sci. 9, 263–266. (doi:10.1038/nature08275) 347–356.(doi:10.1111/1467-9280.00067) 27 Rabbitt, P. M. A. 1966 Errors and error correction in 9 Petrusic, W. M. & Baranski, J. V. 2003 Judging confi- choice-response tasks. J. Exp. Psychol. 71, 264–272. dence influences decision processing in comparative (doi:10.1037/h0022853) Phil.Trans.R.Soc.B(2012)

ebook img

Metacognition in human decision-making: confidence and error monitoring. PDF

2012·0.44 MB·English
by  Yeung, Nick
#journals #pubmed
Save to my drive
Quick download
Download
Upgrade Premium
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.

Preview Metacognition in human decision-making: confidence and error monitoring.

Phil. Trans.R.Soc. B(2012) 367, 1310–1321 doi:10.1098/rstb.2011.0416 Review Metacognition in human decision-making: confidence and error monitoring Nick Yeung* and Christopher Summerfield DepartmentofExperimentalPsychology,UniversityofOxford,SouthParksRoad,OxfordOX13UD,UK People are capable of robust evaluations of their decisions: they are often aware of their mistakes even without explicit feedback, and report levels of confidence in their decisions that correlate with objective performance. These metacognitive abilities help people to avoid making the same mistakestwice,andtoavoidovercommittingtimeorresourcestodecisionsthatarebasedonunreli- ableevidence.Inthisreview,weconsiderprogressincharacterizingtheneuralandmechanisticbasis of these related aspects of metacognition—confidence judgements and error monitoring—and identify crucial points of convergence between methods and theories in the two fields. This con- vergence suggests that common principles govern metacognitive judgements of confidence and accuracy; in particular, a shared reliance on post-decisional processing within the systems res- ponsible for the initial decision. However, research in both fields has focused rather narrowly on simple, discrete decisions—reflecting the correspondingly restricted focus of current models of the decision process itself—raising doubts about the degree to which discovered principles will scale up to explain metacognitive evaluation of real-world decisions and actions that are fluid, temporally extended, and embedded in the broader context of evolving behavioural goals. Keywords: decision-making; confidence; metacognition; error monitoring 1. INTRODUCTION andtheroadbecomesslipperyandwet.Theapproach (a) Metacognition in perceptual choice toeachbendstillrequiresadeftanglingofhandlebars, Imagine yourself cycling along a narrow, winding but now you are less certain about whether each countrylaneonasummer’sday.Successfullynegotiat- chosen action is optimized to steer the bicycle in a ing each twist and turn in the road requires the fashion that appropriately meets the gradient of the interpretation of a variety of subtle sensory cues and curve and the camber of the road. Action selection their conversion into appropriate motor commands. proceeds as before, but you are unsure whether the For example, incoming visual signals that reveal the actions chosen are the most appropriate ones. In curvature of the road, or motion parallax signals aris- fact, you might even experience an immediate sense ing from the trees beyond, allows you to gently that you have just made a poor choice. In short, you adjust the handlebars towards the left or right to are less confident about your decisions. steer the bicycle smoothly round each bend. Curiously, despite universal agreement that an Over past decades, psychologists and neuroscien- accompanying sense of confidence is a subjectively tists have devoted substantial effort to understand salient property of almost all our decisions, there is the neural and computational mechanisms by which currently little consensus about how we might incor- actionsareselected on the basis of a streamof incom- porate decision confidence into formal models of ing sensory information [1]. Because this sort of choice behaviour or explore its biological substrates. sensorimotor control often requires an observer to Fundamental questions remain unanswered. For commit to one discrete action from among several example, is the information that gives rise to the possible candidates, this literature has been informed ‘second-order’estimateofconfidenceinachoiceiden- by computational models that offer a formal account tical to that determining the ‘first-order’ choice itself? of how categorical decisions are made. These Why are we generally more sure that we are correct models,describedinmoredetailbelow,haveprovided than that we have made an error, even for difficult an excellent account of the decisions and decision choices? Why do we sometimes appear to ‘change latencies exhibited by observers selecting actions on ourmind’afteramotorprogrammehasbeeninitiated? the basis of ambiguous sensory information [2–5]. Arethesechangesofmindnecessarilyaccompaniedby However, now imagine that as you are cycling, the awareness that the initial choice was incorrect? skydarkensanditbeginstorain—visibilityisreduced, In what follows, we review the literature that has posed these and related questions about decision con- fidence in perceptual choice tasks. Subsequently, we *Authorforcorrespondence([email protected]). highlight links between this work and a literature that has considered how people monitor whether they One contribution of 13 to a Theme Issue ‘Metacognition: computation,neurobiologyandfunction’. have made an error under conditions of uncertainty 1310 Thisjournalisq2012TheRoyalSociety Review. Confidence and error monitoring N. Yeung & C. Summerfield 1311 time This simple model has much to recommend it both as a normative and descriptive account of categorical choice. The DDM has been widely and very suc- q cessfully applied to decision-making in a range of cognitive domains—from low-levelperceptualdecisions a to retrieval of facts from semantic memory, to econo- e response time for trial 1 mic decision-making under uncertainty [6–8]—and c n accounts neatly for several empirical characteristics of e d vi reaction times (RTs) observed for binary choices in e b response time for trial 2 thesetasks.Firstly,thenaturalgeometryoftheaccumu- lation-to-boundprocesspredictstheobservedrightward –q skew in RT distributions. Secondly, varying the bound offers an elegant account of the economy of speed and accuracy that characterizes mental chronometry tasks. Thirdly,byallowingdriftrateandoriginofaccumulation Figure1.Thedrift-diffusionmodel.Accumulatingevidence tovaryacrosstrials, the model accounts for the relative (thedecisionvariable,y-axis)overtime(x-axis)isshownfor RTof correct and error trials under conditions where twoillustrativetrials(markedaandb,greyandblacklines), speed or accuracy are emphasized [8]. Moreover, the one on which the choiceuis made and the other on which model implements a sequential probability ratio pro- thechoice –uismade.Adecisionistriggeredwhenevidence cedure that optimizes speed for a given error rate, and reachesuor –u.Grey line, trial 1;blackline, trial 2. thus takes its place among a family of statistically optimaldescriptionsofthechoiceprocess[2]. orconflict.Finally,weproposesomepotentiallyfruitful avenues for research that draw upon common themes 2. DECISION CONFIDENCE in these two literatures, building on their shared However, as illustrated above, we not only make strengthsandaddressingtheirsharedlimitations. decisions,butalsoconcurrentlyevaluatethelikelihood that those decisions will result in favourable or unfavourable outcomes. How, then, can we incorpor- (b) Formal models of perceptual choice ate decision confidence into the formal framework