Skip to main content

Levels of Explanation Vindicated


Marr’s celebrated contribution to cognitive science (Marr 1982, chap. 1) was the introduction of (at least) three levels of description/explanation. However, most contemporary research has relegated the distinction between levels to a rather dispensable remark. Ignoring such an important contribution comes at a price, or so we shall argue. In the present paper, first we review Marr’s main points and motivations regarding levels of explanation. Second, we examine two cases in which the distinction between levels has been neglected when considering the structure of mental representations: Cummins et al.’s distinction between structural representation and encodings (Cummins in Journal of Philosophy, 93(12):591–614, 1996; Cummins et al. in Journal of Philosophical Research, 30:405–408, 2001) and Fodor’s account of iconic representation (Fodor 2008). These two cases illustrate the kind of problems in which researchers can find themselves if they overlook distinctions between levels and how easily these problems can be solved when levels are carefully examined. The analysis of these cases allows us to conclude that researchers in the cognitive sciences are well advised to avoid risks of confusion by respecting Marr’s old lesson.

This is a preview of subscription content, access via your institution.


  1. On this assumption, Newell’s knowledge level and Pylyshyn’s semantic level are closely related to (and arguably can be identified with) Marr’s level 1. We thank an anonymous referee for this journal for bringing this point to our attention. Among many others, authors that have provided an intentional/semantic reading of Marr’s topmost level include Bermúdez (1995), Burge (1986), Davies (1991), Kitcher (1988), Morton (1993) and Shapiro (1997).

  2. According to this, cognitive research may, consistently with Marr’s programme, begin with the lowest implementation level. This is so even if it is not possible to make progress when research remains confined to this level; cf. Marr’s analogy: “trying to understand perception by studying only neurons is like trying to understand bird flight by studying only feathers: it just cannot be done” (Marr 1982, p. 27).

  3. Note that, in this context, the specified Marrian function is not merely a function in extension, that is, it not only defines an input/output relation. Systematicity phenomena require also the identification of the (linguistic) structural information that makes an input/output relation a systematic relation. A function that, e.g., goes from the representation of Rab to the representation of Rba would not capture systematicity if Rab and Rba were primitive representations in the system. One such intensional function can be seen as a Marrian function that not only states what the system does but also why the system does it. In Peacockean terms, the function is seen as specifying, at Peacocke’s level 1.5, the information on which the algorithm draws (cf. Peacocke 1986).

  4. This is, on reflection, very striking. The distinctive feature of pictures is that they are structurally homogeneous symbols (see below). But it is very controversial to suppose that homogeneous symbols—that is, symbols with no canonical constituents—can figure in systematicity phenomena, because what systematicity demands, if anything, is that the distinctive (semantic-plus-syntactic) contribution of parts of a symbol can be identified in a large number of other symbols—say, the semantic-plus-syntactic contribution of a found in Fa, Ga, Rab, Rba,…). That is impossible if, as homogeneity demands, symbols do not have distinctive (semantic-plus-syntactic) parts. This puzzle, however, does not affect our present discussion.

  5. Examples of accounts of the structure in visual perception abound in cognitive science. To mention just one, Hummel (2001) offers a synthetic account in terms of static and dynamic binding. In Hummel’s model, only when dynamic binding (which is understood as requiring attention) is involved does explicit representation of (parts of) objects and their spatial relations take place. When the visual information is processed quickly, no explicit representation of the structure is used by the algorithm. The present point is that, which structure (if any) a given representation actually has, is never accounted for in terms of outer reality, but only in terms of inner computational processes. Fodor’s apparent mistake seems thus to be Pylyshyn’s vigorously denounced intentional fallacy, that is, “the fallacy of attributing properties of what is being represented to the representation itself (as if our representation of a red square were itself red and square). Yet so long as we assume that the form of some mental representation must account for the content of the perceptual experience we are inevitably led to postulate a picture-like representation to match a picture-like experience” (Pylyshyn 2007, p. 122).

  6. Not of course in the account of the icon’s structure as a representation of X. The problem is that the specification of the structure of a given representation is something over and above what the representation represents. The point is even clearer in the case of discursive representation. When we say that the discursive representation John loves Mary is constituted by John, loves and Mary it is not on the ground that John loves Mary represents John, the relation of loving and Mary. In fact, that this is exactly what John loves Mary actually does represent is perfectly compatible with this discursive representation having some other constituents; indeed, it is compatible with the discursive representation being primitive.

  7. As an anonymous referee rightly points out, since Bermúdez’ 1995 paper, the debate surrounding nonconceptual content has also taken into account nonconceptual content at a subpersonal level. This fact, however, does not affect the point just made about Fodor’s failure to link subpersonal cases with intentional content of perceptual experience. For one thing, Fodor’s contribution to the debate aims to offer an empirical argument against the a priori approaches of “philosophers in the Sellars tradition” (Fodor 2008, p. 193). That tradition—be it on the conceptualist or on the nonconceptualist side—has invariably been concerned with conceptual/nonconceptual contents of personal perceptual experience. For another thing, the conceptualist/nonconceptualist debate seems far from being settled even if we accept the existence of genuine nonconceptual content at a subpersonal level. This seems to be Bermúdez’ own view, who acknowledges for instance that “it may well turn out that the nonconceptual content of subpersonal information systems is rather different from the nonconceptual content of perception” (Bermúdez 2007, p. 69). Finally, we are here following Toribio (forthcoming), who presents the so-called ‘subpersonal worry’, that is, the worry that Fodor’s line of argument is threatened by considering that iconic representation can only be subpersonal algorithmic representation, and hence, over and above the neo-Fregean conceptualist position focused on the nature of content from the subject’s own point of view. Note also that, under the assumption that Marr’s level 1 is intentional (see section 1 above), the high and personal level of description can be accounted for in terms of Marr’s topmost level (thereby arguably equating in this context Marr’s level 1 with Newell’s knowledge level or Pylyshyn’s semantic level).

