Computational explanations focus on information processing tasks of specific cognitive capacities. In this paper, we argue that there are at least two different kinds of computational explanations; the interlevel and the intralevel ones. Moreover, it will be argued that neither interlevel nor intralevel computational explanations can be subsumed under the banner of standard mechanistic explanations. In the case of interlevel explanations, the problem is the direction of explanation, and in the case of intralevel explanations, the problem are the dependencies that the explanations track. Finally, it is argued that in the context of explanation of cognitive phenomena, it may be necessary to defend more liberal and pluralistic views of explanation, which would allow that there are also some non-mechanistic forms of explanation.
Computational explanations focus on information processing required in exhibiting specific cognitive capacities, such as perception, reasoning or decision making. At an abstract level, these computational tasks can be specified as mappings from one kind of information to another.
These explanations can increase our understanding of a cognitive process at least in three ways: (i) they can explain a certain cognitive phenomenon in terms of fundamental rational or mathematical principles governing the information processing task faced by a system, or (ii) they can explain by describing the formal dependencies between certain kinds of tasks and certain kinds of information processing requirements. Moreover, in many computational accounts1 it is often assumed that (iii) computational explanations can explain the phenomenon in terms of its implementation in more primitive constituent processes.
In recent years, a number of philosophers have proposed
that computational explanations of cognitive phenomena
could be seen as instances of mechanistic explanation
In what follows, we will argue that while fulfilling these epistemic needs is essential in computational explanation in the cognitive sciences, only the last mode of explanation conform to the mechanists’ way of thinking what genuine mechanistic explanation is.
Thus, we conclude that either philosophers of cognitive
science need to embrace non–mechanistic computational
explanations, or extend the scope of what counts as
“mechanistic” explanation in cognitive science.
Within the last ten years, a growing number of philosophers
have defended the view that computational explanations are
mechanistic explanations
According to this mechanistic account, the goal of computational explanation is to characterize the functions that are being computed (the what) and specify the algorithms by which the system computes the function (the how). In other words, the idea is that an information processing phenomenon would be explained by giving a sufficiently accurate model of how hierarchical causal systems composed of component parts and their properties sustain or produce the phenomenon2.
2 Constructing an explanatory mechanistic model thus involves mapping elements of a mechanistic model to the system of interest, so that the elements of the model correspond to identifiable constituent parts with the appropriate organization and causal powers to sustain that organization. These explanatory models should specify the initial and termination conditions for the mechanism, how it behaves under various kinds of interventions, how it is integrated with its environment, and so on.
This kind of mechanistic “computational” explanations
track causal dependencies at the level of cognitive
performances. They correspond to explanations which
David
In Marr´s trichotomy, computational explanations
specifies what are the information processing tasks, and
what is computed and why. Computational explanations
give an account of the tasks that the neurocognitive system
performs, or problems that the cognitive system in question
is thought to have the capacity to solve, as well as the
information requirements of the tasks
This level of explanation is also the level, whereby the
appropriateness and adequacy (for the task) of mappings
from representations to others are assessed
It is important to distinguish two different types of computational explanations. Firstly, there are interlevel computational explanations, which explain by describing, how the possible behavior or processes of a system is governed by certain information processing principles, rather than explain how certain algorithms compute certain functions. These computational explanations display the function that the mechanism computes and they explain and why this function is appropriate for a given cognitive task.
Some of our critics, such as Milkowski, have claimed that
we see these interlevel computational explanations as
“systemic explanations that show how a cognitive system
can have some capacities”
In this sense, interlevel explanations explain the behavior of mechanisms at the algorithmic and implementation levels. In such explanations, the explanans is at the “upper” computational level, and the explananda are at the “lower” algorithmic or performance levels. For example, if one considers, why certain synaptic change is such-and-such, answers are often something like “because it serves to store the value of x needed in order to compute y. Or, why is the wiring in this ganglion such-and-such? Because it computes, or approximates computation of x. In other words, phenomena at the lower levels are explained by their appropriateness of the mechanism for the computational tasks.
Secondly, there are computational explanations, which are rather intralevel than interlevel explanations. In short, these explanations track formal dependencies between certain kinds of information processing tasks, and they explain by describing certain kinds of information processing requirements at the level of cognitive competences.
There are different views about the nature of the formal
dependencies, which are tracked by these computational
explanations. Some take it that the dependencies can be
described intentionally i.e. in terms of informational
content, while some other, such as Egan (1992) argues that
computational explanations track appropriate mathematical
dependencies by specifying the mathematical
input-outputfunctions that is being computed. There are also some
pluralistic views; for instance
So, there are at least two different kinds of computational
explanations; the interlevel and the intralevel ones. In the
following sections, we will argue that neither interlevel nor
intralevel computational explanations can be subsumed
under the banner of standard mechanistic explanations. In
the case of interlevel explanations, the problem is the
direction of explanation
In a nutshell, the problem for standard mechanistic accounts of interlevel explanations goes as follows: In standard accounts (constitutive) mechanistic explanations are characterized in such a way that in inter-level computational explanations, the explanans is at a lower level than the explanandum. For example Craver (2001, p. 70, emphasis added) notes that “ (Constitutive) explanations are inward and downward looking, looking within the boundaries of X to determine the lower level mechanisms by which it can Φ. The explanandum… is the Φ-ing of an X, and the explanans is a description of the organized σ-ing (activities) of Ps (still lower level mechanisms).”