Even under good viewing conditions, visual infor- offered by the DDM and other quantitative models mation is corrupted by multiple sources of noise and of perceptual choice? In what follows, we review uncertainty, arising both in the external world and in recent empirical and modelling work that has the dynamics of neural processing. One sensible way attemptedtoextendthisformalframeworktoaccount to increase the signal-to-noise ratio of visual infor- formetacognitivejudgements.Muchofthisdebatehas mation is to sample the external world repeatedly hinged on a simple question: can confidence be read and integrate this information over time, making a outdirectlyfromtheDVatthetimeofchoice(decisio- decision only when the information is considered to nal locus model), or do confidence judgements be of sufficient quality [5]. This idea forms the basis depend on new information arriving beyond the of a class of model in which binary choices are decision point (post-decisional locus models)? [9]. described via an accumulation-to-bound mechanism, As an aside, we note that this question only arises with successive samples of information totted up if one is concerned with modelling the dynamics until they reach a criterial level or ‘bound’, upon of the decision process, such as with the DDM or which a commitment to action is made (figure 1). relatedevidence-accumulationmodels.Otherdecision- Different versions of this model make divergent theoretic models, such as signal detection theory assumptions about exactly what quantity is integrated (SDT),havegivendetailedtreatmentofhowtomeasure en route to a decision—i.e. about precisely how the observers’ metacognitive sensitivity, specifically, their ‘decisionvariable’(DV)iscomposed.Tosimplifymat- abilitytodistinguishtheirowncorrectandincorrectjud- ters, in this review, we focus on one variant of this gements [10–12]. However, because SDT is a static model—the ‘drift-diffusion’ model, or DDM—in model of the decision process in which the temporal which the DV corresponds to the relative likelihood dynamicsofevidenceaccumulationareignored,primary of the two choices being correct, given the stimulus (‘type I’) and metacognitive (‘type II’) judgements [3]. In the DDM, the DVon sample v is updated on t mustnecessarilybe basedonthesameevidence(albeit each sample t with an increment composed of two potentially corrupted by different sources of noise quantities: d, a linear drift term that encodes the rate [12]), and thusthese models are by definition decision of evidence accumulation, and cW, Gaussian noise locusmodels. with a mean of zero and a variance of c2: n ¼n þdþcW: ð1:1Þ (a) Decisional locus models t t(cid:2)1 The very earliest investigations of decision confidence Decisions are made when the DV exceeds a fixed revealed that, perhaps unsurprisingly, we are more deviation from zero, u, such that during evidence certain about our perceptual choices when sensory accumulation: inputs are stronger [13], and when we are given longer tosamplesensoryinformation[14].Itthusfol- (cid:2)u.v.u: ð1:2Þ lows that confidence should reflect both the quality of Phil.Trans.R.Soc.B(2012) 1312 N. Yeung & C. Summerfield Review. Confidence and error monitoring the evidence (represented by the drift rate,d) and the onlow-confidencetrialsweresubstantiallyattenuated, quantity of evidence at the choice point (represented falling equidistant between those for the chosen and bytheabsolutevalueofthedecisionbound,juj).How- unchosen targets. The authors interpret these data as ever, it is also clear that neither of these quantities showing that confidence is a simple product of quan- aloneissufficient todescribeanobserver’sconfidence tity and quality of the DV, as proposed by early intheirchoices.Anydiffusion-to-boundmodelassum- accounts of decision confidence, obviating the need ing that confidence is directly indexed by the state of for a separate metacognitive monitoring process [20]. evidence at the time of choice inevitably predicts that Kepecs et al. [19]employed an odourdiscrimination all choiceswill be madewith preciselythe same confi- paradigm in conjunction with single-cell recordings in dence(correspondingtotheevidencelevelrequiredto the rat orbitofrontal cortex (OFC; see also Kepecs & reachthebound).Similarly,anymodelproposingthat Mainen [25]). They report an intriguing finding: that confidence reflects evidence quality implicitly assumes OFC firing rates discriminated between correct judge- that observers have direct access to this quantity— mentsanderrorsafterthechoicebutbeforetheoutcome which,iftheyhad,wouldobviatetheneedforasequen- had been revealed, even when objective difficultyof the tialsamplingapproachinthefirstplace(seePleskac& choiceparadigmwascontrolledfor.Becausevariancein Busemeyer[15]foranexcellentrecentreview). theseresponsescouldnotbeexplainedbyotherfactors, Various accounts dating back over 100 years such as the reinforcement history over past trials, the have therefore proposed that confidence reflects authors suggest that these neurons encode confidence some interaction between the quantity and quality of estimates associated with the current decision. They evidence—for example, that under the DDM, confi- showthattheactivityoftheseneuronscanbeexplained dence scales with the product of d and u [16,17]. byaclassofmodelinwhichtwoevidencetalliesraceto This account has attractive properties—for example, thedecision bound (rather than a single quantityrepre- it predicts that when the observer has the option of senting their difference, as in the DDM). Specifically, terminating evidence accumulation with a choice at neuronal responses reflected the difference between any point, RTs are faster for high-confidence trials thetwototalsatthetimeofthedecision—makingthisa [18]. Moreover, this view enjoys support from two decisional locus model, even thoughthe relevant neural prominent neurophysiological studies detailed in §2b. activity was sustained through the post-response period—andthisactivityinturnpredictedthelikelihood thatthechoicewascorrect[19]. (b) Neural substrates of decision confidence At first glance, these two studies offer convincing Two recent single-cell recording studies, one involv- evidence thatneuronsintheparietalandorbitofrontal ing rats [19] and the other monkeys [20], claim to cortices encode the subjective confidence associated have identified neurons encoding subjective decision with achoice: agraded quantitythatcan be estimated confidence. Kiani et al. [20] recorded from neurons directly fromthe evidence on which the original, first- inthelateralintraparietalcortex(LIP)ofmacaquesdis- orderchoice was based. In both cases, the authors are criminating the direction of motion in ambiguous assiduous in attempting to rule out alternative expla- moving-dot arrays [21] with a saccade to one of two nations of their findings, such as attention or learned targets placed either side of the motion stimulus. This reward contingencies. However, to return momenta- experimental approach has previously been used to rily to our discussion of models of categorical choice, demonstrate that LIP neurons whose receptive field two well-replicated behavioural phenomena cast overlaps with the chosen target display a characteristic doubt on any model in which subjective confidence acceleration of spiking activity while the monkey views directly reflects the evidence accumulated up to the the moving-dot stimulus, whereas those overlapping choice point. First, it has consistently been shown with the non-chosen target show a relative dampening thatconfidenceincorrectchoicesisstrongerthancon- of activity [22,23]. This