Theoretical issues in multilevel explanation

There are many philosophical issues that arise in thinking about how to integrate explanations at different levels. These issues are particularly pressing within the general field of cognitive science, since the mind can be studied at so many different levels, from the molecular and even cellular structure of individual neurons to psychological and behavioral studies of higher cognitive functions. Philosophers have developed a number of ways of classifying levels and the relations between them. Questions about multilevel explanation can be divided into three sub-questions, each of which has several possible answers:

1. What are the kinds of things that are sorted into levels?

In order to use the term "level" clearly, we must identify what it is that we are sorting into levels. Answers to this question fall into two general categories:

Oppenheim, P., & Putnam, H. (1958)."Unity of Science as a Working Hypothesis." In H. Feigl, M. Scriven, and G. Maxwell (Eds), Concepts, Theories and the Mind-Body Problem, Minnesota Studies in the Philosophy of Science II. Minneapolis: University of Minnesota Press (pp. 3-36).

Oppenheim and Putnam offer one of the classic discussions of levels of organization in philosophy of science. They presume that there is a neat correspondence between levels of science and levels of nature. They identify six basic levels of nature and argue that there is a unique theoretical vocabulary and set of scientific laws that correspond to each.

Craver, C.F. (2007). "A field guide to levels." In Explaining the Brain: Mechanisms and the Mosaic Unity of Neuroscience (chapter 5). Oxford: Oxford University Press.

The term "level" is used in a myriad of different ways throughout the philosophy of science. Craver offers a survey of these ways of discussing levels, and sorts them into a helpful taxonomy. He goes on to defend a particular understanding of levels (levels of mechanisms) for explanations in neuroscience

a. Levels of science

To say that levels of explanation refer to levels of science is to take the basic relata to be either products or units of science. Products of science can be theories, models, or concepts. Units of science are things like fields of science and research programs.

Darden, L.

(1991). Theory Change in Science: Strategies from Mendelian Genetics. Oxford: Oxford University Press.

(2005). "Relations among fields: Mendelian, cytological, and molecular mechanisms." Studies in History and Philosophy of Biological and Biomedical Sciences, 36, 349-371.

Darden's work offers an example of thinking about levels of science in terms of fields of scientific study. Studies of genetics, for example, are sorted into levels based on the field of study from which the genes are examined. Mendelian, cytological, and molecular approaches thus represent different levels of genetic science. Much of Darden's own work focuses on specifying how these various levels of genetics relate to one another.

Kuhn, T. (1962). Structure of Scientific Revolutions. Chicago: University of Chicago Press.

Kuhn distinguished levels of science in terms of paradigms. Paradigms of science are the set of received beliefs operative in science at a particular time. Thus Kuhn distinguished levels of science temporally. Each scientific paradigm involves its own set of rules, methods, standards for success, and foundational principles. Scientific revolution occurs when one paradigm of science is replaced by another.

Wimsatt, W. (1994). "The ontology of complex systems: Levels, perspectives, and causal thickets." Canadian Journal of Philosophy (Supp), 20, 207-74.

Wimsatt has distinguished levels of science in terms of scientific perspectives. For Wimsatt, a perspective is set of techniques used in a particular approach to scientific question. Each way of investigating a scientific phenomena, such as memory, corresponds to a distinct level of science.

Lakatos, I. (1977). The Methodology of Scientific Research Programmes: Philosophical Papers, Vol. 1. Cambridge: Cambridge University Press.

Lakatos divides levels of science in terms of research programs. A Lakatosian research program is a set of theories and experimental techniques that a scientific field develops over time. In each of these there will be central set of ideas which form the "hard core" of the program. Around the core there are various auxiliary hypotheses. Different levels of science may share some auxiliary elements, but they will be distinguished from one another by their core beliefs and approaches.

b. Levels of Nature

Levels of explanation can also refer directly to the objects in the world that are studied by science. Items in the world include entities, properties of entities, and activities involving entities. Objects of nature are often sorted into levels on the basis of their size or ability to causally interact with one another.

Churchland, P., & Sejnowski, T. (1992). The Computational Brain. Cambridge, MA: MIT Press.

