Download working draft - DAVID KAPLAN | Macquarie University

Neural computation, multiple realizability, and the
prospects for mechanistic explanation
David Michael Kaplan
There is an ongoing philosophical and scientific debate concerning the nature of computational explanation in
the neurosciences. Recently, some have cited modeling work involving so-called canonical neural computations –
standard computational modules that apply the same fundamental operations across multiple brain areas – as
evidence that computational neuroscientists sometimes employ a distinctive explanatory scheme from that of
mechanistic explanation. Because these neural computations can rely on diverse circuits and mechanisms,
modeling the underlying mechanisms is supposed to be of limited explanatory value. I argue that these
conclusions about computational explanations in neuroscience are mistaken, and rest upon a number of
confusions about the proper scope of mechanistic explanation and the relevance of multiple realizability
considerations. Once these confusions are resolved, the mechanistic character of computational explanations can
once again be appreciated.
1. Introduction
Multiple realizability considerations have played a long-standing role in debates about reduction, explanation, and the
autonomy of higher-level sciences including psychology and functional biology. The concept of multiple realizability
(hereafter, MR) was initially introduced in the philosophy of mind to serve as an argument against mind-brain identity
theories, and to motivate a functionalist view about the mind. Since the 1970s, however, there has been a gradual shift
in focus beyond ontological issues raised by MR to epistemological (i.e., explanatory and methodological) ones
including whether psychological explanations and methods are autonomous from or reduce to those encountered in
lower-level sciences such as neuroscience. Fodor (1974) famously appeals to MR in an attempt to block the reduction of
higher-level theories and/or laws of the special sciences to those of lower-level sciences. Putnam (1975) invokes MR to
argue for the “autonomous” character of higher-level explanatory generalizations in psychology. Several decades later,
Fodor explicitly foregrounds the explanatory issues raised by MR, asserting that “the conventional wisdom in the
philosophy of mind is that psychological states are functional and the laws and theories that figure in psychological
explanations are autonomous [i.e., are not reducible to a law or theory of physics]” (1997, 149). These formulations of
the MR argument directly link up with debates in the philosophy of science concerning explanation and theory
reduction in ways that many early construals do not.1
Although philosophical excitement about MR and the prospects for the autonomy of psychology in its
traditional guise have faded to some degree in recent years, MR considerations are once again being invoked in a
different light in order to draw conclusions about the nature of explanation, and in particular, computational explanation
in neuroscience and psychology. More specifically, it has recently been argued by several authors (Carandini 2012;
1
For further discussion of the distinction between epistemological and ontological construals of reductionism, and the importance of the former over
the latter in scientific contexts, see Ayala (1968) Hoyningen-Huene (1989). Although these authors are explicitly concerned with reductionism in
biology, the same kinds of considerations are relevant in the present context. Finally, for examples of work on MR that prioritize explanatory over
ontological issues associated with reductionism, see many contributions from the philosophy of biology (e.g., Kitcher 1984; Rosenberg 2001; Sarkar
1992; Sober 1999; Waters 1990).
1
Carandini and Heeger 2012; Chirimuuta 2014) that certain kinds of computations performed in the brain — so-called
canonical neural computations — cannot be explained in mechanistic terms. Briefly, canonical neural computations
(hereafter CNCs) are defined as “standard computational modules that apply the same fundamental operations in a
variety of contexts” (Carandini and Heeger 2012, 51). According to these authors, the reason why CNCs cannot be
accommodated by the framework of mechanistic explanation – which is dominant across the life sciences – is that these
computations are associated with multiple neural circuits or mechanisms in the brain, which can vary from region to
region and from species to species. In other words, CNCs cannot be analyzed in mechanistic terms because they are
multiply realized. Instead, as Chirimuuta (2014) argues, modelling work involving CNCs must embody an entirely
different, non-mechanistic form of explanation. In this chapter, I argue against the necessity of this claim by showing
how MR fails to block the development of adequate mechanistic explanations of computational phenomena in
neuroscience. In this manner, I show how MR considerations are largely irrelevant to assessing the quality of
mechanistic explanations. Once these confusions are resolved, it becomes clear that CNC explanations can be properly
understood as mechanistic explanations.