build-up scales with signal fidence in incorrect choices, even when choice strength, and terminates when an action (saccadic eye difficulty is controlled. As previously pointed out movement) is selected, prompting the view that LIP [15], this finding cannot be explained by decisional firing rates encode the level of evidence available for a locus models in which confidence is a mere function choice—inotherwords,their firing rates form aneural of diffusion model parameters, because these para- representationoftheDVproposedbytheDDM[1]. meters are invariant across correct and incorrect To investigate decision confidence, Kiani et al. trials (but see Galvin et al. [10] for relevant analysis added a new twist to this protocol. Typically, the usingSDT). Second, andcrucially,subjects occasion- monkey is rewarded for correct but not incorrect ally change their mind between the first-order choice choices, but here, on a fraction of trials, the monkey and the second-order estimation of decision confi- was offered a ‘sure bet’ option such that a certain dence. At least on these trials, the decisions must be but lower-valued reward could be obtained via a sac- influenced by processing that occurs after the first- cade to a third response option. In ethology, it is order choice point. In §2c, we consider models that commonly assumed that an animal choosing the safe propose such a mechanism. option when evidence is scarce must have access to metacognitive information about the likelihood that it will make an error on the main task [24]. The (c) Post-decisional locus models and changes authors found that not only did their monkeys use of mind thisoptionjudiciously—respondingwhenthestimulus Resulaj et al. [26] report a behavioural experiment in wasweakorambiguous—butalsothatLIPfiring rates which human subjects indicated the direction of a Phil.Trans.R.Soc.B(2012) Review. Confidence and error monitoring N. Yeung & C. Summerfield 1313 random dot motion stimulus by moving a handle to a predicting decision confidence might not be confined leftwards or rightwards target some 20cm away. This tothestimulationperiodalone.However,onepotential design allowed the researchers to isolate trials on caveattothisviewisthattheactivityofLIPneuronsis which subjects began to move towards one target but known to drop off sharply once the eye movement is then changed their mind and veered-off towards the made [22,23]. Thus, it remains to be shown whether other. Careful behavioural analyses demonstrated a variation in the post-decisional LIP signal could con- number of interesting phenomena. Firstly, changes of tribute to a later confidence judgement, or whether a mind were not symmetric: subjects switched more separaterepresentationoftheevolvingDVissupporting often from an incorrect to a correct choice. Secondly, theobservedchangesofmind.Thiscaveataside,these change-of-mind trials tended to occur when, owing emerging findings suggest that post-decisional proces- to stochasticity in the stimulus display, motion sing plays a crucial role in metacognitive judgements, energy began by favouring the initial choice, but sub- which can lead to changes of mind or support ratings sequently came to favour the alternative, switched-to ofconfidenceinaninitialdecision. option.Becausethemotionstimulusoffsetoncemove- ment began, subjects must have capitalized on the balance of information in the immediate pre-decision 3. ERROR MONITORING period when deciding to change their mind. Notably, People are often aware of their own mistakes; for although this motion information was available prior example, in choice RT tasks when time pressure is to the decision, the switch occurred only once move- applied to induce errors in simple judgements. Error ment initiation began, suggesting that evidence monitoringisthemetacognitiveprocessbywhichweare accumulation continued beyond the point at which abletodetectandsignalourerrorsassoonasaresponse the initial choice was made. has been made. This process plays a crucial role in To account for these and other data, a number of adaptive human behaviour, allowing our actions to be researchers have proposed models in which, in con- shaped by their outcomes both in the short term, for trast to the classical DDM, evidence accumulation example,byrespondingmorecautiouslytoavoidfurther continues even beyond the choice point, with this errors,andinthelongerterm,throughgraduallearning extra variability in the DV also contributing to esti- ofappropriatestimulus–responsecontingencies. mates of subjective confidence when probed at a Whereas the studies described above have typically later time point [9,15]. Resulaj et al. propose that asked subjects to report their confidence that they their data can be explained by just this type of made the correct choice, studies examining error model, with changes of mind occurring when latent monitoring have tended to ask subjects the converse information in the processing pipeline drives the DV question—i.e. to report the likelihood that they made across a second, ‘change-of-mind’ bound. A related an error. Although these judgements seem like two account, the two-stage dynamic signal detection sides of the same coin, methods and assumptions in (2DSD) model [15], likewise proposes that the diffu- the two literatures have often been quite different. For sion process continues beyond initial choice, with example,decisionconfidencehastypicallybeenstudied decision confidence reflecting the absolute value of usingtasksthatremainchallengingevenwhenextended the DVat the post-decision point at which a second- processing times are permitted, such as challenging orderdecisionisrequired.Thus,incontrasttodiscrete psychophysical discriminations. Under these circum- changesofmindthatareexpressedasovertcorrections stances, subjects are sometimes sure they responded ofaninitialresponse,the2DSDmodelallowsforcon- correctlyandsometimesunsurewhether theyare right tinuously varying levels of confidence that provide an orwrong,butrarelycertainthatthattheymadeamis- explicit judgement about an earlier response. take [15]. In contrast, research on error monitoring Behaviourally,thesepost-decisionprocessingmodels has mostly been studied using simple but time- are able to account for a broad range of findings con- pressured tasks in which subjects are usually aware of cerning decision confidence. Firstly, they correctly their errors and very rarely unsure whether their predict that observers will change their mind more decision was right or wrong. Framed in terms of the often from incorrect to correct responses than vice DDM,thedistinctionconcernswhethererrorsandsen- versa, because beyond the bound the DV on error sitivitytoprocessingnoisearisebecauseoflowdriftrate, trials will tend to regress towards the mean, whereas d,inthecaseofperceptualambiguity,oradoptionofa after correct responses it will continue to grow, driven lowthreshold,u,toengenderfastresponding.Thereis bythetrue underlying drift rate.This observation also nonetheless obvious similarity between metacognitive naturally explains another conundrum associated with judgements of confidence and error likelihood, and it decision confidence: that second-order confidence is isthereforeunsurprisingthat models oferror monitor- generally higher for correct trials than incorrect trials. ing turn out to complement those more recently Indeed, reconsidering for a moment the data from developedinperceptualdecision-makingresearch. Kiani et al., we note that these authors report that neuralactivityinthedelaybetweenstimulusoffsetand (a)Post-decisionprocessinginerror monitoring movement execution exerted