Abstract
Perception purports to help you gain knowledge of the world even if the world is not the way you expected it to be. Perception also purports to be an independent tribunal against which you can test your beliefs. It is natural to think that in order to serve these and other central functions, perceptual representations must not causally depend on your prior beliefs and expectations. In this paper, I clarify and then argue against the natural thought above. All perceptual systems must solve an under-determination problem: the sensory data they receive could be caused by indefinitely many arrangements of distal objects and properties. Using a Bayesian approach to perceptual processing, I argue that in order to solve the under-determination problem, perceptual capacities must rely on prior beliefs or expectations of some kind. I then argue that perceptual states or processes can help ground knowledge of the world whether the ‘beliefs’ necessary for perceptual processing are encoded as (or influenced by) sub-personal states within a perceptual system or cognitive states, such as person-level beliefs. My argument has two main parts. First, I give a preliminary argument that cognitive influence on perception can be appropriate, and I respond to three lines of objection. Second, I argue that cognitively influenced perceptual states can be instances of seeing that p, which makes the relevant states well suited to help ground knowledge that p. I conclude that a cognitively penetrated perceptual state or process can help ground knowledge under some circumstances.
This is a preview of subscription content, log in via an institution to check access.
Similar content being viewed by others
Notes
It should be noted that Siegel does not endorse the claim that cognitive penetration is always epistemically inappropriate. In fact, she explicitly mentions possible cases of epistemically beneficial cognitive penetration with respect to, e.g. evaluative properties.
When I say ‘seem to be corrupted’ here, I mean that the process may seem to the philosopher reflecting on a potentially hypothetical case of cognitive influence. As Siegel (2012) notes, in many cases discussed, the perceiver is herself unaware of the cognitive influence on her own perceptual processing.
I doubt that Fodor would take this line in part because Fodor (1984) is careful to distinguish the claim that perception requires ‘inferential’ processes (a claim he accepts) from the claim that perception is thoroughly penetrated by cognition (a claim he rejects).
Perception scientists and others often use ‘beliefs’ and ‘expectations’ to refer to sub-personal states of a hypothesized perceptual system, systems, or capacities. In that usage, such states are not taken to be paradigmatic, person-level cognitive states. Most of the time when I have person-level cognitive states in mind, I will explicitly indicate that that’s what I mean. So typically, if I use ‘beliefs’ or ‘expectations’ without qualification, I have in mind the disjunction of paradigmatic, person-level, cognitive beliefs and sub-personal beliefs.
Throughout the paper, I will refer to cognition and perception. But the paper’s arguments can be translated into talk that eschews or downplays the cognition/perception distinction, by focusing on the question of whether paradigmatic personal-level beliefs ever have certain kinds of influence on the psychological processes that help yield sense experiences, such as visual experiences. Since nothing in the paper depends on the nature of the cognition/perception distinction (including whether it exists or is important) I will not characterize the distinction.
One could also focus on the practical appropriateness of such influence, e.g. in helping subjects successfully navigate their environments. However, that is not my focus here.
The definition of ‘epistemically appropriate’ is intended to be neutral on a number of controversial issues. I use ‘states or processes’ because what provides justification in various circumstances is controversial. For example, when you see that there’s an apple ahead, some views entail that your visual experience is the source of the relevant justification to believe that there’s an apple ahead. Other views entail that the nature of the process (or process type) that causes your belief is what provides the relevant justification. In this paper, I want to remain neutral on the issue of what exactly provides justification, with the exception that I argue against the restriction that the relevant state or process be free from a certain kind of cognitive influence. For ease of exposition, I will sometimes omit the reference to states or processes. In addition, I say ‘help provide…justification’ and ‘help ground S’s knowledge’ rather than ‘provide justification’ or ‘ground knowledge’, because it is controversial whether such states and processes can provide justification or ground knowledge all on their own, or whether other conditions must be in place. Again, I want to remain neutral on that issue.
I avoid offering a precise characterization of epistemic justification in order to remain neutral on its nature, which is controversial. Roughly, the idea is that one has justification to believe that p in proportion to the strength of reasons one has indicating that p is true.
