CoChat
Sign Up
Cited by 307Bayesian Modeling and Causal InferenceDecision-Making and Behavioral EconomicsCognitive Science and Education Research

Bayesian Brains without Probabilities

Open access
Sanborn, Adam N., Chater, Nick·2016·Trends in Cognitive Sciences
DOI 10.1016/j.tics.2016.10.003
Open this paper in CoChat — free accountOpen PDF

Abstract

Bayesian explanations have swept through cognitive science over the past two decades, from intuitive physics and causal learning, to perception, motor control and language. Yet people flounder with even the simplest probability questions. What explains this apparent paradox? How can a supposedly Bayesian brain reason so poorly with probabilities? In this paper, we propose a direct and perhaps unexpected answer: that Bayesian brains need not represent or calculate probabilities at all and are, indeed, poorly adapted to do so. Instead, the brain is a Bayesian sampler. Only with infinite samples does a Bayesian sampler conform to the laws of probability; with finite samples it systematically generates classic probabilistic reasoning errors, including the unpacking effect, base-rate neglect, and the conjunction fallacy.

AI snapshot

Written from the paper’s full text.

TL;DR

Bayesian cognitive models do not require the brain to represent or calculate probabilities; instead, the brain can be a Bayesian sampler. With finite samples, such a sampler systematically produces classic probabilistic reasoning errors, while only infinite samples would conform to the laws of probability.

Key findings

  • Bayesian brains need not represent or calculate probabilities at all; sampling can approximate Bayesian inference using only local relative posterior probabilities, and only with infinite samples does a Bayesian sampler conform to probability laws.
  • Finite sampling generates classic reasoning fallacies, including the unpacking effect, base-rate neglect, and the conjunction fallacy. For example, explicit Bayesian calculation over vocabulary would require posterior probabilities over tens of thousands of words, ideally the 600,000 words in the Oxford English Dictionary.
  • Rich, contextually constrained real-world tasks can be easier for a Bayesian sampler than abstract lab tasks, because context focuses sampling on relevant regions of the probability landscape.
  • Sampling predicts behavioral traces such as stochasticity and autocorrelation; the paper reinterprets anchoring and adjustment as insufficient sampling from a starting point.

Methods

  • Theoretical opinion article; no empirical sample or dataset is reported.
  • The argument is developed conceptually and illustrated with simulations of sampling algorithms, including Metropolis-Hastings and JAGS, in Figures 1 and 2; R code for the plots is included as supplemental material.
  • Main analysis is a conceptual reconciliation of Bayesian cognitive models with probabilistic reasoning errors, using examples such as word-pattern judgments, puzzle-piece conjunctions, and cross-domain probability comparisons.

Limitations

  • Paper-stated: the Outstanding Questions section lists open issues, including whether sampling is sequential or parallel, how sampling is neurally implemented, how autocorrelated samples are integrated, whether samples come from the correct complex distribution or a simplified one, and how sampling handles counterfactuals and causality.
  • Evident from the text: this is an opinion/theoretical paper with no new empirical data or formal model-fitting; the sampling simulations are illustrative rather than quantitative tests of human behavior.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1016/j.tics.2016.10.003

Cites

100 of 100 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 65 cited sources checked · 0 supported · 0 flagged

  • Findingcitation
    § Box 1. Sampling and Autocorrelation

    Participants given clumpier reward distributions in a 2D computer foraging task later behaved as if tuned to clumpier distributions in semantic space in a word-generations task.

    “Participants given 'clumpier' reward distributions in a 2D computer foraging task later behaved as if 'tuned' to clumpier distributions in semantic space in a word-generations task.”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [78][79][80]

      Not checked in the free run.

  • Interpretationreasoning
    § Bayesian Brains without Probabilities

    Using a number of samples much smaller than the number of hypotheses makes the computations feasible.

    “Using a number of samples much smaller than the number of hypotheses makes the computations feasible.”
    Quote in the paperSubstantial claim
  • Interpretationcitation
    § Unavoidable Limitations of Sampling

    The problem of missing peaks applies to any sampling algorithm without prior knowledge of the location of the peaks.