The chapter is organized as follows. In Section 2, I discuss some traditional MR-based arguments for the
explanatory autonomy of psychology and briefly highlight their well-known limitations. In Section 3, I introduce the
concept of canonical neural computation and outline how MR considerations are supposed to raise problems for
mechanistic explanations of these phenomena. In Sections 4 and 5, I describe two different mechanistic explanations in
neuroscience that invoke the same neural computation to show how MR considerations, although operative in these
cases, are irrelevant to assessing the quality of the explanations provided. In Section 6, I show how persistent
confusions over the proper scope of mechanistic explanations serve to underwrite MR-based challenges to mechanistic
explanation of computational phenomena. Specifically, I elaborate how scope is not an appropriate norm on mechanistic
explanation and therefore even mechanistic models with highly restricted scope can in principle be explanatorily
adequate. This conclusion provides a foothold the view that mechanistic models of computational phenomena can
provide legitimate explanations, even in the presence of multiple realizability. In Section 7, I return to the specific
debate concerning the nature of CNC explanations and expose these same confusions about the relationship between
scope and mechanistic explanation. In Section 8, I summarize the mechanistic perspective on neural computation
defended in this chapter.
2 Traditional MR-based arguments for explanatory autonomy
The central idea behind traditional, MR-based arguments for autonomy is that the phenomena or kinds identified in the
theories of a higher-level science such as psychology can be multiply realized by heterogeneous sets of lower-level
realizers, and so will invariably be cross-classified by, and therefore cannot be explained in terms of, the phenomena or
kinds invoked in lower‐level theories (e.g., Fodor 1974, 1997; Putnam 1975). In virtue of this, higher-level phenomena
or kinds (and the theories and generalizations that invoke them) are said to enjoy a certain degree of explanatory
autonomy or independence from lower-level phenomena or kinds (and the theories or generalizations that invoke them).
Before proceeding, it is important to distinguish this kind of autonomy from methodological autonomy. A scientific
discipline X is methodologically autonomous from some lower-level discipline Y if X’s investigative methods, discovery
procedures, and most importantly, its taxonomic categories can vary independently of Y’s, or vice versa.2 Some
advocates of autonomous computational psychology implicitly embrace methodological autonomy without taking a
stance on explanatory autonomy. For example, the psychologist Philip Johnson-Laird embraces methodological
2
For a similar characterization of methodological autonomy, see Feest (2003).
2
autonomy several decades when boldly pronouncing that “[t]he mind can be studied independently from the brain.
Psychology (the study of the programs) can be pursued independently from neurophysiology (the study of the machine
and the machine code)” (Johnson-Laird 1983, 9).
It is important to keep these two notions of autonomy distinct because the evidence against the methodological
autonomy of psychology from neuroscience is relatively strong, whereas the case concerning the explanatory autonomy
of psychology remains far more uncertain. Keeley (2000), for example, argues compellingly that the discovery and
categorization of psychological kinds frequently depends on structural information about the brain delivered from the
neurosciences. Similarly, Bechtel and Mundale (1999) make a powerful case that the taxonomic enterprises of
psychology and neuroscience are not independent in the ways implied by the thesis of multiple realizability. In
particular, they show how the taxonomic identification of brain areas in neuroscience relies intimately on the
deployment of functional criteria and prior categorizations from psychology. They also demonstrate how efforts to
decompose psychological capacities or functions often depends on, and is sometimes refined by, structural information
about the brain and its organization. The functional and structural taxonomies of psychology and neuroscience thus
appear to reflect the complex interplay and interaction between the disciplines. The resulting picture is arguably one of
methodological interdependence, not autonomy. Consequently, even if the ontological thesis that psychological states
are (or can in principle be) multiply realized in different substrates, there is good reason to think this does not support
the methodological claim that neuroscience lacks usefulness in guiding psychological investigations, and vice versa. By
contrast, the status of the explanatory autonomy of psychology is far less settled.3 In what follows, the focus will
exclusively be on this form of autonomy.