  8. It seems clear to us that, even if Fodor is right that some iconic representation is structurally homogeneous, this may be far from being the general case. To mention one well-known example, according to Marr (1982), representations in the earliest stage of visual processing consist of primal sketches which are certainly structurally heterogeneous: they involve the processing of geometrical information and intensity changes in light from the two-dimensional retinal image so as to detect edges, bars, ends and blobs. Therefore, as far as Marr’s classic model is concerned, there are distinctive semantic/syntactic parts in iconic representation right from the start.


  • Bermúdez, J.L. 1995. Nonconceptual content: From perceptual experience to subpersonal computational states. Mind & Language 10(4): 333–369.

    Article  Google Scholar 

  • Bermúdez, J.L. 2007. What is at stake in the debate on nonconceptual content? Philosophical Perspectives 21(1): 55–72.

    Article  Google Scholar 

  • Burge, T. 1986. Individualism and psychology. Philosophical Review 95: 3–45.

    Article  Google Scholar 

  • Cummins, R. 1996. Systematicity. Journal of Philosophy 93(12): 591–614.

    Article  Google Scholar 

  • Cummins, R., J. Blackmon, D. Byrd, P. Poirier, M. Roth, and G. Schwarz. 2001. Systematicity and the cognition of structured domains. Journal of Philosophy 98(4): 167–187.

    Article  Google Scholar 

  • Cummins, R., J. Blackmon, D. Byrd, A. Lee, and M. Roth. 2005. What systematicity isn’t: Reply to Davis. Journal of Philosophical Research 30: 405–408.

    Google Scholar 

  • Davies, M. 1991. Individualism and perceptual content. Mind 100: 461–484.

    Article  Google Scholar 

  • Egan, F. 1996. Intentionality and the theory of vision. In Perception, ed. K. Akins, 232–247. Oxford: Oxford University Press.

    Google Scholar 

  • Fodor, J.A. 2008. LOT2: The language of thought revisited. Oxford: Oxford University Press.

    Google Scholar 

  • Fodor, J.A., and Z. Pylyshyn. 1988. Connectionism and cognitive architecture: A critical analysis. Cognition 28: 3–71.

    Article  Google Scholar 

  • García-Carpintero, M. 1995. The philosophical import of connectionism: A critical notice of Andy Clark’s associative engines. Mind & Language 10(4): 370–401.

    Article  Google Scholar 

  • Hummel, J.E. 2001. Complementary solutions to the binding problem in vision: Implications for shape perception and object recognition. Visual Cognition 8: 489–517.

    Article  Google Scholar 

  • Julesz, B. 1971. Foundations of cyclopean perception. Chicago: Univ. of Chicago Press.

    Google Scholar 

  • Kitcher, P. 1988. Marr’s computational theory of vision. Philosophy of Science 55: 1–24.

    Article  Google Scholar 

  • Marr, D. 1982. Vision. San Francisco: Freeman.

    Google Scholar 

  • Morton, P. 1993. Supervenience and computational explanation in vision theory. Philosophy of Science 60: 86–99.

    Article  Google Scholar 

  • Newell, A. 1986. The symbol level and the knowledge level. In Meaning and cognitive structure, ed. Z.W. Pylyshyn and W. Demopoulos, 31–39. Norwood: Ablex.

    Google Scholar 

  • Peacocke, C. 1986. Explanation in computational psychology: Language, perception and level 1.5. Mind & Language 1: 101–123.

    Article  Google Scholar 

  • Pylyshyn, Z.W. 1984. Computation and cognition. Cambridge: MIT.

    Google Scholar 

  • Pylyshyn, Z.W. 2007. Things and places. Cambridge: MIT.

    Google Scholar 

  • Segal, G. 1989. Seeing what is not there. Philosophical Review 98(2): 189–214.

    Article  Google Scholar 

  • Shapiro, L.A. 1997. A clearer vision. Philosophy of Science 64(1): 131–153.

    Article  Google Scholar 

  • Smolensky, P., G. Legendre, and Y. Miyata. 1992. Principles for an Integrated Connectionist/Symbolic Theory of Higher Cognition. Institute of Cognitive Science, University of Colorado, Technical Report 92–08.

  • Sperling, G. 1960. The information available in brief visual presentations. Psychological Monographs, 74(11): 1–29.

    Google Scholar 

  • Toribio, J. (forthcoming). Compositionality, Iconicity and Perceptual Nonconceptualism. Philosophical Psychology.

Download references


We would like to thank Josefa Toribio, Christopher Evans and two anonymous referees for their helpful comments and suggestions on earlier drafts. This research has been partially funded by the MICINN, Spanish government, under the research project FFI2008-06164-C02-02, the CONSOLIDER INGENIO 2010 Program, grant CSD2009-0056, and the Catalan government, via the consolidated research group GRECC, SGR2009-1528.

Author information

Authors and Affiliations


Corresponding author

Correspondence to Víctor M. Verdejo.

Rights and permissions

Reprints and Permissions

About this article

Cite this article

Verdejo, V.M., Quesada, D. Levels of Explanation Vindicated. Rev.Phil.Psych. 2, 77–88 (2011).

Download citation

  • Published:

  • Issue Date:

  • DOI:


  • Perceptual Experience
  • Linguistic Structure
  • Representational Scheme
  • Iconic Representation
  • Nonconceptual Content