    “it applies to any sampling algorithm without prior knowledge of the location of the peaks [28,30]”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [28] source

      Not checked in the free run.

    • cites [30] source

      Not checked in the free run.

  • Interpretationreasoning
    § Unavoidable Limitations of Sampling

    A Bayesian sampler that misses peaks of probability will produce reasoning fallacies.

    “A Bayesian sampler that misses peaks of probability will produce reasoning fallacies.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: The brain is a Bayesian sampler.
  • Interpretationreasoning
    § Unavoidable Limitations of Sampling

    A Bayesian sampler's progress around a probability landscape is biased by its starting point.

    “The android's progress around the probability landscape is, first and most obviously, biased by its starting point.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: The brain is a Bayesian sampler.
  • Classified as probably not a substantial claim. 31
  • May combine two claims. 22

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Findingown data
    § Box 1. Sampling and Autocorrelation

    Samples produced by both MCMC methods have low autocorrelations for independent unimodal distributions, but for highly correlated variables and bimodal distributions autocorrelations are more prevalent, even when the aggregated samples match the overall distribution.

    “Samples produced by both MCMC methods have low autocorrelations for independent unimodal distributions (top row of Figure 1B), but for highly correlated variables and bimodal distributions autocorrelations are more prevalent, even when the aggregated samples match the overall distribution (middle rows of Figure 1B).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingnone given
    § Box 1. Sampling and Autocorrelation

    When two modes are far apart, these samplers are very unlikely to sample both modes.

    “And of course, when two modes are far apart, these samplers are very unlikely to sample both modes (bottom row of Figure 1B).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingnone given
    § Box 1. Sampling and Autocorrelation

    The extent of autocorrelation depends on both the algorithm and the probability distribution.

    “The extent of autocorrelation depends on both the algorithm and the probability distribution (see Figure 1B).”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingcitation
    § Box 1. Sampling and Autocorrelation

    More recently developed sampling methods such as those employed by Stan can also show sample autocorrelations, but more complex distributions are needed to induce them.

    “More recently developed sampling methods such as those employed by the program Stan can also show sample autocorrelations, but more complex distributions are needed to induce them [94].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [94]

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Findingnone given
    § Outstanding Questions

    Sampling algorithms have difficulties with isolated modes and produce autocorrelations.

    “Sampling Algorithms Have Difficulties with Isolated Modes and Produce Autocorrelations.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data
    § Outstanding Questions

    Not all of the modes are always sampled, even when thousands of samples are drawn.

    “These show that not all of the modes are always sampled, even when thousands of samples are drawn (i.e., in the bottom row).”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Premisecitation
    § TICS 1618 No. of Pages 11

    In the unpacking effect, participants judge a packed description (e.g., 'being a lawyer') to be less probable than an unpacked description (e.g., 'being a tax, corporate, patent, or other type of lawyer').

    “In the unpacking effect, participants judge, say, 'being a lawyer', to be less probable than the 'unpacked' description 'being a tax, corporate, patent, or other type of lawyer' [44][45][46].”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [44] source

      Not checked in the free run.

    • cites [45] source

      Not checked in the free run.

    • cites [46] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given
    § Unavoidable Limitations of Sampling

    Any distribution that can be sampled can be approximated arbitrarily closely in the limit of infinite samples.

    “Any distribution that can be sampled can, in the limit, be approximated”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning
    § Bayesian Brains without Probabilities

    Bayesian brains need not represent or calculate probabilities at all and are poorly adapted to do so.

    “we propose a direct and perhaps unexpected answer: that Bayesian brains need not represent or calculate probabilities at all and are, indeed, poorly adapted to do so.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: The brain is a Bayesian sampler.

  • May combine two claims.

    Interpretationreasoning
    § Unavoidable Limitations of Sampling

    Finite samples from a distribution are imperfect and hence potentially misleading.