Churchland and Sejnowski distinguish levels of nature in terms of levels of processing. In such a classification levels are temporally distinguished: they include stages of a task or process. For example, in visual processing, different areas of the brain are divided into different levels according to their place in the temporal sequence. Processing in the lateral geniculate nucleus of the thalamus is considered "low level" because it occurs early on in visual processing, and the V2 area of cortex is considered "high level" because it is one of the later places that visual information is processed.

Grush, R. (2004). "The emulation theory of representation: motor control, imagery, and perception." Behavioral and Brain Sciences, 27, 377-396.

Grush offers a discussion of theories of motor control and visual imagery in which he describes levels of nature as levels of control. This way of characterizing the objects sorted into levels is similar to levels of processing, only here the levels are individuated by relations of subordination not chronology. For example, Grush explains how control theory offers a distinction between "controllers" and "plants" that can be seen as dividing levels of nature in terms of which levels regulate others.

For discussion of levels of control between different areas of cortex, see: Posner and DiGirolamo (1998). "Executive attention: Conflict, target detection, and cognitive control," In R. Parasuraman (ed), The Attentive Brain. Cambridge, MA: MIT Press (pp. 401-423).

Churchland, P., & Sejnowski, T. (2000). "Perspectives on neuroscience" in M.S. Gazzaniga (ed), Cognitive Neuroscience. Oxford: Blackwell (pp. 14-24).

Churchland and Sejnowski distinguish levels of neuroscience in terms of the size of the entities involved. These levels range from molecules to neurons to systems to the entire central nervous system. Levels can be individuated by the units in which the objects at a particular level are measured. For example, molecules are measured in Angstrom units whereas neural networks are measured in millimeters.

Wimsatt, W. (1976). "Reductionism, levels of organization, and the mind-body problem," In G. Globus, I. Savodnik, and G. Maxwell (eds), Consciousness and the Brain. New York: Plenum Press (pp. 199-267).

Here Wimsatt offers a discussion of levels of nature in terms of causal interactions among the units. That is, two objects exist at the same level if they interact causally. This way of characterizing levels will often conform to levels of size, but it need not. Objects in science are thus divided by the regularities of interaction between them, as objects of a particular level are those acted on by the same forces.

2. How might the levels be classified?

Philosophers and scientists have offered a variety of distinct ways of characterizing the levels of explanation involved in describing scientific phenomena. Specific classification issues arise when characterizing cognitive phenomena, as there will be explanations in terms of rationality and intelligence that may not be relevant at the lower levels. Three general approaches to classifying levels:

a. Marr's Tri-level Theory of Explanation

David Marr was a neuroscientist whose book on vision (published posthumously in 1982) is credited with pioneering the field of computational neuroscience and establishing a tripartite model of explanation that has come to serve as a central methodology for cognitive science.Marr argued that explanation of any cognitive phenomenon must include answers to questions formulated at three different levels:

  • Computational level: formulation of the computational task at hand. What is the task? Why is it done?
  • Algorithmic level: How is the task carried out? Which algorithms characterize the transformations involved? What is the nature of the representations involved?
  • Implementation level: How are the representations and processes realized in a physical system?

Additionally, Marr argued that relations between the levels should proceed in a top-down fashion, with questions at the computational level being settled first and then used to aid in answering questions at the lower levels. A complete explanation is achieved once answers have been given to the questions at each level. See Marr (1977) for a defense of the claim that explanations of cognitive phenomena must involve multiple levels.

Marr, D. (1982). "The philosophy and the approach." In Vision: A Computational Investigation into the Human Representation and Processing of Visual Information (chapter 1). San Francisco, CA: W.H. Freeman.

Marr, D. (1977). "Artificial Intelligence: A Personal View." Artificial Intelligence, 9, 37-48.

Dawson, M.R.W. (1998). Understanding Cognitive Science. Oxford: Blackwell.

Dawson provides a cognitive science textbook that is organized in accordance with Marr's view of explanation, which Dawson terms Marr's tri-level hypothesis. Specifically, Dawson argues that Marr's requirement that explanation involve answers at three different levels, with their distinct projects and vocabularies, requires that cognitive science be an interdisciplinary field.

Kitcher, P. (1988). "Marr's computational theory of vision." Philosophy of Science, 88, 1-24.