a separate, independent Rabbittandco-workers’pioneeringworkbeginning in influence over the decision to choose the ‘sure bet’ the 1960s established that error monitoring relies on option [20]. Thus, although the authors argue that a post-decision processing. Their experiments showed mechanistic description of decision confidence does that people can very reliably detect and correct their notrequireustoinvokeadistinctmetacognitiveprocess own errors without requiring explicit feedback [27], separate from evidence accumulation, the evidence butthatthisabilityisimpairedwhenstimulusduration Phil.Trans.R.Soc.B(2012) 1314 N. Yeung & C. Summerfield Review. Confidence and error monitoring time uncertainty—or conflict—in the decision process after an initial response [34], or as inconsistency between the outcomesofparalleldecisionprocessesat different (a) processing stages (e.g. perceptual categorization and q response selection) [35]. While varying in their details and precise predictions [32,34], common to all pro- ce (b) qcom posals is the claim that metacognitive accuracy en judgementsdependonpost-decisionprocessing. vid (c) In effect, applying the framework provided by the e DDM and its variants, one is obliged to assume that –q error detection reduces to error correction: that errors are detected whenevercorrective activity reaches a cri- (d) teria level—the change-of-mind bound. However, this conclusion is difficult to reconcile with evidence that errorcorrectionanddetectionareatleastpartlydissoci- Figure2.TheoriesoferrordetectionwithintheDDMframe- able, a finding that was again presciently reported in work. The drift-diffusion process is illustrated schematically Rabbitt’sseminalwork.Rabbitt’sstudiesdemonstrated fortwotrials,oneinwhichdecisionuisthecorrectresponse that error corrections are a relatively automatic and andonetrialinwhichthisdecisionisincorrect.Bothdecisions unreflective consequence of post-decision processing. occur at the same time point (a). Following the correct response (grey line), post-decision processing continues to Thus, they can be extremely fast, occurring within accumulate in favour of the decision just made. Following 10–20msoftheinitialerror[36],andmaybeproduced errors (black line), the drift rate regresses to its true mean, evenwhensubjectsareinstructedtoavoiddoingso[37]. causing the DV to re-cross the decision bound (b), sub- Incontrast,explicitdetectionandsignallingoferrorsis sequently cross a change-of-mind bound (c), and finally muchslower,voluntary,more pronetointerference by crosstheoriginallycorrectdecisionbound, –u(d).Thegrey distractingtasks,andmoresensitivetocognitivedecline shaded area indicates a period of uncertainty, or conflict, in normal ageing [29]. Indeed, people sometimes betweenthere-crossingoftheubound(b)andlatercrossing remainunawareoferrorsthattheyneverthelesscorrect ofthe –ubound(d). [38].Collectively,thesedifferencessuggestthatexplicit error detection cannot be a mere consequence of is reduced [28], suggesting its dependence on contin- post-decision correction: further processing or evalu- ued processing of the stimulus after an initial error ation must intervene between initial correction and (which is curtailed when stimuli are very briefly pre- later explicit awareness that an error has occurred. sented). Error monitoring is also impaired when Consistent with this analysis, recent investigations subsequent stimuli appear soon after the initial have identified dissociable neural correlates of error response [29], and responses to those later stimuli correctionanddetection. are themselves postponed following errors [30], con- sistent with the notion that this monitoring involves (b) Neural substrates of error monitoring the same mechanisms as the initial decision. Researchinterestinerrormonitoringincreasedsubstan- Summarizing these findings, Rabbitt likened evi- tially following the discovery of scalp EEG potentials dence accumulation in decision-making to votes in a that reliably occur time-locked to decision errors. committee,inwhichincorrectdecisionsaresometimes Most studies have used the Eriksen flanker task, in made on the basis of incomplete information, but ‘as which subjects perform a speeded categorization of a subsequent votes come in, a more accurate consensus centraltarget(e.g.HorS),whileignoringflankingdis- will accumulate and the earlier mistake will become tractorsthataresometimescompatible(e.g.HHH)and apparent’[28].Thus,errorsarecharacterizedbybipha- sometimes incompatible (e.g. SHS) with that target. sicevidenceaccumulation,withinitialaccumulationin Withmodestspeed pressure,error rates onincompati- favour of the incorrect response followed by later drift ble trials can exceed 20 per cent. Following such towards the correct decision (as the trial-average drift errors,anegativeevent-relatedpotentialoverfrontocen- rateregressestothetruemean).Bycontrast,continued tral sites is observed within 100ms of the incorrect evaluation following correct responses tends simply response, followed by a later positive wave peaking to reinforce the original decision. This model is an 200–400ms later over parietal cortex [31]. Labelled obvious precursor to more recent accounts of decision theerror-relatednegativity(ERN/Ne)anderrorpositiv- confidence[15]andchangesofmind[26]. ity(Pe),respectively,theseEEGcomponentshavebeen All subsequent models of error detection have widelystudiedtoprovideinsightintoerror monitoring adoptedRabbitt’sbroadframework,withdebatefocus- inhealthyandclinicalpopulations. ing instead on the precise mechanism by which post- Converging evidence has identified anterior cingu- decision processing leads to error detection. Figure 2 late cortex (ACC) as the source of the ERN. For illustrates some key model variants. Within a standard example, in simultaneous EEG–fMRI recordings, DDMframework,errorscouldbedetectedassuccessive single-trial ERN amplitude correlates reliably only crossingsofdecisionboundariesforthetwocompeting with activity in a focused ACC source [39]. The responses [31,32] or as ‘double crossings’ of a single sourceofthePeislesswellcharacterized,butevidence decision bound [33]—both close relatives of Resulaj that it is a variant of the well-studied P3 component et al.’s notion of a change-of-mind bound. Errors can [40] would imply widely distributed neural generators also be detected in terms of the occurrence of inparietalandprefrontalcortex[41].TheP3hasbeen Phil.Trans.R.Soc.B(2012) Review. Confidence and error monitoring N. Yeung & C. Summerfield 1315 suggested to reflect the distributed action of norepi- focused on the finding that subjects usually respond nephrine released by the locus coeruleus brainstem more slowly on trials immediately following errors nucleus in response to motivationally salient events [27].Althoughthiseffectatleastpartlyreflectsthedis- during decision-making [42]. tracting occurrence of a rare event [50], as errors The functional significance of the ERN and Pe is a typically are, a consensus view holds that post-error matter of ongoing debate. Competing theories of the slowing reflects strategic adaptation to prevent further ERN propose a role in error detection, reinforcement errors [51]. EEG studies have subsequently shown learning and conflict monitoring; while theories of that the degree of observed slowing scales with the the Pe include conscious awareness of errors, affective magnitudeoferror-relatedERN/Peactivity[46].Com- responses and behavioural adjustments to avoid putational models implementing error-related control further errors [43]. These debates notwithstanding, over distance-to-bound, juj, account for