    “finite samples will be imperfect and hence potentially misleading.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Premisecitation
    § TICS 1618 No. of Pages 11

    Accuracy is improved in base-rate neglect studies when a causal link is provided between the individual events, and experiencing outcomes together weakens conjunction fallacies.

    “accuracy is improved in base-rate neglect studies when a causal link is provided between the individual events [55,56], and experiencing outcomes together weakens conjunction fallacies [57]”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [55] source

      Not checked in the free run.

    • cites [56] source

      Not checked in the free run.

    • cites [57] source

      Not checked in the free run.

  • Classified as probably not a substantial claim.

    Premisenone given
    § TICS 1618 No. of Pages 11

    The conjunction of the correct locations of all puzzle pieces cannot be more probable than the correct location of a single piece.

    “The conjunction of the correct locations of all the puzzle pieces cannot, of course, be more probable that the correct location of a single piece.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning
    § TICS 1618 No. of Pages 11

    Sampling cannot accurately assess base rates or estimate the probabilities of conjunctions or disjunctions.

    “the inability to bridge between the two phenomena means that sampling cannot accurately assess base rates or estimate the probabilities of conjunctions or disjunctions.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: The brain is a Bayesian sampler.

  • May combine two claims.

    Interpretationreasoning
    § Sampling and Task Richness

    In complex real-world problems, climbing a rugged probability landscape to find a 'good-enough' hypothesis is crucial; linking to numerical probabilities, even approximately, is not.

    “In complex real-world problems, then, climbing a rugged probability landscape to find 'good-enough' hypothesis is crucial; linking to numerical probabilities, even approximately, is not.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: The brain is a Bayesian sampler.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § Sampling and Task Richness

    Bayesian models of cognition have primarily flourished as explanations of the complex dependency of behavior on the environment in domains including intuitive physics, causal learning, perception, motor control, and language.

    “Bayesian models of cognition have primarily flourished as explanations of the complex dependency of behavior on the environment in domains including intuitive physics, causal learning, perception, motor control, and language [1-10]”
    Quote in the paperNot a substantial claim0 of 10 sources checked
    • cites [1] source

      Not checked in the free run.

    • cites [2] source

      Not checked in the free run.

    • cites [3] source

      Not checked in the free run.

    • cites [4] source

      Not checked in the free run.

    • cites [5] source

      Not checked in the free run.

    • cites [6] source

      Not checked in the free run.

    • cites [7] source

      Not checked in the free run.

    • cites [8] source

      Not checked in the free run.

    • cites [9] source

      Not checked in the free run.

    • cites [10] source

      Not checked in the free run.

  • May combine two claims.

    Interpretationreasoning
    § Sampling and Task Richness

    Learning the structure of the world, or interpreting an image or a sentence, involves finding 'good-enough' hypotheses to usefully guide our actions, which can be achieved by local sampling in the probability landscape.

    “the problem of learning the structure of the world, or interpreting an image or a sentence, involves finding 'good-enough' hypotheses to usefully guide our actions, which can be achieved by local sampling in the probability landscape.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: The brain is a Bayesian sampler.

  • May combine two claims.

    Interpretationcitation
    § TICS 1618 No. of Pages 11

    In these cases, participants are left to combine non-normalized sampled probability estimates for each hypothesis manually, and perhaps choose just one or combine them inappropriately.

    “participants are left to combine the non-normalized sampled probability estimates for each hypothesis 'manually', and perhaps choose just one of the probabilities or perhaps combine them inappropriately [52][53][54].”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [52] source

      Not checked in the free run.

    • cites [53] source

      Not checked in the free run.

    • cites [54] source

      Not checked in the free run.

    • The authors state this as an assumption: The brain is a Bayesian sampler.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Interpretationnone given
    § Sampling and Task Richness

    These computational models generally do not involve explicit probability calculations, but apply Bayesian sampling, using methods such as MCMC.