In those explanations, phenomena at a higher level of
hierarchical mechanistic organization are explained by their
lower-level constitutive causal mechanisms but not vice
versa
One might argue that this analysis ignores the possibility
that computational explanations are contextual rather than
constitutive mechanistic explanations. In the mechanistic
terminology, the contextual explanations explain how the
“higher-level” mechanism constrains what a lower level
mechanism does, and one computational mechanism can be
a component of a larger computational system, while the
latter serves as the contextual level for the former. For
example Bechtel seems to accept this position, when he
remarks that “since (marrian) computational explanations
address what mechanisms are doing they focus on
mechanisms “in context””
Now, if computational explanations was contextual explanations, then our argument would fail. Namely, if computational-level explanations were contextual explanations, and if contextual explanation is a subspecies of standard mechanistic explanations, then computational level explanations would be a subspecies of mechanistic explanations.
However, it is possible to argue that computational explanations are not contextual explanations in the standard mechanistic sense. For instance, Craver characterizes contextual explanations as explanations, which “refer to components outside of X” and are “upward looking because they contextualize X within a higher level mechanism”. On this view, a description of how a cognitive system “behaves” in its environment, or how an organization of a system constraints the behavior of its components, require a spatiotemporal interpretation for the mechanisms. But, as we argued in 2011, computational explanations do not necessarily refer to spatiotemporally implemented higherlevel mechanisms, and they do not involve spatiotemporally implemented components “outside of (spatiotemporally implemented) X”. Instead, they refer to abstract “mechanisms”, which are not causally or spatiotemporally implemented.
In other words, the problem is that in standard mechanistic accounts, in contextual explanations the “contexts” are expressed in causal and spatiotemporal terms, not in terms of information processing at the level of computational competences. Crucially, this kind of view conceives contextual explanations as a kind of systemic explanations, in which the uppermost level of the larger mechanism will still remain non-computational in character.
For this reason, computational explanations are not these “systemic” contextual explanations. In contrast, we claim, computational explanations involve abstract mechanisms, which are not causally, but logically governing the behavior of the mechanisms at the lower levels.
Now, let´s move to the intralevel computational explanations. Why cannot they be seen as standard mechanistic explanations? Well, the answer is that they simply track different kinds of dependencies. While algorithmic and implementation level explanation track causal or constitutive dependencies at the level of cognitive or neural performances, intra-level computational explanations track formal dependencies between certain kinds of information processing tasks at the level of cognitive competences.
Because of this, these different modes of explanation are not necessarily logically dependent on each other. Thus the computational explanations at the highest level may be formulated independently of assumptions about the algorithmic or neural mechanisms which perform the computation.
Some of our critics, such as
Indeed, we defend the view that computational
explanations can be in principle - if not in practice
constructed largely autonomously with respect to the
algorithmic or implementation levels below. That is:
computational problems of the highest level may be
formulated independently of assumptions about the
algorithmic or neural mechanisms which perform the
computation
However,
Is this, really, an argument against our position? As
Kaplan himself remarks, Marr rejects “the idea that any
computationally adequate algorithm (i.e., one that produces
the same input-output transformation or computes the same
function) is equally good as an explanation of how the
computation is performed in that particular system”
But then, we are not talking about competence level explanations anymore. When the issue is how the computation is performed in the particular system, such as in human brains, then the explanation is given in terms of algorithmic or neural processes, or mechanisms, if you will. Then, naturally, the crucial issue is what kinds of algorithms are possible for a certain kind of system, or whether the system has structural components that can sustain the information processing that the computational model posits at the neural level. If one aims to explain how our brains are able to perform some computations, then – of course – one should take the actual neural implementation and the constraints of the possible neurocognitive architecture into account as well.
But given this, these kinds of explanations are
explanations at the algorithmic or performance level, not at
the computational or competence level. Because of this, we
also find position defended by
For this reason, the remark that computational explanations do not describe how the computational system “actually works” is not an argument against the logical independence of the computational level explanations.
A more problematic issue is to what extent computational
explanations are explanatory after all. Although Milkowski
may partially misinterpret our position, he still raises an
important question concerning the explanatory character of
computational explanations
If computational explanations are characterized as explanations which answer questions such as: “What is the goal of this computation?”, it may be claimed that we fail to make a distinction between task analysis and genuine explanations.
A task analysis breaks a capacity of a system into a set of sub-capacities and specifies how the sub-capacities are (or may be) organized to yield the capacity to be explained. Obviously, if computational explanations are mere descriptions of computational tasks, then they are not explanations at all.
However, computational explanations are clearly more than mere descriptions of computational tasks, because they describe formal dependencies between certain kinds of tasks and certain kinds of information processing requirements. If these formal dependencies are such that descriptions of them not only offer the ability to say how the computational layout of the system actually is, but also the ability to say how it would be under a variety of circumstances or interventions, they can be counted as explanatory4.