Kitcher provides an overview of Marr's theory of vision and its importance for philosophy of psychology. She identifies three important relations between Marr's work and various philosophies of psychology: 1) the contrast between Marr's commitment to multiple levels of explanation and the trend toward exclusively neurophysiological accounts of cognitive posited by Eliminativists; 2) Marr's violation of Fodor's methodological solipsism in favor of the inclusion of factors external to the mind in explanations of cognitive phenomena (Fodor, 1980); and 3) Marr's support of a modularist view of perceptual processes versus the claim that vision is a belief-driven process.

Fodor, J. (1980). "Methodological solipsism as a research strategy in cognitive psychology." In Representations. Cambridge, MA: MIT Press.

Sterelny, K. (1990). "Marr on vision; Fodor on the mind's organization: The theory in action." The Representational Theory of Mind (chapter 4).

Sterelny offers a review of Marr's work and situates Marr's view within the framework of philosophy of psychology/philosophy of mind. Marr's theory of vision is shown to be an explication of a computational theory of mind, and Sterelny offers further discussion of Marr's claims about modularity, nativism, and realizability. The chapter concludes with a discussion of the influence of Marr's theory in Fodor's The Modularity of Mind.

Fodor, J. (1983). The Modularity of Mind. Cambridge, MA: MIT Press.

Gardner, H. (1985). The mind's new science: A history of the cognitive revolution. New York: Basic Books.

Gardner presents a discussion of Marr's theory of vision at a level of abstraction from the technical details of the visual system. Gardner provides a key discussion of the contrasts between Marr's approach to vision and the noncomputational approach of J.J. Gibson.

Peacocke, C. (1986). "Explanation in computational psychology: Language, perception, and level 1.5" Mind and Language, 1, 101-123.

Peacocke argues for the need of an additional level in Marr's hierarchy of cognitive explanations. Peacocke calls this 'level 1.5,' as it exists between the computational and algorithmic levels. Level 1.5 characterizes the information over which the computational function is specified and from which the algorithm draws. There is a need for this additional level, he claims, because the same function can be computed across many different bodies of information and, similarly, two different algorithms can draw on the same body of information.

Horgan, T., & Tienson, J. (1994). "A nonclassical framework for cognitive science, Synthese, 101, 305-345.

Horgan and Tienson contrast Marr's tri-level view with connectionist approaches to cognitive science. They detail various ways in which connectionist and other approaches to cognitive science deviate from the classical methodology (as taken from Marr), concluding with explication of the view they favor, which rejects the assumption that cognitive abilities are tractably computable functions.

Egan, F. (1995). "Computation and content." Philosophical Review, 104, 181-203.

Egan offers a characterization of the top level of Marr's hierarchy: the computational level. She rejects accounts of Marr's theory that construe the theory as intentional (such as Burge, 1986). Instead, she claims the computational level of Marr's theory is intended to be purely mathematical. This view has implications for how the lower levels of Marr's theory – particularly the algorithmic level – should be understood.

Burge, T. (1986). "Individualism and psychology." Philosophical Review, 95, 3-45.

Newell, A. (1982). "The knowledge level." Artificial Intelligence, 18, 87-127.

Newell defends a multilevel view of explanation in cognitive science that bears close resemblance to Marr's. Specifically, Newell's knowledge level closely corresponds with Marr's computational level. For a more complete explication of Newell's multilevel view, see: Newell, A. (1990). Unified Theories of Cognition. Cambridge, MA: Harvard University Press.

Franks, B. (1995). "On Explanation in the Cognitive Sciences: Competence, Idealization, and the Failure of the Classical Cascade." British Journal for the Philosophy of Science, 46, 475-502.

Franks argues that Marr-style explanation in cognitive science – sometimes termed the "Classical cascade" – fails to provide adequate explanations of the cognitive phenomena in question. This failure is illuminated through an analogy between the competence/performance distinction in Chomskian linguistics and the distinction between the computational and algorithmic levels in Marr's hierarchy. He concludes that the methodology provided by Marr fails to establish a sound methodology for cognitive science. Patterson (1998) offers a rebuttal of Franks' intended analogy. See also a reply by Franks (1999) arguing that even without the analogy, the consequences for cognitive science still hold.

Patterson, S. (1998). "Competence and the classical cascade: A reply to Franks." British Journal for the Philosophy of Science, 49, 625-636.