As indicated above, many of the hallmark defenses of the explanatory autonomy of psychology and other
higher-level sciences invoke MR considerations in order to deny the possibility of reducing the theories or laws of the
higher-level science to those of the lower-level science (e.g., Fodor 1974; Putnam 1975).4 Proponents of these views
(and their critics) have generally assumed a traditional conception of theory reduction, according to which some higherlevel theory (or law) is said to be reducible to (and also explained by) some lower-level theory (or law) just in case the
former can be logically derived from the latter, together with appropriate bridge principles and boundary conditions
(e.g., Nagel 1961; Schaffner 1967). Since the terminology invoked by the higher- and lower-level theories often differ
in appreciable ways, so-called bridge principles connecting the terms or predicates of the two theories in a systematic
(one-to-one) manner are necessary to ensure the derivation can go through.5 The classical antireductionist strategy thus
involves showing how the bridge principle-building enterprise breaks down because higher-level phenomena are often
multiply realized by heterogeneous sets of lower-level realizers, implying an unsystematic (many-to-one) mapping
between the terms of the two theories. For example, Fodor, who famously frames the issue in terms of the challenges
MR poses for the logical positivist goal of the unity of science, states that:
The problem all along has been that there is an open empirical possibility that what corresponds to the natural kind
predicates of a reduced science may be a heterogeneous and unsystematic disjunction of predicates in the reducing
3
Before proceeding it is worth preempting a certain mistaken conclusion that this analysis might elicit. Although I am focusing here on challenges
raised about the standard conclusion of the MR argument — the autonomy of psychology — this does not imply that this is the exclusive or even
primary focus in the critical literature. Many authors have also challenged the MR argument by denying the MR premise itself. Although MR once
held the status as an unquestioned truth, as philosophers started to pay closer attention to empirical research in neuroscience, the evidence for multiple
realization began to appear less clear-cut than had been initially assumed (e.g., (Bechtel and Mundale 1999; Polger 2009; Shapiro 2000, 2008).
4
According to the traditional model, theory reduction implies that all the higher-level laws and observational consequences can be derived from
information contained in the lower-level theory. Strictly speaking, then, the higher-level (reduced) theory can make no non-redundant or
“autonomous” informational or explanatory contribution beyond that made by the lower-level (reducing) theory.
5
Because the lower-level theory will typically only apply over a restricted part of the domain of the higher-level theory (Nagel 1961) or at certain
limits (Glymour 1970; Batterman 2002), boundary conditions that set the appropriate range for the reduction are also typically required for the
derivation to succeed. For further discussion, see Kaplan (2015).
3
science, and we do not want the unity of science to be prejudiced by this possibility. Suppose, then, that we allow that
bridge statements may be of the form […] Sx ⇆P1x ∨ P2x ∨ … ∨ Pnx, where ‘P1x ∨ P2x ∨ … ∨ Pnx’ is not a
natural kind predicate in the reducing science. I take it that this is tantamount to allowing that at least some 'bridge laws'
may, in fact, not turn out to be laws, since I take it that a necessary condition on a universal generalization being lawlike
is that the predicates which constitute its antecedent and consequent should pick natural kinds.” (Fodor 1974, 107-108).
The major difficulty with this and many other traditional positions staked out on both sides of the debate over
explanatory autonomy is that they all commonly assume the appropriateness of what is now widely recognized as an
outdated and inapplicable law-based model of theory reduction and explanation (Rosenberg 2001).6 The problem, now
widely recognized among philosophers of the special sciences including biology and psychology, is the conspicuous
absence of lawful generalizations, either at the level of the reducing theory or the reduced theory (e.g., Machamer,
Darden, and Craver 2000; Woodward 2000).7 This is problematic for traditional reductionists because there is no scope
for a reduction without laws in either theory. Interestingly, the absence of laws also raises difficulties for traditional
defenders of autonomous psychological explanation because all recourse to laws, even at the level of the reduced
theory, becomes empty in this nomological vacuum. For this reason, philosophical interest in these debates has
diminished in recent years.
Despite diminished attention of late, MR considerations are once again being invoked in a different light in
order to draw conclusions about the nature of explanation, and in particular, computational explanation in neuroscience.