    “these computational models generally do not involve explicit probability calculations, but apply Bayesian sampling, using methods such as MCMC.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: The brain is a Bayesian sampler.

  • May combine two claims.

    Interpretationreasoning
    § Box 1. Sampling and Autocorrelation

    Asking the gestation time of an elephant will bias a person's estimate because they begin with the common reference point of 9 months; the starting point is an anchor and sampling adjusts from that anchor but insufficiently when small samples are drawn.

    “Thus, asking, for example, the gestation time of an elephant will bias a person's estimate because they begin with the common reference point of 9 months: the starting point is an 'anchor' and sampling 'adjusts' from that anchorbut insufficiently when small samples are drawn.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § Box 1. Sampling and Autocorrelation

    Models of free recall such as Search of Associative Memory and SIMPLE use the previously recalled word to cue the next word, and have been productively used to account for the dependencies seen when people attempt to recall a list of words.

    “Models of free recall such as Search of Associative Memory [95] and SIMPLE [96] use the previously recalled word to cue the next word, and has been productively used to account for the dependencies seen when people attempt to recall a list of words.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [95] source

      Not checked in the free run.

    • cites [96] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisereasoning
    § Box 1. Sampling and Autocorrelation

    Because MCMC algorithms look locally when producing the next sample, autocorrelation can result.

    “Because these algorithms look locally when producing the next sample, autocorrelation can result.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § Box 1. Sampling and Autocorrelation

    In decision making, the ballistic accumulator model could be considered an extreme version of autocorrelation in which each internal time step produces the same strength of sampled evidence within a trial.

    “In decision making, the ballistic accumulator model could be considered an extreme version of autocorrelation in which each internal time step produces the same strength of sampled evidence within a trial [97].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [97] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisenone given
    § Box 1. Sampling and Autocorrelation

    The best settings of these tuning parameters are not known in advance, but they can be learned by drawing warm-up samples to get a sense of what the distribution is like.

    “Of course the best settings of these tuning parameters are not known in advance, but they can be learned by drawing 'warm-up' samples to get a sense of what the distribution is like.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisenone given
    § Box 1. Sampling and Autocorrelation

    When sampling from complex distributions, it is often impossible to draw samples directly, and sampling algorithms such as MCMC are used instead.

    “When sampling from complex distributions, it is often impossible to draw samples directly, and sampling algorithms such as MCMC are used instead.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § Box 1. Sampling and Autocorrelation

    Although outnumbered by models of behavior that assume independent sampling, several models of memory and decision making do assume there are autocorrelations in sampling.

    “Although they are outnumbered by models of behavior that assume independent sampling, several models of memory and decision making do assume there are autocorrelations in sampling.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § Box 1. Sampling and Autocorrelation

    Explicit models of autocorrelated sampling have been used to account for perceptual switching times of multistable stimuli and anchoring effects in reasoning tasks.

    “Explicit models of autocorrelated sampling have been used to account for perceptual switching times of multistable stimuli [18] and anchoring effects in reasoning tasks [25].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [18] source

      Not checked in the free run.

    • cites [25]

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Premisenone given
    § Box 1. Sampling and Autocorrelation

    Sampling algorithms often have tuning parameters which are chosen to minimize autocorrelation and to better explore multiple peaks.

    “Sampling algorithms often have tuning parameters which are chosen to minimize autocorrelation and to better explore multiple peaks.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based). May combine two claims.

    Interpretationreasoning
    § Outstanding Questions

    When comparing 'Things fall apart; the centre cannot hold' to the eclipse, the quote may seem more likely as it is common among quotes, yet this neglects the base rates: most websites do not have literary quotations, and there are many chances for an eclipse each year.

    “In particular, when comparing 'Things fall apart; the centre cannot hold' to the eclipse, the quote may seem more likely as it is a common among quotes, yet this neglects the base rates: most websites do not have literary quotations, and there are many chances for an eclipse each year.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim. May combine two claims.