Here ‘sensory-fixed’ and ‘attention-fixed’ indicate that the states in question can have their influence via a causal chain that does not trace solely through a shift in the sensory data or a shift in where the subject attends. For similar definitions of ‘cognitive penetration’ see Pylyshyn (1999), MacPherson (2012), Siegel (2012), Stokes (2013), and Vance (2014).
This example is drawn from Balcetis and Dunning (2010). Note that Balcetis and Dunning do not control for attention shifts. So their results do not provide clear evidence of cognitive penetration as defined here.
These lists are not intended to be exhaustive. It is controversial whether perceptual experience ever actually represents properties outside the sensory core, such as the property of being an oak leaf. For arguments in favor of visual contents such as those in the example, see Siegel (2006) and Masrour (2011). Nothing in the main texts turns on whether such properties are in fact represented in perceptual experience. My use of ‘sensory core’ follows Masrour (2011).
As another example of the sort of worries in the vicinity, Zeimbekis (2013) worries about the justificatory impact that would result if cognitive penetration occurred. He claims that the existence of cognitive penetration would have “far-reaching consequences: it implies that the way we think of objects determines how we see them, thus threatening the role of perception in justifying beliefs.” Zeimbekis’s phrasing could be interpreted to imply that some cognitive penetration of perception would threaten perception’s ability to provide justification for beliefs in general. That view is not plausible. The more cautious, plausible view in the text entails that cognitive penetration of some token perceptual state s with respect to some core content p threatens s’s ability to provide justification for p.
Throughout I treat only the retinal images as the relevant sensory data. However, the argument is compatible with claims that the sensory data include more. For example, the argument is compatible with embodied cognition approaches to perception. The argument requires only the assumption that the sensory data under-determine which are the distal stimuli, which is true on most accounts of the sensory data.
My use of the term ‘sensory data’ for vision to include only retinal images is not intended to imply any substantive claims. The problem is for the perceptual system to generate a determinate percept given the retinal images and other input information, regardless of what we call the images and information. In particular, my talk of sensory data does not imply the sense data views of perception popular in the early and mid 20th century (e.g. Ayer 1956).
See also Rock (1983) for a non-Bayesian development of a constructivist approach to perception.
Evidence for Bayesian approaches comes from a range of empirical work in perception science. For example, McDermott (2004) provides evidence that it is difficult for visual systems to determine whether there is an edge present using only local sensory data. Tu and Zhu (2002) provide further evidence that regional cues are not sufficient for edge detection. Some of the best evidence that perceptual processes rely on prior assumptions in approximate accordance with Bayesian probability theory comes from work on motion perception (see especially Weiss et al. 2002). See Yuille and Kersten (2006) for further discussion of the reliance on prior assumptions in a Bayesian framework for resolving stimulus ambiguity in the sensory core.
Although priors and likelihoods are sometimes referred to as ‘prior beliefs or knowledge’ this need not imply that the priors or likelihoods are represented as paradigmatic person-level, cognitive states, such as subjects’ beliefs. They may instead be represented as sub-personal states within a perceptual system. Compare Rescorla (forthcoming).
p(D) is the sum of the products of the prior and likelihood over all the hypotheses: p(D|H1*p(H1) + p(D|H2)*p(H2) + … p(D|Hn)*p(Hn) It serves as a normalizing constant, which is required to generate an unconditional posterior probability used in updating the priors via conditionalization. The added step of generating an unconditional posterior via conditionalization will be important in the section on double counting below. However, for the purposes of generating the percept on a given occasion, what’s important is which hypothesis in H has the highest posterior conditional probability given the sensory data D. So, except where the discussion turns to updating the priors, the denominator of Bayes’ theorem can be ignored.
Most models also include a utility function as part of the rule for generating a percept. Depending on the utility function and the perceptual task, the resulting percept need not represent the MAP in every case. These complications do not affect my main argument. In what follows, I set them aside.