Critically, if these arguments about the distinctness and autonomy of computational explanation in neuroscience can
secure a foothold, then broader conclusions about the autonomy of computational explanations in psychology are also
within reach. In what follows, I will argue that even when rehabilitated from the outmoded framework of laws, MR
considerations are orthogonal to issues concerning explanation.
3 Canonical neural computation explanations and multiple realizability
Several recent authors (e.g., Carandini 2012; Carandini and Heeger 2011; Chirimuuta 2014) claim that certain kinds of
computations performed in the brain – canonical neural computations (CNCs) – cannot be explained in mechanistic
terms because they rely on a diversity of neural mechanisms and circuits. According to Carandini, a leading visual
neuroscientist, the brain “relies on a core set of standard (canonical) neural computations: combined and repeated across
brain regions and modalities to apply simpler operations to different problems.” (Carandini 2012, 508). The key idea is
that these neural computations are repeatedly deployed across different brain areas, different sensory modalities, and
even across different species.8 Prominent examples of CNCs include linear filtering and divisive normalization. Linear
filtering involves a computation of the weighted sum of synaptic inputs by linear receptive fields. Divisive
6
Explanation and theory reduction are closely linked in the traditional Nagelian model. Just as the D–N model ties successful explanation to the
presence of a deductive or derivability relationship between explanans and explanandum statements, so too successful reduction is tied to the presence
of a deductive relationship between two theories. Intertheoretic reduction is thus naturally treated as a special case of deductive-nomological
explanation. For further discussion, see Kaplan (2015).
7
To the extent that the notion of “law” continues to find uses in these scientific disciplines it is perhaps best characterized in terms of describing the
pattern, effect, or regularity to be explained, rather than as providing the explanation (e.g., Bechtel 2008; Cummins 2000; Wright and Bechtel 2006).
For example, Fitts’ law describes the tradeoff (negative correlation) between speed and accuracy in goal-directed, human motor behavior. The WeberFechner law describes the robust psychological finding that the just-noticeable difference between two sensory stimuli is proportional to their
magnitudes. These well-known generalizations mathematically describe but do not explain what gives rise to these widespread patterns or regularities
in human behavior and perception.
8
One assumption not addressed in this chapter is the foundational one concerning whether neural systems genuinely compute and process
information. While most computational neuroscientists assume that neural systems perform computations (canonical or otherwise), and so go beyond
what many other computational scientists assume about the phenomena they investigate when building computational models (e.g., computational
climate scientists do not think that climate systems compute in the relevant sense), this remains open to challenge. For further discussion of this
important issue, see Piccinini and Shagrir (2014).
4
normalization involves the activity of one neuron (or neural population) being normalized or adjusted by the activity of
other neurons in the same region. Below, I will focus exclusively on the example of divisive normalization.
The divisive normalization model was initially introduced to account for certain response properties of neurons
in primary visual cortex (V1) including contrast saturation and cross-orientation suppression (Heeger 1992; Carandini
and Heeger 1994; Carandini et al. 1997), which could not be handled by linear models of orientation selectivity (e.g.,
Hubel and Wiesel 1968). Cross-orientation suppression, for example, describes the phenomenon in which so-called
simple cells in V1 exhibit weaker responses to the superposition of two orientation stimuli than would be predicted by
the sum of the responses to each individual stimulus alone, even when one of the component stimuli evokes no response
at all (Figure 1).