Franks, B. (1999). "Idealizations, competence, and explanation: A response to Patterson." British Journal for the Philosophy of Science, 50, 735-746.

b. The Intentional Stance

Dennett proposed a tri-level theory of explanation that bears a resemblance to Marr's. The key difference comes in the formulation of the top level of explanation. Dennett characterized explanations at this level as involving the "intentional stance" where the cognitive system is approached as if it were a rational agent.

Dennett, D.

(1971). "Intentional systems." Journal of Philosophy, 68, 87-106.

(1981). "True Believers: The Intentional Strategy and Why it works." In A.F. Heath (Ed) Scientific Explanation, Oxford: Oxford University Press, reprinted in W. Lycan (ed), Mind and Cognition, 2nd edition, Oxford: Blackwell, 1999.

(1987). The Intentional Stance. Cambridge, MA: MIT Press.

Dennett proposed a tri-level theory of explanation similar to Marr's. Dennett characterizes the three levels in terms of stances taken toward the explanandum. Dennett's theory includes a physical stance and a design stance, roughly analogous to Marr's implementation and algorithmic levels. The top level of Dennett's explanatory hierarchy is the intentional stance. From this stance, the cognitive system is approached as if it were a rational agent: its behaviors are characterized in terms of the agent's beliefs and desires. This top level differs from Marr's in many respects. First, Dennett claims that the characterizations at this level will be vague, rough-and-ready approximations, as opposed to the precise computational questions required by Marr. Second, Dennett supports an instrumentalist view of the generalizations made through the intentional stance, in contrast to Marr's realism about computational level explanations. According to Dennett, the generalizations afforded by the intentional stance will not be literally true, although they are indispensable for describing human behavior.

Stich, S. (1981). "Dennett on intentional systems." Philosophical Topics, 12, 39-62.

Stich offers a critique of Dennett's intentional stance. Stich details both a hard and soft line view of Dennett's position on intentional systems, neither of which – he claims – offers a tenable view of our practices of intentional attribution.

Baker, L. (1989). "Instrumental intentionality." Philosophy of Science, 56, 303-316.

Baker argues that Dennett's treatment of intentional explanations as instrumental is an untenable approach to characterizing intentional states in physicalist terms. The difficulty comes from tensions in Dennett's own view, between the structure of the theory of intentional systems, and the characterization of key terms.

Dennett, D. (1991). Real Patterns. Journal of Philosophy, 88, 27-51.

Dennett strengthens the realism of his intentional level through a discussion of how the design and intentional levels relate to one another. The psychological attributions made through the intentional stance are not merely instrumental. Instead, they are a real part of the system, insofar as they are a pattern than emerges from simpler interactions at the design level. Dennett makes heavy use of an analogy with Conway's Game of Life (see Holland 1998 for further discussion of the Game of Life).

Holland, J. (1998). Emergence: From chaos to order. Reading, MA: Addison-Wesley.

Heitner, R. (2000). "Is design relative or real? Dennett on intentional relativism and physical realism." Minds and Machines, 10, 267-283.

Heitner argues that Dennett's attempt to provide physical realism through his intentional instrumentalism fails because of conflicts within his own theory. Specifically, Dennett's claim that intentional states emerge as real patterns from physical systems does not work because he offers conflicting arguments on the status of design. Heitner concludes that the relation between the design and intentional levels can only be achieved by intentional realism.

Ratcliffe, M. (2001). "A Kantian Stance on the Intentional Stance." Biology and Philosophy, 16, 29-52.

Ratcliffe argues that the relation between Dennett's levels of intention and design is best understood as an attempt to naturalize intentionality. Ratcliffe goes on to make an analogy between this approach and Kant's in The Critique of Judgment, which is meant to show the failure of Dennett's attempt.

c. The Personal/Subpersonal distinction

One of the central distinctions in philosophy of psychology is the distinction between personal and subpersonal levels of explanation. Subpersonal explanations of cognitive phenomena are given in terms of the parts of the organism that carry out the given behavior. Personal level explanations, in contrast, make reference to the whole person, explaining behaviors in terms of the intentional actions and mental phenomena of the agent involved.

Bermúdez, J. (2005). Philosophy of Psychology. New York: Routledge.