For the claim that priors are updated in accord with Bayes, see Shea (2014). Currently, evidence for relevant updating is perhaps not strong (Thanks to David Bennett for discussion of this point). But some such evidence comes from a series of studies by Thorsten Hansen, Karl Gegenfurtner, and colleagues, who found evidence that memory modulates color experience (Hansen et al. 2006; Olkkonen et al. 2008). One of the studies in the series is particularly important for the question of whether the priors are updated over relatively short time-scales (e.g. days, months or years). Witzel et al. (2011) found memory color-modulating effects for human-made artifacts. For example, they found that in some circumstances, artifacts with a characteristic color, e.g. blue Smurfs, appeared bluer than objects with same surface reflectance properties but no characteristic color (and similarly for other artifacts with characteristic color, such as red Coca Cola logos). Witzel et al.’s results can be explained in a Bayesian framework as follows. With repeated exposure to blue Smurfs, the subject (or her perceptual system) updates the priors to increase the probability that Smurfs (or Smurf-shaped objects) are blue. When the subject is later exposed to images of Smurfs under unusual conditions (where the surface reflectance of the Smurf is shifted significantly toward achromaticity), the perceptual system relies more heavily on the prior expectation that Smurfs are blue. As a result of relying more heavily on the priors, the perceptual inference generates a blue-shifted percept relative to that generated under the same viewing conditions for objects with the same surface reflectance properties but no characteristic color. Since the artifacts were recently invented and do not occur in nature, it is implausible that the priors were set via an evolutionarily old process. Rather the priors were likely updated over a relatively short period of time as the subject learned Smurfs’ characteristic color.
The presentation in the text is an oversimplification. In a more accurate presentation, the priors would represent the distribution of some property in the environment, where the hypotheses partition the space. In addition, for ease of presentation I also mostly bracket issues concerning noise in perceptual processing. For an accessible introduction to how noise is accounted for in a Bayesian framework, see Bennett et al. (2014).
The example is based on one from Bennett et al. (2014).
It would be possible for someone to argue here that, on the assumption that priors are neither cognitive states nor influenced by cognitive states, Bayesian perceptual inference and the resulting percepts necessarily provide the relevant threshold justification because (allegedly) inferences involving no influence by cognitive states are not epistemically evaluable in the ways that could prevent them from providing such justification. Please note that that is not the argument I have presented here. My arguments in the main text are neutral on the question of whether states or processes not involving influence by cognitive states are rationally evaluable in various ways. Thanks to Susanna Siegel for pressing me to clarify this point.
Because likelihoods refer in part to the sensory data themselves (e.g. retinal images) it is implausible that likelihoods are stored as cognitive states. However, it is possible that (some) priors could be, since priors denote only states of the world. As noted in the main text, all that is needed for cognitive penetration of perceptual states to occur is that cognitive states have an attention-fixed, sensory-fixed influence on such states. In principle, cognitive states could influence either likelihoods or priors, either of which could constitute cognitive penetration.
For many initially plausible examples of cognitive penetration (e.g. the Necker cube or the duck/rabbit ambiguous figure) it has been argued that variations due to cognitive influence are entirely mediated by shifts in attention or sensory data or both. As a result, the examples do not meet the criteria for cognitive penetration as standardly defined. Some instances are difficult or impossible to explain away by appeal to attention shifts. For example, in a series of recent studies, Thorsten Hansen, Karl Gegenfurtner, and colleagues have found evidence that memory modulates color experience (Hansen et al. 2006; Olkkonen et al. 2008). MacPherson (2012) argues that they present the most likely examples of cognitive penetration. However, Deroy (2013) argues that the influencing states may not be cognitive, but instead may be internal to a perceptual module (perhaps one that straddles different perceptual modalities, but nevertheless does not extend to cognition). Because there are no clear cases of cognitive penetration, the present paper is oriented toward the potential consequences of prior beliefs and expectations (including, but not limited to possible cognitive penetration) with respect to contents in the sensory core. The partial focus on possible cognitive penetration does not rob the paper of theoretical interest. Considerations of realistic but merely possible cases of cognitive influence can still shed light on the roles of perception and whether or not views like Independent Tribunal are true.
Roughly, reliabilism about justification entails that a belief that p is justified to the extent that it is the product of a reliable type of process. For a classic defense of the view see Goldman (1986).
