1. Author Response:

    Reviewer #2 (Public Review):

    [...] The key analyses focus on a distinction between decision and confidence encoding in the EEG data. The main approach here was to identify trials where computational model and behavioural data diverged - cases where behaviour (either choice or confidence, or both) either matches or mismatches the model on individual trials. By applying decoding techniques the authors were able to identify neural correlates of these suboptimalities. One concern here is that if the behavioural data deviates from a noise-free ideal-observer model, it's not clear what neural correlates of these deviations mean. One interpretation could be that they indicate subjects are using a different model - in which case identifying neural correlates of deviations are less informative. Another interpretation is that they are deviating from the assumed model on a fraction of trials, but if this is the case, analysing these deviations will not be able to identify neural correlates of the latent variables of the (otherwise well-functioning) model. In other words, it is not clear whether these analyses are identifying latent states tracking noise in a confidence representation (confidence in confidence?), latent states underpinning (psychological) confidence, or something else.

    We thank the reviewer for this comment, which shows us where we missed some important details in the previous manuscript. To clarify, we do not compare computational model predictions to behaviour, we compare behaviour to the optimal observer who perfectly encodes the presented orientations and perfectly estimates the decision evidence so as to maximise the probability of making a correct perceptual decision. We define this better now in the first part of the Results (P4, L135).

    The computational model estimates the internal evidence the observer is using to make their decisions, which differs from the optimal evidence. We assume that when the observer makes a response that is different from that predicted by the optimal presented evidence, their internal evidence is more different from optimal than when they make a response that is in agreement with the optimal response. Given that observers’ responses are well predicted by the optimal evidence (further details in response to E3), we can say that the internal evidence the observer is using is some form of approximation of the optimal evidence. It may be that the observer is using an approximation that is different from the one specified by our computational model (although extensive analysis in previous research has suggested this model is a good approximation; Drugowitsch et al., 2016), in this case, responses that differ from the optimal response would still on average be due to evidence that is more different from optimal than the evidence that leads to responses that match the optimal response.

    Our neural analysis aims at identifying whether the neural representation differs from the optimal evidence on trials where the response also differs from the optimal response. The clusters of neural signals we isolate are those where deviations from optimal in the neural representation predict whether the observer will make an optimal response. In other words, the isolated clusters of neural signals follow the observers’ internal evidence L* more closely than the optimal presented evidence L. We make this clearer now at P10L, 302 and on P12, L352.

    This analysis was based on the trial-by-trial level assumption that the internal representation used to accumulate evidence (also known as the ‘decision variable’ in previous work) is further from optimal on trials where the observer does not give the optimal response. Because this evidence is accumulated over several samples, it will deviate more or less from optimal across samples (and trials). We therefore estimate the sample-wise ‘error’ (i.e., difference from the optimal evidence) associated with: 1. the neural representation, and 2. the computational model of behaviour, and we test whether there exists a significant correlation between these two neural and behavioural errors. We explain this section in further detail on P12, L382.

    We also see how the reader may want to understand this in terms of the actual confidence response, i.e. confidence magnitude. We therefore performed a further analysis, using the source localised signals, inspired by the reviewer’s comment. We show that the accumulated evidence reflected in the signals localised to the orbitofrontal cortex predict confidence magnitude in the lead up to and following the perceptual decision response (P13, L638).

    We note here that although the implementation of this analysis is novel, the reasoning behind it is not. Van Bergen et al. (2015), for example, use the variability in the decoded representation of stimulus orientation to index internal uncertainty, and relate this to behavioural biases in orientation estimation (we now reference this in the manuscript, P12, L357). The rationale is the same: the neural representation deviates from the optimal presented evidence in a way that predicts behavioural deviations, therefore, these processes index information important for behaviour.

    In summary, we are tracking the internal evidence on which observers base their confidence reports, assuming this is identifying neural signature associated with the computation of confidence as opposed to correlated with the eventual magnitude (either due to upstream processes such as the presented evidence, or downstream processes such as emotional responses to decision accuracy), and we now show that these signatures of the computation of confidence do indeed reflect the eventual confidence report.

    Read the original source
    Was this evaluation helpful?
  2. Evaluation Summary:

    This paper is of interest to neuroscientists and psychologists working on perceptual decision-making and metacognition. Using a novel task varying the timing of covert decisions, together with sophisticated computational modelling, allowed identifying neural correlates of latent states related to confidence. The conclusions are in line with other papers identifying a dissociation between brain activity supporting performance and confidence, but provide a novel lens through which to understand these differences by focusing on confidence noise. An open issue is how to interpret conclusions about neural correlates of deviations from an ideal-observer model.

    (This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. Reviewer #1 agreed to share their name with the authors.)