Bermúdez's introduction to the philosophy of psychology takes the relation between personal and subpersonal levels of explanation to be the central question for this area of inquiry. He terms the question of relating the personal and subpersonal "the interface problem" and distinguishes views in the philosophy of psychology by the way in which they respond to this question.

Dennett, D. (1969). "Personal and subpersonal levels of explanation." Content and Consciousness. New York: Routledge.

In this work, Dennett introduces a distinction between personal and subpersonal levels of explanation. Subpersonal explanations of cognitive phenomena are often given in terms of the parts of the organism that carry out the given behavior. These explanations are mechanical, typically referencing the brain and the networks through which it interacts with the rest of the body. Personal level explanations, in contrast, make reference to the whole person, explaining behaviors in terms of the intentional actions and mental phenomena of the agent involved. For organisms with the ability for rational or intentional action, the personal level is important and it seems fundamentally different than lower levels of explanation. Characterizing this level and how it relates to subpersonal levels of explanation is one of the central questions of philosophy of psychology.

Pylyshyn, Z. (1984). "Computation and Cognition: Towards a Foundation for Cognitive Science. Cambridge, MA: Harvard University Press.

Here Pylyshyn offers an instance of a common argument for the indispensability of personal level explanations. The argument is that personal level explanations of behavior are required because there is no way to pick out all and only the lower level descriptions which will be relevant to explaining personal level behavior in a particular situation. The description at the personal level can be realized in many different ways. Pylyshyn gives the example a person viewing a car crash. There is a general expectation of that people who are among the first on the scene of an accident will engage in helping behavior. This generalization might be relevant to predicting how a given person will behave in such a situation. But there will be no one particular low level realization of this behavior that could provide a complete explanation of the person's helping. The actual physical act of 'helping' can be carried out in many different ways, using many different muscle systems, sequences of behavior, etc. Thus, there are certain generalizations that can only be captured with personal, or high level, explanations.

For similar arguments on the indispensability of the personal level, see:
Putnam, H. (1960). "Minds and machines." In S. Hook (ed) Dimensions of Mind. New York: New York University Press.
Fodor, J. (1975). The Language of Thought. New York: Crowell.

Stich, S. (1978). "Belief and subdoxastic states," Philosophy of Science, 45.

In this article, Stich offers a way of characterizing the distinction between personal and subpersonal level explanations. In so doing, he argues for a principled distinction between the two. Personal level explanations – explanations involving what Stich terms "doxastic states" – are those that are accessible to consciousness and capable of inferential integration with other doxastic states. Subdoxastic states – those existing at the subpersonal levels lack such accessibility and integration.

Bermúdez, J. (2000). Personal and subpersonal: A Difference without a Distinction." Philosophical Explorations, 3, 63 – 82.

In this paper Bermúdez argues that the difference between personal and subpersonal explanation is not strong enough to maintain the autonomy of personal level explanations. Personal level explanations are quite often necessary for explaining behavior, but that does not warrant treatment of the personal level as a distinct, independent domain of explanation. Bermúdez argues that the autonomy of the personal level is compromised in two ways. First, subpersonal explanations will often play a critical role in explaining personal level behavior. And second, explanations at the personal level cannot always be understood without the incorporation of information from subpersonal levels.

Hornsby, J. (2000). "Personal and Sub-personal: A Defence of Dennett's Early Distinction." Philosophical Explorations, 3, 6-24.

Hornsby's paper offers an explication of Dennett's personal/subpersonal distinction and how it changed after his establishment of the intentional stance. Initially, the personal level was meant to be entirely autonomous from subpersonal levels of explanation: the requirements of explaining persons were argued to be incommensurable with talk in terms of their realization in the lower levels. This way of understanding the personal level fell away as Dennett worked to explicate the intentional stance and its to the design and physical levels. Hornsby concludes that a return to Dennett's early formulation of the distinction results in a coherent and plausible anti-physicalist anti-dualist philosophy of mind.

Hornsby, J. (1986). "Physicalist thinking and conceptions of behavior." In P. Pettit and J. McDowell (Eds), Subject, Thought, and Context. Oxford: Oxford University Press.