detailed prop- it is now clear that the ERN and Pe dissociably map erties of empirically observed post-error slowing. In onto processes related to error correction and error one class of model, detection of response uncertainty detection, respectively. Thus, ERN onset coincides (conflict) immediately following error commission with the onset of error-correcting activity as revealed leads to an increase in the bound—and, hence, more through EMG recordings [44], typically around the cautious responding—on subsequent trials [52]. time of error commission [31], and its amplitude Recent extensions of this idea suggest that conflict varies with both the speed [45] and probability [46] detection may also be used to adjust the bound oferrorcorrection. In contrast, Pe amplitude isinsen- dynamicallyevenasadecisionisbeingmade[51,53]. sitive to the strength of correcting activity when error Error signals not only support subtle adjustments detection rates are controlled [47]. Conversely, that optimize online decision-making, they also play a although both the ERN and Pe are found to covary key role in longer term adjustments during learning. with subjective ratings of response accuracy [46,48], Holroyd & Coles [35], for example, suggest that the correlations involving the ERN disappear when the ERN reflects reinforcement learning of action values. two components are carefully dissociated. In antisac- They showed that the ERN migrates in time as new cade tasks in which subjects correct all of their errors stimulus–responsemappingsarelearned,frominitially butdetectonlyhalfofthem,ERNamplitudeisequiv- being triggered by environmental feedback to later alent for aware and unaware errors, whereas the Pe is being driven by internal representations of the learned robustly observed only when subjects detect their mappings, a pattern that mimics the migration of errors [38]. Taken together, these findings suggest dopaminergic reward prediction error signals from that whereas the ERN directly indexes automatic unconditionedrewardstopredictivestimuliduringcon- post-decision processes leading to rapid error correc- ditioning [54]. Meanwhile, fMRI activity in ACC and tion, the later Pe is selectively associated with explicit neighbouringcortexatthetimeofanincorrectresponse detection and signalling of errors. These results thus have been shown to predict response accuracyon later provide converging evidence for the view that error presentationsoftherelevantstimulus[55,56]. correction and detection reflect distinct processes. Moststudiesofpost-erroradjustmentshavefocused Arecentstudyhasshedlightonspecificallyhowthe on the ERN and ACC, reflecting wide interest in the Perelatestoerrordetection[49].Subjectsperformeda role of ACC in reinforcement learning [35,57], difficultbrightnessdiscriminationunderspeedpressure rather than on the later Pe component. However, the toinduceamixtureoferrorsresultingfromperceptual ERN and Pe typically covary across conditions and, ambiguity and decision urgency. After each response, when the two components are dissociated, post-error they made a binary correct/error judgement, with adjustments are only observed following detected monetary incentives varied across blocks to encourage errors on which a Pe component is present [38], either liberal or conservative error signalling. Signal suggesting that the latter may be a more direct corre- detectionanalysisindicatedthatsubjects’accuracyjud- late of the learning mechanisms by which future gements could be well fit by assuming that these behaviour is adapted following an error. judgementsreflectedacontinuumofconfidence(from sure correct to sure incorrect), with subjects applying a criterion—a metacognitive u—that varied according 4. INTEGRATIVE MODELS OF DECISION to the incentive regime. Critically, error signallingper- CONFIDENCE AND ERROR MONITORING formancewascloselyrelatedtobetween-conditionand The discussion above highlights convergent evolution trial-by-trial variation in Pe amplitude (but not the in work on decision confidence and error monitoring, ERN). That is, Pe amplitude appeared to provide a suggesting that common principles govern metacogni- direct neural index of continuously varying decision tive judgements regardless of whether errors arise confidenceonwhichsubjectsbasedtheirmetacognitive because of intrinsic task difficulty (low drift rate, d) judgements,withcategoricalsignallingoferrorsoccur- or because of externally imposed decision urgency ring when confidence that the response was wrong (low threshold, u). In what follows, we consider the exceededsomecriteriallevel. implications of this convergence for future research, both positive (in terms of mutually informative les- (c) Impact of error monitoring on behaviour sons) and negative (in terms of shared limitations). Theresearchdescribedabovedocumentsthemechan- istic and neural basis of error monitoring. A parallel (a) Converging themes lineofresearchhasconsideredtheimpactoferrormoni- Theories of decision confidence and error monitoring toring on future behaviour. Much of this work has each emphasize the importance of post-decision Phil.Trans.R.Soc.B(2012) 1316 N. Yeung & C. Summerfield Review. Confidence and error monitoring processing, and likewise point to a key dissociation behaviourasaseriesofdiscretedecisions,eachsubject between, on the one hand, continued processing of toindependentmetacognitivescrutiny,isausefulcon- stimulus information within the decision-making venience when developing experimental tests and system (leading to changes of mind and error correc- formal models. However, it remains unclear whether tions) and, on the other, the formulation and the findings will scale up to explain real-world expression of explicit judgements of confidence and decisions and actions that are fluid, temporally accuracy. The literatures have developed these ideas extended and embedded in the broader context of in complementary directions, such that each literature evolving behavioural goals. offers valuable lessons for the other. First, whereas Letusrevisitforamomentouropeningexampleof confidence is typically characterized as varying along cycling along on a winding country lane. Obviously, a continuum, and formalized as such in accounts we do make occasional discrete categorical such as the 2DSD model [15], error detection is decisions—for example, when choosing between two often characterized as an all-or-none process [32,33]. available paths at a fork in the road. However, most Thus, according to many current theories of error of our decisions and actions unfold gradually, shaped monitoring, binary yes/no error judgements are an by our interactions with the environment and an intrinsic feature of the monitoring system rather than ever-changing stream of incoming sensory infor- a reflection of the arbitrary metacognitive decision mation, as exemplified by the continuous, subtle thatsubjectsareaskedtomake.Assuch,thesetheories adjustmentsofhandlebarsandbrakesneededtomain- cannotexplainhowsubjectsareabletoexpressgraded tain direction and balance when cycling. Current confidence in their accuracy judgements [48,49]. theories of metacognition require a clear division in Conversely, the preceding review of work on error time between cognitive decision-making and meta- monitoring identifies important limitations in current cognitive evaluation, with the latter beginning when theories of decision confidence. This review suggests the former ends (‘post-decision processing’). However, that the current debate between decisional locus and for continuous and extended actions, there is no post-decisional models of decision