Accuracy and reliability are distinct properties. Accuracy pertains to how closely perceptual representation matches reality on an occasion. Reliability concerns propensity for accuracy or accuracy over time.
The example is due to Ophelia Deroy. Like Fodor, Deroy argues that perceptual processes are largely encapsulated from cognitive influence, and she too thinks that such encapsulation is a good thing. Here is the relevant passage: “[T] he fact that perceptual experiences are determined independently of our background beliefs grounds the idea that these experiences can justify or ground our beliefs. If the reverse was true, that is, if our previous belief that most apples are red partly determines our present seeing an apple being red, then our seeing this apple being red suddenly looks like a bad way to justify or increase the degree of justification in our belief that most apples are red.”(Deroy 2013)
In principle, it is possible for the perceptual system to ignore the sensory data, but such cases are rare if they ever occur. Only if the organism or perceptual system expects the sensory signal to be so noisy as to be uninformative, will the sensory data fail to constrain the perceptual inference.
Conditionalization takes a conditional probability p(H|D) and data D and yields the unconditional probability p(H).
There is experimental evidence that the perceptual system—or nervous system generally—updates likelihoods when engaged in sensorimotor learning (Körding and Wolpert 2004; Seydell et al. 2008). These updates of the likelihoods occur over short time-scales (e.g. the duration of several hundred trials at a sensorimotor task). However, these updates don’t entail double counting according to the objection. Bayesian conditionalization applies only to updates of the priors (not the likelihoods).
Recall that some Bayesian models (e.g. Clark 2013; Hohwy 2013) postulate that there are no boundaries to sensory-fixed, attention-fixed influence on perceptual processing in the sensory core by a subject’s beliefs. On these views, it is possible that some priors are stored as person-level beliefs and will affect processing that generates percepts.
There are numerous conceptions of seeing that p, many of which are compatible with one another. I do not claim that the notion of seeing that p addressed in the main text is the only such notion. I merely aim to lay down the sense of seeing that p on which I will focus. I argue below that seeing that p in the sense I define is connected to the target sense of epistemic appropriateness.
Please note that my usage of ‘seeing that p’ does not imply the disjunctivism about perceptual states defended by Hinton (1967), Snowdon (2005) and others. My argument does not imply--nor does it get any help from—the idea that there is nothing in common between states of seeing that p, hallucinating that p, or having an illusory experience as of p. In fact, as Rescorla (forthcoming) argues, a Bayesian approach to perception does not fit well with disjunctivism. I focus on seeing that p because, as I argue below, seeing that p is in an important sense a successful case of perceptual experience. Evidence that cognitive penetration can result in seeing that p is evidence that cognitive penetration can be epistemically appropriate in the target sense. Thanks to an anonymous referee for suggesting this clarification.
Strictly speaking, the causal condition entails the accuracy condition. I state them as separate conditions for clarity. For discussion of the causal condition, see Ryle (1949), Grice (1961), Strawson (1974), and Goldman (1977). The condition is often motivated by appeal to Grice’s (1961) pillar example. Suppose that there are two pillars, Pillar 1 and Pillar 2. Pillar 1 is twenty feet directly ahead of S. There is a mirror directly ahead of S, and the mirror blocks S’s view of Pillar 1. The mirror also reflects the image of Pillar 2 so that it appears to S as if there is a pillar twenty feet directly ahead. Suppose further that the reflection of Pillar 2 in the mirror appears to S exactly as Pillar 1 would appear if the mirror were removed. Finally, suppose that for any change made to Pillar 2, there would be a corresponding change to Pillar 1. In this case, the accuracy constraint is met for the claim that S sees that there is a pillar twenty feet directly ahead. But intuitively S does not see that there is a pillar twenty feet directly ahead. Nor does S see the pillar that’s ahead or its location. An inference to the best explanation of each of these facts is that meeting the accuracy constraint is not sufficient for seeing that p. A causal constraint is also necessary.
For familiar reasons having to do with unusual counterfactual scenarios, the counterfactual condition is strictly neither necessary nor sufficient for meeting the causal condition. But under most circumstances it is a good test for meeting the causal condition.