Hornsby characterizes the distinction between personal and subpersonal levels of explanation as a difference between explananda. That is, the two types of explanation attempt to explain different things, instead of two differing explanations of the same thing. Hornsby argues that, despite the brute similarity between the explananda of personal and subpersonal explanations, personal explanations explain actions whereas subpersonal explanations explain movement. Recognition of this distinction is meant to provide grounds for maintaining both forms of explanation.

3. How do the levels relate to each other?

Once various levels of explanation have been identified we need to explore the relations between those levels. How do the levels fit together? To answer this question we must first decide what type of relation we are looking for between the two levels. Two basic types of relation will be addressed here:

a. Metaphysical Relations

To ask about the metaphysical relation between levels is to wonder about the nature of the relation between the things at each level. Things at lower levels can be thought to either realize or constitute higher-level phenomena.

i. Role-Realizer Relations

These are metaphysical relations of realization or implementation. Views in this area are broadly functionalist, making a distinction between the functional role and what occupies or instantiates that role. In philosophy of mind, the distinction is made between the roles played by psychological states and the brain states that realize them. A common example of this relation is that between pain and c-fibers. A simplified example, to be sure, but we can think of pain as a psychological state and c-fiber firing as the physical realization of that psychological pain.

Putnam, H. (1967). "Psychological Predicates." Reprinted as "The Nature of Mental States" in N. Block (ed) Readings in Philosophy of Psychology, vol. 1, Cambridge, MA: Harvard University Press (pp. 223-231).

Putnam frames a functionalist view of the mind through an explication of the functional states in probabilistic automata. The role-realizer relation is meant to be analogous to the software-hardware distinction in computer science. The program states of a machine are implanted by – but not identical to – the states of the physical device. In much the same way, mental states are (in some sense) brain states, but they are not identical to brain states. The relation between them is one of realization.

Fodor, J. (1968). Psychological Explanation. New York: Random House.

Fodor endorses a view of functionalism that is focused on the style of explanation found in cognitive psychology. Brain states are individuated by the role(s) they play in the cognitive system of the organism. This allows states that otherwise look very different to be considered the same (insofar as they fulfill the same function). This marks the emergence of psychological functionalism, insofar as it bases its individuation of the functional roles of the cognitive system on the explanations given by the best psychological (scientific) explanations available.

Lycan, W. (1987). "Continuity of the levels of nature." Reprinted in W. Lycan (Ed) Mind and Cognition, Oxford: Blackwell (pp. 49-63).

Lycan supports a version of functionalism often known as homunculuar functionalism. His differs from other functionalist views in that he rejects the idea of a two-part role-realizer relation. Instead, Lycan argues for a series of functional levels. The assignment of a state to the category role or realizer is relative: a given state may be a realizer for a higher state, but also be playing a functional role that is itself realized in lower states. This decomposition is meant to bottom out in mechanisms that are playing no further functional role. Lycan's view is also referred to as teleological functionalism, as the functions in question will be evolutionarily specified.

Cummins, R. (2000). "'How does it work?' versus 'What are the laws?'" In F. Keil and R. Wilson (Eds) Explanation and Cognition. Cambridge, MA: MIT Press.

Cummins defends a view of homuncular functionalism that is similar to Lycan's. Cummins motivation for the view is providing an adequate description of current psychological practice. Cummins claims that the project of psychology is decomposing our general psychological abilities into smaller, less sophisticated capacities. The job of psychology, he claims, is to perform such functional analysis, not to discover laws.

ii. Mereological Relations

Mereological Relations are part-whole relations. They occur when a combination of things at one level gives rise to a larger thing at a higher level. How exactly the parts are related to one another can differ across types of mereology. The parts can be merely aggregative, or the parts can be structured in their relations to one another.

Bechtel, W. & Richardson, R. (1993). Discovering Complexity: Decomposition and Localization as Strategies in Scientific Research. Princeton: Princeton University Press.

Bechtel and Richardson make a distinction between decomposition and localization that is helpful for thinking about the various ways in which wholes can be divided into parts. Decomposition, on their view, is the differentiation of a system into various components. Localization, on the other hand, assigns the tasks individuated by the decomposition to various parts of the system. Decomposition is a functional relationship of parts to wholes; localization is a relationship of spatial containment.

Glennan, S. (2002). "Rethinking mechanistic explanation." Philosophy of Science, 69 (supp), S342-S353.