confidence will definable time-point at which a specific decision is very likely be resolved in favour of the latter, but that finalized and metacognitive evaluation begins. Most even these post-decisional theories will need modifi- current theories are therefore ill-suited to describing cation to accommodate evidence that metacognitive thetemporallyextendedactions,comprisinghundreds awareness (cf. error detection and the Pe) cannot be or thousands of micro-decisions, that characterize reduced simply to post-decision processing (cf. error everyday behaviour. correction and the ERN)—the two are at least partly Indeed, decisions that initially appear discrete and dissociable. Perhaps more informative still will be categorical may turn out, on closer inspection, to be exploration of the role of confidence judgements in graded and transitory. For example, in the Eriksen guiding future actions: whereas research on decision flanker task, overt responses occur tightly time- confidencehaslargelyfocusedonhowconfidenceesti- locked to the point at which lateralized activity in mates are derived, a major focus of error-monitoring motor cortex exceeds a threshold value [58], just as research has been on how this kind of metacognitive saccadiceyemovementsappeartobetriggeredaccord- information might be used to modify behaviour both ing to a threshold firing rate for LIP neurons [1], in the short- [27,52,53] and long-term [35,55,56]. suggestive of a fixed decision point after which an Borrowing these insights, we might predict that para- action is initiated. However, fine-grained analyses metric estimates of confidence could support finer- reveal graded and continuous information flow at grained control of behaviour than can be achieved every stage, even downstream of motor cortex. Thus, through binary categorization of responses as correct within a single trial, motor cortex activity may latera- or incorrect—for example, by allowing parametric lize first towards one response then towards the variation in post-error slowing or by providing a other; small EMG twitches in one finger may be fol- scalar predictionerrorsignaltosupportreinforcement lowed by full movements of another; and overt learning. Thus, confidence estimates could provide actions themselves may vary in force in a graded useful information in optimizing the rate of learning: manner, for example, with incorrect actions executed onemightpredictthatpeoplewillpaygreaterattention less forcefully than correct ones [44,58,59]. Mean- to environmental feedback following responses in while, categorical or economic judgements about which they have less confidence. visual information are often preceded by exploratory eye movements, which may themselves constitute interim decisions en route to the eventual choice [60]. (b) Shared limitations In such systems, there is no single, final decision As well as sharing complementary strengths, the the- point that could mark the beginning of metacognitive ories also share common weaknesses. In particular, evaluation. Detailed analyses of neural markers of like the models of decision-making on which they are metacognition point towards a similar conclusion: based, current theories of confidence and error moni- error-relatedEEGactivityisnotonlyobservedfollow- toring have focused almost exclusively on decisions ing overt incorrect actions, but also following ‘partial’ that are discrete and punctate in time: a decision is errors in which the incorrect muscles twitch, but the made when a boundary is reached [3]; errors are incorrect action is not produced [61] and, crucially, detected when a second change-of-mind bound is in a graded fashion as a function of the level of sub- crossed [26]; and confidence is estimated at the time threshold cortical activity favouring an incorrect of a later metacognitive probe [15]. Characterizing response [34]. Metacognition appears to be graded Phil.Trans.R.Soc.B(2012) Review. Confidence and error monitoring N. Yeung & C. Summerfield 1317 and continuous in just the same manner as the revealed that human observers are exquisitely sensi- underlying decision process. tivity to evidence reliability when sampling from Human decision-making also has a continuous multiple sources of information [67]. quality whenviewedover longer timescales, with indi- Yet, most currently popular models of perceptual vidual decisions chained into sequences that serve decisions offer no way of expressing the trust or dis- longer-term behavioural goals. Thus, actions that trust associated with the evidence accumulated; all reflect definitive choices at a lower goal level (e.g. sourcesofevidencearecombinedinthecommoncur- movements of the arm, shifts of posture) may consti- rency of the DV, which then gives the strength of tute reversible, interim choices at a higher goal level evidence as a simple scalar value (the vertical location (e.g.turningleftataforkintheroad),andevenacom- ofthediffusingparticleintheDDM,orthemagnitude mitment to an individual right or left turn might be of the evidence variable in SDT). Once factored into just a partial commitment to a yet higher level goal the DV in this way, the reliability of evidence has no (e.g. reaching a specified destination) [62]. This further impact on the decision-making process. This form of hierarchical structure is built into many feature contrasts with some accounts of economic recent theories of the computational and neural basis choices, in which sensitivity to risk (i.e. outcome var- of action selection [63], in which common principles iance) as well as value (i.e. outcome mean) has been of selection and control are held to operate at each documentedatthebehaviouralandneurophysiological level of hierarchical abstraction [64]. Recent findings levels [68,69]. indicatethatmetacognitiveprocessesaresimilarlysen- Mathematically, the reliability of perceptual evi- sitive to this hierarchical structure [65]. For example, dence is orthogonal to its strength, because the mean errors that are equally discrepant in terms of low- and the variance represent different moments of a level actions are treated very differently according to probability distribution. Bayesian models that exploit their impact on global task performance [66]. Little this point—representing evidence as a probability dis- is currently known about the mechanisms by which tribution with a given mean (evidence strength) and metacognitive judgements might be embedded in variance (evidence reliability)—have been used to cal- ongoing higher level behaviour in this way. culate ideal estimates of expected value in economic Thus, a crucial limitation of current metacognitive choice tasks [70]. Applied to perceptual decision theoriesisthattheydonotreflectthewaythatourcon- tasks, the notion would be that the diffusing parti- fidence in our actions contributes to fluid, structured cle of the DDM is more accurately conceived of as a sensorimotor behaviour. Rather, they consider that probability distribution that evolves across samples errors are processed in an all-or-nothing fashion, for (figure 3), with the central tendency of that distri- example,whenapost-decisionprocesscrossesameta- bution analogous to the vertical location of the cognitivedecisionbound.Inwhatfollows,weconsider particle. Crucially, the variance of this distribution anotherwayofthinkingaboutdecisionconfidencethat provides additional information—specifically, a rep- is not subject to these limitations. resentation of evidence reliability in terms of the precision of the mean—that is not made explicit in the standard DDM (by precision, we mean the inverse 5. FUTURE DIRECTIONS of