In other words, if these descriptions answer questions such as “Why does this kind of task create this kind of constraint rather than that kind of constraint?” by tracking such formal dependencies which can explain what makes the difference, then these descriptions can be explanatory.
Obviously, computational explanations of this sort are not causal explanations. However, in the context of explanation of cognitive phenomena, it may be necessary to defend more liberal and pluralistic views of explanation, which would allow that there are also some non-causal forms of explanation.
We agree with mechanists that when we are explaining how cognitive processing actually happens for example in human brains, it is a matter of causal explanation to tell how the neuronal structures sustain or produce the information processing in question. However, we still defend the view that there are other modes of explanation in cognitive sciences as well.
Some explanations of cognitive phenomena can be
subsumed under the banner of “mechanistic explanation”.
Typically those explanations are neurocognitive
explanations of how certain neurocognitive mechanisms
produce or sustain certain cognitive phenomena, but also
some psychological explanations can be seen as instances of
mechanistic explanations. Moreover, if a more liberal
interpretation for the term mechanism is allowed, then some
computational or competence level explanations may also
qualify as mechanistic explanations
4 This is a non-causal modification of the Woodward´s
manipulationist account of explanation
Nevertheless, we think that there are compelling reasons to doubt whether mechanistic explanation can be extended to cover all cognitive explanations. There are several reasons for this plea for explanatory pluralism: Firstly, it is not clear whether all cognitive systems or cognitive phenomena can be captured mechanistically. Mechanistic explanations require that the system can be decomposed i.e. analyzed into a set of possible component operations that would be sufficient to produce or sustain the phenomenon in question (Bechtel & Richardson 1993). Typically a mechanism built in such a manner will work in a sequential order, so that the contributions of each component can be examined separately (Bechtel & Richardson 1993).
However, in cognitive sciences there are examples of systems – such as certain neural nets – which are not organized in such a manner. As Bechtel and colleagues remark, the behavior of these kinds of systems cannot be explained by decomposing the systems into subsystems, because the parts of the networks do not perform any activities individually that could be characterized in terms of what the whole network does (Bechtel & Richardson 1993; Bechtel 2011, 2012). Hence, it is an open question to what extent the behavior of these kinds of systems can be explained mechanistically. At the very least, it will require adopting a framework of mechanistic explanation different from the one that assumes sequential operation of decomposable parts (Bechtel 2011, 2012; Bechtel & Abrahamsen 2011).
Moreover, Von Eckardt and Poland (2004) raise the question to what extent the mechanistic account is appropriate for those explanations which involve appeal to mental representations or to the normative features of certain psychopathological phenomena. Although we find Von Eckardt and Poland´s argumentation slightly misguided, we still think that it is important to consider the normative aspects of cognitive phenomena. Cognitive systems are, after all, adaptive systems which have a tendency to seek “optimal”, “rational” or “best possible” solutions to the information processing problems that they face. Because of this, cognitive processes are not only goaldirected, but also normative. It is not clear how well this normative aspect of cognitive systems can be captured by mechanistic explanations.
Thirdly, some philosophers have paid attention to the fact
that there are examples of explanatory computational
models in cognitive sciences which focus on the flow of
information through a system rather than the mechanisms
that underlie the information prosessing
Finally, in recent years cognitive scientists have raised the
possibility that there are some universal, law-like principles
of cognition, such as the “principle of simplicity”,
“universal law of generalization” or the “principle of
scalevariance”
These explanations are “principle based” rather than mechanistic explanations. Moreover, Chater and colleagues seem to suggest the mechanistic models of these phenomena may actually be derived from these general principles, and explanations that appeal to these general principles provide “deeper” explanations than the mechanistic explanations (Chater & Brown 2008). It is possible, that many of the so called computational level explanations turn out to be instances of these principle-based explanations rather than instances of mechanistic explanations.
In sum, taken together these diverse claims seem to imply that there is not a single, unified mode of explanation in cognitive sciences. Instead, they seem to suggest that cognitive sciences are examples of those sciences which utilize several different modes of explanation, only some of which can be subsumed under the mechanistic account of explanation.
Obviously, mechanistic explanation is a powerful framework for explaining the behavior of complex systems, and it has demonstrated its usefulness in many scientific domains. Also, many successful theories and explanations in cognitive sciences are due to this mechanistic approach. However, this does not imply that it would be the only way to explain complex cognitive phenomena.
In this paper, we have argued that there are at least two different kinds of computational explanations; the interlevel and the intralevel ones. Moreover, we have argued that neither interlevel nor intralevel computational explanations can be subsumed under the banner of standard mechanistic explanations. In the case of interlevel explanations, the problem is the direction of explanation, and in the case of intralevel explanations, the problem are the dependencies that the explanations track.
Obviously, computational explanations of this sort are not causal explanations. However, in the context of explanation of cognitive phenomena, it may be necessary to defend more liberal and pluralistic views of explanation, which would allow that there are also some non-causal forms of explanation.