Glennan defends a view of mechanisms where mechanisms are understood as coherent organizations of parts. On his view a mechanism is a complex system whose parts comprise the mechanism insofar as they are organized together in such ways that they produce the activities characteristic of that mechanism.

Wimsatt, W. (1986). "Forms of aggregativity," In A. Donagan, A Perovich & M. Wedin (eds) Human Nature and Natural Knowledge. Boston: D. Reidel (pp. 259-291).

In this article, Wimsatt offers a discussion of the ways in which parts can combine to form aggregates. According to Wimsatt, aggregation occurs when the whole is a simple sum of the parts: the same whole would arise if the parts were disaggregated and reaggregated over and over again. This type of aggregation pays no attention to activities that might occur between parts in ways that are dependent upon their structure or organization.

Kauffman, S. (1971). "Articulation of parts explanation in biology and the rational search for them." In R. Buck and R. Cohen (eds), PSA 1970, Dordrecht: Reidel.

Kauffman provides an explanation of why decomposition of a system into parts must respect the nature of the phenomenon to be explained. That is, there is no decomposition simpliciter, but only decomposition relative to the behavior/phenomenon/activity one is trying to explain. In this way, decomposition of a system into parts travels top-down. What counts as an interesting division of parts at the lower level is dependent upon formulation of the system at the higher level.

b. Explanatory Relations

Explanatory relations between levels of explanation ask about the impact of explanations given at one level for the explanations at other levels. These explanations can be independent of one another, mutually exclusive, or mutually allowable.

i. Autonomy

Views about the autonomy of various levels of explanation presume that the explanations given at one level do not influence the explanations given at another. On such views there is a principled distinction to be drawn between (at least some) levels of explanation. In philosophy of psychology, this line of distinction is typically made between the personal and subpersonal levels of explanation.

Davidson, D. (1980). "Mental Events." In Essays on Actions and Events, Oxford: Oxford University Press.

Davidson defends a view in the philosophy of psychology known as anomalous monism. The view involves two basic claims. First, Davidson argues that all mental states are identical to physical states. Herein lies the monism. Second, Davidson claims that all laws exist at the physical level – there are no laws that exist at the level of mental phenomena. The separation of physical and mental explanations is what makes Davidson's monism anomalous. The mental is governed by principles of rationality and coherence, which have no analogue at the lower levels. Thus, personal level explanations (given at the mental level) are autonomous from those given at lower levels.

McDowell, J. (1994). "The Content of Perceptual Experience." Philosophical Quarterly, 44, 190-205.

McDowell offers a different defense of the autonomy of personal level explanations. He argues that the projects of personal and subpersonal explanation (folk psychology and cognitive science, respectively) are fundamentally different and should be understood at distinct levels of explanation. In defense of this view, McDowell formulates the difference between the personal and subpersonal levels of explanation as one between syntax and semantics. He makes a distinction between what enables and what constitutes the personal level. The subpersonal (syntactic) level provides the structure needed to give rise to the semantic level, but it does not involve any form of content. Thus, McDowell criticizes Dennett's caharacterization of subpersonal states as involving content. For a reply, see: Bermúdez (1995). "Syntax, semantics, and levels of explanation." Philosophical Quarterly, 45, 361-367.

Dennett, D. (1969). "Personal and subpersonal levels of explanation." Content and Consciousness. New York: Routledge.

Dennett's distinction between personal and subpersonal explanations, at least in this early formulation, is meant to show the autonomy of personal level explanations from the lower levels. Subpersonal explanations of cognitive phenomena are often given in terms of the parts of the organism that carry out the given behavior. These explanations are mechanical, typically referencing the brain and the networks through which it interacts with the rest of the body. Personal level explanations, in contrast, make reference to the whole person, explaining behaviors in terms of the intentional actions and mental phenomena of the agent involved. For organisms with the ability for rational or intentional action, the personal level is important and is fundamentally different than (and autonomous from) lower levels of explanation.

ii. Multi-level/Interlevel

Many philosophers and psychologists think that explanations should be given at many different levels. This claim encompasses a wide variety of disparate views. There are some who think that there will be different levels of explanation required for giving explanations of different phenomena, and there are others who think that multiple levels of explanation will be relevant to one singular phenomenon.