the standard deviation). Addressing the challenges outlined above will require Recently, it has been shown that neural network the development of new models that consider not modelsinwhichLIPneuronsencodethefullposterior just discrete, binary choices, but also the kinds of probability distribution associated with a stimulus can extended, goal-oriented actions that characterize capture behaviour and neural dynamics occurring everydayhuman behaviour.We conclude by consider- duringpsychophysicaldiscriminationtasksinprimates ing one promising extension of current decision [71]. This result suggests that precision may be models, which addresses the issue of the reliability of encoded in the variance of firing rates across a neural evidence in a manner that opens intriguing new ave- population, thus signalling evidence reliability in nues for understanding how metacognitive evaluation much the sameway thatmean firing rateof the popu- might occur for continuous, extended actions. lation encodes evidence strength [72]. Relatedly, Ratcliff & Starns [73] have recently proposed a (a) Evidence reliability model of decision confidence in which the decision- Formal accounts of categorical decisions, such as the relevant evidence on a single trial is not a scalar esti- DDM, are often illustrated by analogy to court of mate, but a distribution of possible values, allowing law, in which the jury weighs up evidence favouring parametric estimates of confidence in the choice to the innocence or guilt of the defendant [1]. However, be calculated asthe integral of this distribution falling this analogy also highlights an inconsistency between within graded confidence bounds. Together, these current decision models and choices made in the real findings open promising new avenues for research world: that in the latter, we usually consider the into the neural basis of categorical choice. extent to which we trust the evidence relevant for a decision. For example, in a law court, evidence from a trustworthy source (for example, official telephone (b) Confidence revisited records)mightweighmoreheavilyinajury’sdelibera- We propose that evidence strength and reliability are tions than evidence from an unreliable source (for encoded in parallel during evidence accumulation, example, a witness with a vested interest). Corre- and that this framework provides an intriguing new spondingly, careful signal detection analyses have way of thinking about decision confidence—as the Phil.Trans.R.Soc.B(2012) 1318 N. Yeung & C. Summerfield Review. Confidence and error monitoring (a)ata) time t time t +1 Swehciochndelyv,idthenecme oqduealliitsyavbalerietso deveesncribweithsiitnuaatiosninsgilne d H | trial [71], something that standard models cannot p ( achieve. In fact, by keeping trackof the likely variabil- ity of information in the external world, Bayesian perceptual hypotheses perceptual hypotheses accounts can optimally distinguish true state changes (b) in the generative information giving rise to the senses ses low variability high variability p (H | data) fromnoise[74].Precisionestimatesarethusparticula- e h ot high rly useful in situations where the causes of perceptual p y evidence may change unpredictably over time, and as al h low such may provide a better account of the sort of fluid, u pt ongoing sensorimotor integration that characterizes e c everydayactivitiessuchasridingabicycle. er p The hypothesis also leads to clear and testable time time predictions about the sensitivity of human decision- Figure3.Schematicofamodelinwhichboththemeanand making to evidence reliability. Firstly, it predicts that variance of information in an array are estimated in a serial variabilityofevidencewithinatrialshouldbothreduce sampling framework. (a) The left panel shows the posterior accuracy and lengthen RT, because during sequential probability distribution p (H j data) over a continuous sampling, the precision of the mean increases more space of possible perceptual hypotheses (e.g. these dots are slowlywhenthesamplesaredrawnfromamorevariable moving to the right with 30% coherence; this signal is 40% distribution. This prediction has recently been con- visible; etc.) at a given time, t. This distribution reflectsthe firmed in an experiment in which observers made evidence sampled from the stimulus thus far, i.e. between discrimination judgements on the average feature (e.g. onset and time t (grey dots). The new sample received at colour) of an array of multiple elements presented timetisshowninred.Rightpanel:attimetþ1,thisdistri- simultaneously on the screen [75]. Critically, this bution (light grey) is updated in the light of the newly sampled information, giving rise to anewprobabilitydistri- multi-elementaveragingtaskallowedtheexperimenters bution.Inthismodel,confidenceisreflectedintheprecision tomanipulatethemeanandthevarianceoftherelevant oftheposteriordistribution,i.e.thereciprocalofitsstandard featureinanorthogonalfashion.Theresultsshowedthat deviation.(b)Theevolvingposteriorprobabilitydistribution observers were slower to discriminate more variable over perceptualhypotheses(y-axis)foreachsuccessivetime arrays,aresultthatispredictedbytheprecisionaccount point (x-axis; blue–red colourmap; red values indicate but not by standard accumulation models such as the higher probabilities). The posterior distribution is updated DDM. Moreover, observers in this study tended to followingthearrivalofsuccessivesampleswithlowvariance downweight outlying or otherwise untrustworthy evi- (left panel) or high variance (right panel). Precision of the dence,muchlikeastatisticianmightexcludeanoutlier probabilistic representation of evidence strength increases from a sample of data. Although this study did not morerapidlyfor thelow-variabilitysamples. assess subjects’ second-order confidence in their decision, the precision accountmakesthe clear predic- precisionofevidenceaccumulatedduringasingletrial. tion that in the multi-element task, both subjective Thisrepresentationofconfidencediffersintwocritical confidence and the rate of occurrence of changes of respects from the models described above: it is based mindwilldependonarrayvariabilitytoagreaterextent on an intrinsic feature of the decision process—var- thanonitsmean,apredictionripeforfuturetesting. iance in the evidence encountered—and it is Finally, this conception of decision confidence available continuously and instantaneously. In con- makes direct contact with broader theories of the trast, existing models suggest that metacognitive role of metacognitive evaluation in behavioural con- evaluations are derived indirectly by comparing rep- trol. In particular, because precision is closely related resentations of the DV from two (or more) discrete to the concept of decision conflict [52]—greater var- time points. Precision thus provides a more suitable iance in evidence should result in greater conflict basisfor metacognitiveevaluationinthekindsoftem- between competing response options—the theory can porally extended tasks discussed above, in which no inherit ideas from research on decision conflict about discrete decision point divides cognitive decisions how precision estimates might be used to guide both from metacognitive evaluation. current performance (e.g. through dynamic modu- This hypothesis, though speculative at this point, lation of decision bounds) [51,53] and future has several attractive features. Firstly, it is consistent behaviour (e.g. through modulation of learning rate with the evidence described above on post-decision in relation to environmental signals of success or fail- processing because instantaneous