See Goldman (1977) for more on the relevant counterfactual relation, though Goldman’s discussion is in terms of seeing an object, not seeing that p.
For a small sample, see Bonjour and Sosa (2003), Feldman and Conee (1985), Huemer (2001), Pryor (2000), Peacocke (2004), Silins (2007), Wright (1985), Goldman (1986), and Steup (2004). Some of these theories are strictly silent on the precise question of whether S’s belief that p is justified at or above the threshold. For example, Pryor’s (2000) dogmatism is only a view about when a subject receives some propositional justification from an experience. But given how Pryor applies the view to various cases, he is committed to saying that S’s belief is justified at or above the threshold. The same goes for the others on the list.
I argue for this point in Vance (2014).
References
Ayer, A.J. 1956. The problem of knowledge. London: Macmillan.
Balcetis, E., and D. Dunning. 2010. Wishful seeing: more desired objects are seen as closer. Psychological Science 21(1): 147–152.
Bennett, D., J. Trommershauser, and L.C.J. van Dam. 2014. Bayesian modeling of perceiving: a guide to basic principles. In Sensory integration and the unity of consciousness, ed. D. Bennett and C. Hill, 3–14. Cambridge: MIT Press.
Bonjour, L., and E. Sosa. 2003. Epistemic justification: internalism vs. Externalism, foundations vs. Virtues. Malden: Blackwell Publishing.
Clark, A. 2013. Expecting the world: perception, prediction, and the origins of human knowledge. Journal of Philosophy 110(9): 469–496.
Deroy, O. 2013. Object-sensitivity versus cognitive penetrability of perception. Philosophical Studies 162(1): 87–107.
Feldman, J. 2014. Bayesian models of perception: a tutorial introduction. In Oxford handbook of perceptual organization, ed. J. Wagemans. Oxford: Oxford University Press.
Feldman, R., and E. Conee. 1985. Evidentialism. Philosophical Studies 48: 15–34.
Fodor, J. 1983. The modularity of mind: an essay on faculty psychology. Cambridge: The MIT Press.
Fodor, J. 1984. Observation reconsidered. Philosophy of Science 23–43.
Goldman, A.I. 1967. A causal theory of knowing. The Journal of Philosophy 64(12): 357–372.
Goldman, A.I. 1977. Perceptual objects. Synthese 35(3): 257–284.
Goldman, A.I. 1986. Epistemology and cognition. Cambridge: Harvard University Press.
Grice, H.P. 1961. Symposium: the causal theory of perception. Proceedings of the Aristotelian Society, Supplementary Volumes 35: 121–168.
Hansen, T., M. Olkkonen, et al. 2006. Memory modulates color appearance. Nature Neuroscience 9: 1367–1368.
Helmholtz, H. 2000/1867. Treatise on physiological optics. Bristol: Thoemmes.
Hinton, J.M. 1967. Visual experiences. Mind 76: 217–227.
Hohwy, J. 2013. The predictive mind. Oxford: Oxford University Press.
Huemer, M. 2001. Skepticism and the veil of perception. Lanham: Rowman & Littlefield.
Körding, K.P., and D.M. Wolpert. 2004. Bayesian integration in sensorimotor learning. Nature 427(6971): 244–247.
MacPherson, F. 2012. Cognitive penetration of colour experience: rethinking the issue in light of an indirect mechanism. Philosophy and Phenomenological Research 84(1): 24–62.
Mamassian, P., M. Landy, et al. 2002. Bayesian modeling of visual perception. In Probabilistic models of the brain, ed. R. Rao, B. Olshausen, and M. Lewicki. Cambridge: MIT Press.
Masrour, F. 2011. Is perceptual phenomenology thin? Philosophy and Phenomenological Research 83(2): 366–397.
McDermott, J. 2004. Psychophysics with junctions in real images. Perception 33: 1101–1127.
Nozick, R. 1981. Philosophical explanations. Oxford: Oxford University Press.
Olkkonen, M., T. Hansen, et al. 2008. Color appearance of familiar objects: effects of object shape, texture, and illumination changes. Journal of Vision 8(5): 1–16.