Craver, C. (2007). "Nonfundamental explanation." In Explaining the Brain: Mechanisms and the Mosaic Unity of Neuroscience (chapter 6). Oxford: Oxford University Press.

In this chapter, and throughout the entirety of the book, Craver defends the claim that explanations in neuroscience span multiple levels. Explanations in neuroscience are mechanistic explanations, he claims, which require formulation of relations between entities and activities at various levels of organization. In this chapter, Craver defends the existence of causal relations at higher levels of explanation.

Schaffner, K. (1993). "Theory structure, reduction and disciplinary integration in biology." Biology and Philosophy, 8, 319-347.

Schaffner offers an argument in favor of multilevel explanations in biology. Laws in biology do not have the simple, axiomatic structure of the basic laws of physics. Instead, they are what Schaffner terms "middle range" theories: theories that exist at a central level of explanation, and actively incorporate information from both higher and lower levels. Schaffner uses an example of learning in the Aplysia to demonstrate why such interlevel explanations are required in the biological sciences.

Craver, C. (2002). "Interlevel experiments and multilevel mechanisms in the neuroscience of memory." Philosophy of Science (supp), 69, S83-97.

Craver details a variety of interlevel experimental strategies used in the neuroscientific study of memory and shows how our resultant understanding of memory is inherently interlevel. Explanation of spatial memory, in particular, involves a range of entities from molecules to brain regions to the entire organism. Thus, neuroscientists engage in many forms of activation and inhibition methods to probe the mechanism at these various levels.

Marr, D. (1977). "Artificial Intelligence: A Personal View." Artificial Intelligence, 9, 37-48.

In this article, Marr defends the need of multiple levels of explanation for cognitive phenomena. Specifically, he addresses the importance of understanding the informational processing problem at hand before any further fruitful work can be done. He criticizes the contemporary state of artificial intelligence and neurophysiology, which are focused on the algorithmic and implementational levels of explanation, respectively, without a proper formulation of the computational questions at hand. Only by identifying such questions, he claims, will the study of intelligence move forward.

iii. Reduction

Reductionist approaches to multilevel explanation claim that the true explanation exists at one particular level, and that explanations at other levels are, at best, superfluous. Most commonly, the appropriate or fundamental level of explanation is the lowest.

Coltheart, M., & Langdon, R. (1998). "Autism, modularity, and levels of explanation in cognitive science." Mind & Language, 13, 138-152.

Coltheart and Langdon argue that autism, and other cognitive neuropsychiatric syndromes should be understood at the cognitive level of explanation. Because the symptoms of autism are caused by a myriad of different factors in different cases, explanation of this disorder at the neural level will not be possible. Coltheart and Langdon's view thus differs from other reductionist views in that they offer defense of a higher level of explanation over lower ones.

Owens, D. (1989). "Levels of Explanation." Mind, 98, 59-79.

Owens argues in favor of the claim that physics is the fundamental science. Physics has causal and ontological priority over other sciences, however, explanations at the other levels are still relevant for a variety of purposes. Owens view thus exemplifies the possibility of reductionism without eliminativism.

Churchland, P.M. (1981). "Eliminative Materialism and the Propositional Attitudes," Journal of Philosophy, 78, 67-90.

Churchland argues, here and elsewhere, for the elimination of folk psychological explanations in favor of the more scientifically respectable explanations of cognitive science. Churchland defends this eliminativism by claiming that the commonsense theory of the mental that exists at the personal level is a false theory, and thus will have no place in a rigorous theory of cognition developed by a mature scientific psychology.

Ramsey, W., Stich, S., & Garon, J. (1990). "Connectionism, eliminativism, and the future of folk psychology." Philosophical Perspectives, 4, 499-533.

The authors evaluate the impact of the burgeoning field of connectionist modeling on commonsense theories about the mind. Folk psychology, on their view, is committed to propositional modularity. That is, giving commonsense explanations of human behavior is a practice predicated upon the propositionally structured states being explicitly represented in the heads of cognizers. Connectionist modeling shows how it might be possible to achieve the cognitive sophistication of human beings without such localization and symbolic representation, as connectionist models are systems of distributed non-symbolic representations. Thus, the authors conclude that the truth of connectionism would require the elimination of higher-level explanations of human behavior.