    Read the original source
    Was this evaluation helpful?
  3. Reviewer #1 (Public Review):

    In this paper, Balsdon and colleagues test the hypothesis that perceptual decisions and confidence judgments are regulated by distinct (though overlapping) neural processes. Participants were required to make a categorical judgment regarding the orientation of a series of gabor stimuli while EEG data were recorded. In one condition participants made speeded judgments and in the other they were required to withhold their responses until the end of the sequence. The sequence length was varied to be shorter, the same or longer than the average response time of the speeded condition. In the delayed response condition, behavioural modelling indicated that participants tended to commit to their choice before the end of the longer trial sequence consistent with the imposition of a decision bound. In contrast, confidence judgments were better explained by an unbounded model in which evidence accumulation continued after the first-order choice. Neural signal analyses were also interrogated. Decoding analyses highlight earlier motor preparation in the longer condition consistent with premature commitment while regression analyses highlighted a corresponding attenuation of the neural representation of accumulated evidence. Meanwhile a distinct neural representation was associated with variations in the accuracy/optimality of confidence judgments and was found to persist throughout longer trials and was associated with distinct cortical generators.

    This paper covers a very topical issue that would appeal to a wide range of readers and applies a number of very rigorous and sophisticated analyses.

    At the outset the authors suggest that there are two major viewpoints on confidence representations - the first being that it is 'a mere consequence of perceptual processes' and the second being that it recruits specialised metacognitive resources. It is unclear whether this statement is made with reference to mathematical models of the decision process and/or our current understanding of the neural circuitry. In the case of the latter, the authors do not mention the extensive fMRI and lesion work pointing to a dissociation of the two. In the case of the former, the authors highlight the possibility that confidence judgments may be informed by additional sources of information that do not influence the initial perceptual choice. Here they include post-decisional accumulation as an example of 'an additional source of noise that does not influence the perceptual decision' but even 'first-order' models that are devised to jointly account for choice and confidence judgments through a single accumulation process assign a key role to post-decisional accumulation. Some reframing of the introduction would be beneficial to better clarify what the novel contributions of this paper really are with respect to this previous.

    Read the original source
    Was this evaluation helpful?
  4. Reviewer #2 (Public Review):

    This paper uses a clever manipulation of perceptual evidence streams to encourage people to make covert early decisions in a categorisation task (whether a stream of Gabors was tilted clockwise or counterclockwise). Modeling and behavioural data indicate that stimuli arriving after a covert decision continue to affect confidence, offering an opportunity to dissociate neural signatures of evidence accumulation and confidence formation. A "Free response" task allowed inference onto the status of (neural) evidence accumulation in the corresponding fixed-response conditions, via cross-classification analyses of EEG data. The results are consistent with previous findings that confidence is affected by post-decisional evidence accumulation. Sophisticated computational modelling here allowed the disentangling of neural representations of distinct phases of the perceptual decision process, from stimulus encoding to response preparation.

    The key analyses focus on a distinction between decision and confidence encoding in the EEG data. The main approach here was to identify trials where computational model and behavioural data diverged - cases where behaviour (either choice or confidence, or both) either matches or mismatches the model on individual trials. By applying decoding techniques the authors were able to identify neural correlates of these suboptimalities. One concern here is that if the behavioural data deviates from a noise-free ideal-observer model, it's not clear what neural correlates of these deviations mean. One interpretation could be that they indicate subjects are using a different model - in which case identifying neural correlates of deviations are less informative. Another interpretation is that they are deviating from the assumed model on a fraction of trials, but if this is the case, analysing these deviations will not be able to identify neural correlates of the latent variables of the (otherwise well-functioning) model. In other words, it is not clear whether these analyses are identifying latent states tracking noise in a confidence representation (confidence in confidence?), latent states underpinning (psychological) confidence, or something else.

    Understanding this is important to set the current findings in the context of previous work. For instance, classical lesion approaches have also revealed a distinction between confidence and performance in both humans and animals (eg Lak et al., 2014 Neuron). Others have used neural decoding approaches to dissociate confidence and performance-related activity (eg Cortese et al., 2016 Nat Comms). And other work has examined the relationship between post-decisional evidence samples and neural encoding of confidence (eg Murphy et al., 2015 eLife; Fleming et al. 2018 Nat Neuro). A key advantage of the current work in relation to previous studies is the use of sophisticated computational modelling to identify neural correlates of latent model variables. But it's ambiguous as to whether that approach is actually telling us about confidence, given the focus on suboptimalities. The discussion of the paper states clearly that the focus here is on confidence precision not magnitude, but the findings are then interpreted as revealing neural correlates of confidence (eg in the title and abstract, "computation of confidence"). To support this claim (and tackle issues highlighted above) it may be useful to decode confidence magnitude from the same data, and ask whether confidence precision modulates these or different signals.

    Read the original source
    Was this evaluation helpful?