    Interpretationreasoning
    § Concluding Remarks and Future Perspectives

    It is possible that the brain represents probability distributions over small collections of variables or variational approximations to probability distributions, but this would not affect the key argument, which stems from the unavoidable difficulty of finding isolated peaks in, and calculating the volume of, complex probability distributions.

    “Of course, it is possible that the brain represents probability distributions over small collections of variables [90,91], or variational approximations to probability distributions [92], but this would not affect our key argument, which stems from the unavoidable difficulty of finding isolated peaks in, and calculating the volume of, complex probability distributions.”
    Quote in the paperSubstantial claim0 of 3 sources checked
    • cites [90] source

      Not checked in the free run.

    • cites [91] source

      Not checked in the free run.

    • cites [92] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § Concluding Remarks and Future Perspectives

    The sampling viewpoint fits with current thinking in computational neuroscience, where an influential proposal is that transitions between brain states correspond to sampling from a probability distribution or multiple samples are represented simultaneously.

    “The sampling viewpoint also fits with current thinking in computational neuroscience, where an influential proposal is that transitions between brain states correspond to sampling from a probability distribution [26,27,81,82] or multiple samples are represented simultaneously (e.g., using particle filters [83][84][85][86]).”
    Quote in the paperNot a substantial claim0 of 5 sources checked
    • cites [26] source

      Not checked in the free run.

    • cites [27] source

      Not checked in the free run.

    • cites [81]

      Not checked in the free run.

    • cites [82] source

      Not checked in the free run.

    • cites [83][84][85][86]

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § Concluding Remarks and Future Perspectives

    Neural networks dating back to the Boltzmann machine take a sampling approach to probabilistic inference.

    “Moreover, neural networks dating back to the Boltzmann machine [87] take a sampling approach to probabilistic inference.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [87] source

      Not checked in the free run.

  • May combine two claims.

    Premisereasoning
    § Concluding Remarks and Future Perspectives

    As with human performance, deep belief networks readily sample highly complex probability distributions, but can only calculate explicit probabilities with difficulty.

    “As with human performance, these networks readily sample highly complex probability distributions, but can only calculate explicit probabilities with difficulty.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation
    § Concluding Remarks and Future Perspectives

    Deep belief networks have been highly successful in vision, acoustic analysis, language, and other areas of machine learning.

    “For example, consider deep belief networks, which have been highly successful in vision, acoustic analysis, language, and other areas of machine learning [88].”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [88] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given
    § Concluding Remarks and Future Perspectives

    Sampling provides a natural and scalable implementation of Bayesian models.

    “Sampling provides a natural and scalable implementation of Bayesian models.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given
    § Concluding Remarks and Future Perspectives

    The questions easier to answer with sampling are contextually rich, highly complex, real-world questions on which people show surprising competence.

    “and those easier to answer with sampling are contextually rich, and highly complex, real-world questions on which people show surprising competence.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing), S4 (based). May combine two claims.

    Interpretationnone given
    § Outstanding Questions

    Comparing the probability of seeing the two astronomical events in a year, or the probability of the two quotations appearing on a random website, are both relatively easy.

    “Comparing the probability of seeing the two astronomical events in a year, or the probability of the two quotations appearing on a random website, are both relatively easy.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationnone given
    § Outstanding Questions

    Evaluating the probability of a conjunction will be easier.

    “The bottom row illustrates that evaluating the probability of a conjunction will be easier.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning
    § Concluding Remarks and Future Perspectives

    Sampling generates reasoning fallacies, and leaves traces, such as stochasticity and autocorrelations, which are observed in human behavior.

    “Sampling generates reasoning fallacies, and leaves traces, such as stochasticity and autocorrelations, which are observed in human behavior.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S4 (based).

    Interpretationnone given
    § Concluding Remarks and Future Perspectives

    The questions that are difficult to answer with sampling correspond to those with which people struggle.

    “Moreover, the questions that are difficult to answer with sampling correspond to those with which people struggle,”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).