precision will tend ure) [70]. As such, the precision model not only to be highly correlated with later variability in the provides a formally specified account of decision con- DV(thebasisofmostcurrentmodelsofmetacognitive fidence,butalsoleadstoimmediatesuggestionsabout judgements): variable evidence will give rise to fre- the use of confidence judgements in the optimization quent errors, and these errors will be accompanied of behaviour. by low estimates of precision at the time of the initial response and a high probability of later changes of mind or error correction. Thus, the precision account 6. CONCLUSION agrees that existing theories fit the data well, but Formal models such as the DDM have proved extre- suggests that they may do so for the wrong reasons. mely valuable in understanding human and animal Phil.Trans.R.Soc.B(2012) Review. Confidence and error monitoring N. Yeung & C. Summerfield 1319 decision-making,bysituatingexperimentalobservations judgments. Psychon. Bull. Rev. 10, 177–183. (doi:10. ofbehaviourandneuralactivitywithinapreciselyspeci- 3758/BF03196482) fied and normatively motivated framework. Direct 10 Galvin, S. J., Podd, J. V., Drga, V. & Whitmore, J. extensions of these models have proved equally useful 2003 Type 2 tasks in the theory of signal detectability: discrimination between correct and incorrect decisions. in probing the metacognitive processes by which we Psychon. Bull. Rev. 10, 843–876. (doi:10.3758/BF0 evaluate and express our degree of confidence in our 3196546) decisions. In particular, significant convergence in 11 Evans,S.&Azzopardi,P.2007Evaluationofa‘bias-free’ methods and theories of decision confidence and error measure of awareness. Spat. Vis. 20, 61–77. (doi:10. monitoring suggest that common principles may 1163/156856807779369742) governdifferenttypesofmetacognitivejudgements. 12 Maniscalco,B.&Lau,H.2012Asignaldetectiontheor- There is nonetheless important scope for current etic approach for estimating metacognitive sensitivity modelstoconsiderdecision-makingandmetacognitive from confidence ratings. Conscious. Cogn. (doi:10.1016/ evaluationinsituationsthatencompassnotonlysimple, j.concog.2011.09.021) punctate choices, but also the kinds of extended, 13 Peirce, C. S. & Jastrow, J. 1884 On small differences of sensation. Mem.NatlAcad. Sci.3,73–83. goal-directed decisions and actions that typify human 14 Vickers,D.,Burt,J.,Smith,P.&Brown,M.1985Exper- behaviour outside the experimental laboratory. We imental paradigms emphasizing state or process have proposed one such extension: the hypothesisthat limitations: 1. Effects on speed accuracy tradeoffs. Acta peoplearesensitivenotonlytothestrengthofevidence Psychol.59,129–161. theyencounterastheymakeadecision,butalsotothe 15 Pleskac, T. J. & Busemeyer, J. R. 2010 Two-stage reliability of that evidence. This simple proposal has dynamic signal detection: a theory of choice, decision far-reaching implications: it immediately suggests a time, and confidence. Psychol. Rev. 117, 864–901. novel source of information—evidence precision—that (doi:10.1037/a0019737) could guide metacognitive evaluation. Future develop- 16 Link, S. W. 2003 Confidence and random walk theory. ments in theories of human decision-making promise In Modeling choice, decision time, and confidence 893 (ed. to have similarly profound implications for our under- B. B. E. Borg). Stockholm, Sweden: International Society for Psychophysics. standing of the way in which people evaluate their 17 Heath, R. A. 1984 Random-walk and accumulator decisions in the service of adapting and optimizing models of psychophysical discrimination—a critical those decisions in the face of an uncertain, complex evaluation.Perception13,57–65.(doi:10.1068/p130057) andever-changingenvironment. 18 Vickers,D.&Packer,J.1982Effectsofalternatingsetfor speedoraccuracyonresponse-time,accuracyandconfi- N.Y. was supported by a grant from the U.S. National dence in a unidimensional discrimination task. Acta Institutes of Health (P50-MH62196). C.S. was supported by a grant from the Wellcome Trust (WT092646AIA). Psychol. 50, 179–197. (doi:10.1016/0001-6918(82) Thoughtful commentaries from Vincent de Gardelle and 90006-3) twoanonymous reviewersaregratefullyacknowledged. 19 Kepecs,A.,Uchida,N.,Zariwala,H.A.&Mainen,Z.F. 2008 Neural correlates, computation and behavioural impact of decision confidence. Nature 455, 227–231. (doi:10.1038/nature07200) REFERENCES 20 Kiani, R. & Shadlen, M. N. 2009 Representation of 1 Gold, J. I. & Shadlen, M. N. 2007 The neural basis of confidence associated with a decision by neurons in the decision making. Annu. Rev. Neurosci. 30, 535–574. parietal cortex. Science 324, 759–764. (doi:10.1126/ (doi:10.1146/annurev.neuro.29.051605.113038) science.1169405) 2 Bogacz,R.,Brown,E.,Moehlis,J.,Holmes,P.&Cohen, 21 Britten, K. H., Shadlen, M. N., Newsome, W. T. & J. D. 2006 The physics of optimal decision making: a Movshon, J. A. 1992 The analysis of visual motion: a formalanalysisofmodelsofperformanceintwo-alterna- comparison of neuronal and psychophysical perform- tive forced-choice tasks. Psychol. Rev. 113, 700–765. ance. J. Neurosci.12,4745–4765. (doi:10.1037/0033-295X.113.4.700) 22 Roitman,J.D.&Shadlen,M.N.2002Responseofneur- 3 Ratcliff, R. & McKoon, G. 2008 The diffusion decision ons in the lateral intraparietal area during a combined model: theory and data for two-choice decision tasks. visual discrimination reaction time task. J. Neurosci. 22, Neural Comput. 20, 873–922. (doi:10.1162/neco.2008. 9475–9489. 12-06-420) 23 Shadlen, M. N. & Newsome, W. T. 2001 Neural basis 4 Smith, P. L. & Ratcliff, R. 2004 Psychology and neuro- of a perceptual decision in the parietal cortex biology of simple decisions. Trends Neurosci. 27, (area LIP) of the rhesus monkey. J. Neurophysiol. 86, 161–168.(doi:10.1016/j.tins.2004.01.006) 1916–1936. 5 Wald,A.&Wolfowitz,J.1949Bayessolutionsofsequen- 24 Smith, J. D., Couchman, J. J. & Beran, M. J. 2012 The tial decision problems. Proc. Natl Acad. Sci. USA 35, highs and lows of theoretical interpretation in animal- 99–102. (doi:10.1073/pnas.35.2.99) metacognition research. Phil. Trans. R. Soc. B 367, 6 Ratcliff, R. 1978 A theory of memory retrieval. Psychol. 1297–1309.(doi:10.1098/rstb.2011.0366) Rev. 85,59–108. (doi:10.1037/0033-295X.85.2.59) 25 Kepecs,A.&Mainen,Z.F.2012Acomputationalframe- 7 Busemeyer, J. R. & Johnson, J. G. 2004 Computational workforthestudyofconfidenceinhumansandanimals. models of decision making. In Handbook of judgment Phil.Trans.R.Soc.B367,1322–1337.(doi:10.1098/rstb. and decision making (eds. D. Koehler & N. Harvey), pp. 2012.0037) 133–154.Oxford, UK:Blackwell. 26 Resulaj,A.,Kiani,R.,Wolpert,D.M.&Shadlen,M.N. 8 Ratcliff, R. & Rouder, J. N. 1998 Modeling response 2009 Changes of mind in decision-making. Nature 461, times for two-choice decisions. Psychol. Sci. 9, 263–266. (doi:10.1038/nature08275) 347–356.(doi:10.1111/1467-9280.00067) 27 Rabbitt, P. M. A. 1966 Errors and error correction in 9 Petrusic, W. M. & Baranski, J. V. 2003 Judging confi- choice-response tasks. J. Exp. Psychol. 71, 264–272. dence influences decision processing in comparative (doi:10.1037/h0022853) Phil.Trans.R.Soc.B(2012)

See more

The list of books you might like

Upgrade Premium
Most books are stored in the elastic cloud where traffic is expensive. For this reason, we have a limit on daily download.
DMCA & Copyright: Dear all, most of the website is community built, users are uploading hundred of books everyday, which makes really hard for us to identify copyrighted material, please contact us if you want any material removed.
Copyright @ PDFdrive.to, 2022 | We love our users