Peacocke, C. 2004. The realm of reason. Oxford: Oxford University Press.
Pryor, J. 2000. The skeptic and the dogmatist. Noûs 34(4): 517–549.
Pylyshyn, Z.W. 1980. Computation and cognition: Issues in the foundations of cognitive science. Behavioral and Brain Sciences 3(01): 111–132.
Pylyshyn, Z.W. 1999. Is vision continuous with cognition? the case for cognitive impenetrability of visual perception? Behavioral and Brain Sciences 22: 341–423.
Quine, W.V.O. 1980. Two dogmas of empiricism. In From a logical point of view, 20–46. Cambridge: Harvard University Press.
Rock, I. 1983. The logic of perception. Cambridge: MIT Press.
Ryle, G. 1949. The concept of mind. Chicago: University of Chicago Press.
Seydell, A., B.C. McCann, et al. 2008. Learning stochastic reward distributions in a speeded pointing task. Journal of Neuroscience 28: 4356–4367.
Shea, N. 2014. Distinguishing top-down from bottom-up effects. In Perception and its modalities, ed. S. Biggs, M. Matthen, and D. Stokes. Oxford: Oxford University Press.
Siegel, S. 2006. Which properties are represented in perception? In Perceptual experience, ed. T. Szabo Gendler and J. Hawthorne, 481–503. Oxford: Oxford University Press.
Siegel, S. 2012. Cognitive penetrability and perceptual justification. Noûs 46(2): 201–222.
Silins, N. 2007. Basic justification and the Moorean response to the skeptic. Oxford Studies in Epistemology 2(2): 108.
Snowdon, P.F. 2005. The formulation of disjunctivism: a response to Fish. Proceedings of the Aristotelian Society 105: 129–141. New Series.
Sosa, E. 1999. How to defeat opposition to Moore. Philosophical Perspectives 13: 141–54.
Steup, Matthias. 2004. Internalist reliabilism. Philosophical Issues 14: 403–425.
Stokes, D. 2013. Cognitive penetrability of perception. Philosophy Compass 8(7): 626–663.
Strawson, P.F. 1974. Causation in perception. In Freedom and resentment and other essays, 66–84. London: Methuen.
Tu, Z., and S.C. Zhu. 2002. Image segmentation by data-driven Markov chain monte Carlo. IEEE Transactions on Pattern Analysis and Machine Intelligence 24(5): 657–673.
Vance, J. 2014. Emotion and the new epistemic challenge from cognitive penetrability. Philosophical Studies 169(2): 257–283.
Weiss, Y., E.P. Simoncelli, et al. 2002. Motion illusions as optimal percepts. Nature Neuroscience 5(6): 598–604.
Witzel, et al. 2011. Object knowledge modulates colour appearance. i-Perception 2(1): 13–49.
Wright, C. 1985. Facts and certainty. Proceedings of the British Academy 71: 429–72.
Yuille, A., and D. Kersten. 2006. Vision as Bayesian inference: analysis by synthesis? Trends in Cognitive Sciences 10(7): 301–308.
Zeimbekis, J. 2013. Color and cognitive penetrability. Philosophical Studies 165(1): 167–175.
Acknowledgments
Thanks to Brian Ballard, Adam Bendorf, David Bennett, Jeff Downard, Carl Ginet, Eric Mandlebaum, Luke Maring, Jason Matteson, Jessie Munton, George Rudebusch, Nico Silins, Lu Teng, and an anonymous referee for helpful feedback. Thanks also to audiences at Northern Arizona University and Harvard University, where earlier versions of this material were presented. Finally, special thanks to the editors of this special issue, Susanna Siegel and Zoe Jenkin, for extensive, helpful feedback.
Author information
Authors and Affiliations
Corresponding author
Rights and permissions
About this article
Cite this article
Vance, J. Cognitive Penetration and the Tribunal of Experience. Rev.Phil.Psych. 6, 641–663 (2015). https://doi.org/10.1007/s13164-014-0197-0
Published:
Issue Date:
DOI: https://doi.org/10.1007/s13164-014-0197-0