    Interpretationreasoning
    § Outstanding Questions

    Comparing the probability of seeing one of the astronomical events in a year to the probability of seeing one of the quotations on a random website is more difficult.

    “Comparing the probability of seeing one of the astronomical events in a year to the probability of seeing one of the quotations on a random website is more difficult.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation
    § Concluding Remarks and Future Perspectives

    Applications of deep belief networks include generating facial expressions and recognizing objects.

    “Applications of deep belief networks include generating facial expressions and recognizing objects [34,89].”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [34]

      Not checked in the free run.

    • cites [89]

      Not checked in the free run.

41 of 41

  • The authors state this as an assumption.

    The brain is a Bayesian sampler.

    “Instead, the key assumption is that the brain is a Bayesian sampler (see Glossary).”
1 of 1

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • When behavioral data fail to conform to Bayesian optimality predictions, researchers frequently appeal to auxiliary assumptions such as suboptimal priors, cognitive capacity limitations, or approximation algorithms to account for deviations.

    “when behavioral data fail to conform to the predictions of Bayesian optimality, researchers frequently appeal to auxiliary assumptions—such as “suboptimal” priors, cognitive capacity limitations, or the use of approximation algorithms—to account for these deviations (Rahnev and Denison 2018; Sanborn and Chater 2016).”

    From The myth of the Bayesian brain (2025) · cites this paper as Sanborn and Chater 2016

    In “Bayesian Brains without Probabilities” and “Unavoidable Limitations of Sampling” (page unavailable), the paper explains departures from ideal probability judgments through finite sampling and computational limits. It does not say that researchers frequently invoke auxiliary assumptions in general, or identify suboptimal priors as one of those explanations. (LLM, conf=0.83)

    Experimental — accuracy not yet measured

  • Cognitive experiments have revealed signatures of sampling and amortized inference in human behavior.

    “In particular, cognitive experiments have revealed signatures of sampling111 and amortized inference112 in human behavior.”

    From Cognitive computational neuroscience (2018) · cites this paper as [111]

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Two different references are numbered 24

    Minor

    Reference 24 is assigned to two different papers, making citations to [24] ambiguous.

1 of 1

Cites

Flagged first, then by citations

Cited by

Sorted by citations

Related

Authors
Sanborn, A. N., & Chater, N.
Affiliations
University of Warwick (GB)
Corresponding author
Sanborn, Adam N.
ORCID
Sanborn, Adam N.
Venue
Trends in Cognitive Sciences
Publisher
Elsevier BV
Published
2016-12
Volume
20
Issue
12
Pages
883-893
ISSN
1364-6613 (ISSN-L)
DOI
10.1016/j.tics.2016.10.003
OpenAlex
W2541895270
Semantic Scholar
35b39fd2c568d37d1a17f678d6f13646e12e1d6a
Cited by
307
References
100
Citations by year
2026: 16 · 2025: 27 · 2024: 28 · 2023: 60 · 2022: 33 · 2021: 32 · 2020: 34 · 2019: 30 · 2018: 26 · 2017: 20 · 2015: 1
FWCI
4.25
Type
journal article
Language
en
Open access
Licence
Locations
  • Trends in Cognitive Sciences (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • Warwick Research Archive Portal (University of Warwick) (acceptedVersion)
Funders
  • Engineering and Physical Sciences Research Council (EP/K039830/1)
  • Economic and Social Research Council (ES/K002201/1, ES/K004948/1)
Topics
  • Physical Sciences › Computer Science › Artificial Intelligence › Bayesian Modeling and Causal Inference
  • Social Sciences › Decision Sciences › General Decision Sciences › Decision-Making and Behavioral Economics
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Cognitive Science and Education Research
Keywords
Bayesian inference, probability judgment, Bayesian sampling, base rate neglect, conjunction fallacy, finite sample size, cognitive adaptation, causal learning, perception
MeSH terms
Bayes Theorem, Brain, Humans, Probability, Thinking, Cognitive Science

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-09.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-09.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review