Figure 1. Example of cross-orientation suppression in V1 neurons. (A) Grating stimuli are presented in
alignment with the neuron’s preferred orientation (top row; test stimuli), orthogonal to the neuron’s preferred
orientation (left column; mask stimuli), or by superimposing the two (remaining panels; plaid stimuli). (B)
Corresponding responses of a typical V1 neuron to the same stimulus set. Cross-orientation suppression is most
evident when non-preferred orientation grating stimuli are presented at high contrasts (bottom rows). For this
neuron, the superposition of two gratings (e.g., bottom right panel in A), one at the preferred-orientation (e.g.,
top right panel in A) and one in the null or orthogonal orientation (e.g., bottom left panel in A), evokes a smaller
response than does the preferred-orientation grating alone, even though the null orientation stimulus has no
effect on firing as compared to baseline levels. (Modified from Freeman et al. 2002)
Models of orientation selectivity such as that of Hubel and Wiesel, in which converging feedforward inputs from the
lateral geniculate nucleus (LGN) are combined by taking a weighted sum, fail to account for this and other non-linear
response properties. The divisive normalization model accommodates these non-linearities by including not only a term
for the feedforward or excitatory inputs (sometimes called the summative or simply receptive field), but also a term in
the divisor reflecting activity pooled from a large number of nearby neurons (sometimes called the suppressive field)
(equation 1). The overall neural response (output) thus reflects the ratio of activity between these two terms, and the
essential computational operation being performed is division. A generalized version of the normalization model is:
[1]
where Rj is the normalized response of neuron j; Dj in the numerator represents the neuron’s driving input; ∑j Dk
represents the sum of a large number of inputs, the normalization pool. The constants y, 𝜎, and n represent free
parameters fit to empirical data: y scales the overall level of responsiveness, 𝜎 prevents undefined values resulting from
division by zero, and n is an exponential term that modulates the values of individual inputs.
5
The model successfully accounts for response saturation in V1 neurons with increasing stimulus contrast
because terms in both the numerator and denominator positions scale proportionally with changes in stimulus contrast.
Increasingly larger driving/excitatory responses elicited by increasing contrast are kept neatly in check/effectively
suppressed (are “normalized” in the mathematical sense) because they are being divided by an increasingly larger
number in the denominator reflecting inputs from the entire suppressive field. The model also predicts cross-orientation
suppression effects: Responses of a V1 neuron to an optimally oriented stimulus (e.g., a grating stimulus aligned with
the neuron’s preferred orientation) are suppressed by superimposing a mask stimulus of different orientation. Even a
sub-optimal visual stimulus that is completely ineffective in eliciting a response such as a grating stimulus with an
orientation orthogonal to the neuron’s preferred orientation (the null orientation); Figure 1, left column) nonetheless
suppresses the response to a stimulus that optimally drives the neuron when the two stimuli are presented
simultaneously (Figure 1, top row). The normalization model accounts for this puzzling effect because suppression in
the denominator increases as a function of increasing contrast in both the preferred and non-preferred orientation
stimuli, whereas excitation in the numerator only increases with increasing contrast in the preferred orientation. While
the model's predictive successes in V1 are noteworthy, interest in divisive normalization has expanded considerably
because of its apparent ubiquity in nervous systems.
Critically, the normalization equation has since been applied successfully to a wide range of systems well
beyond mammalian V1. It has been used to characterize responses in invertebrate olfactory system (Olsen, Bhandawat,
and Wilson 2010); mammalian retina (Carandini and Heeger 2011); sub-cortical structures including the LGN (Bonin,
Mante, and Carandini 2005; Mante et al. 2005); higher visual cortical areas including V4 (Reynolds, Pasternak, and
Desimone 2000; Reynolds and Heeger 2009), area MT/V5 (Simoncelli and Heeger 1998; Rust et al. 2006), and IT
(Zoccolan, Cox, and DiCarlo 2005); and in non-visual cortical areas including auditory cortex (Rabinowitz et al. 2011).
The model has even been applied to describe certain operating characteristics of the invertebrate olfactory system.
Simply stated, the same computation is performed in multiple neural circuits across a diversity of brain regions and
species. This, we are told, seriously limits the prospects of providing a suitable mechanistic explanation of this
phenomenon. Carandini has the following to say about canonical neural computation and the relevance of underlying
mechanistic details:
Crucially, research in neural computation does not need to rest on an understanding of the underlying biophysics. Some
computations, such as thresholding, are closely related to underlying biophysical mechanisms. Others, however, such as
divisive normalization, are less likely to map one-to-one onto a biophysical circuit. These computations depend on
multiple circuits and mechanisms acting in combination, which may vary from region to region and species to species.
In this respect, they resemble a set of instructions in a computer language, which does not map uniquely onto a specific
set of transistors or serve uniquely the needs of a specific software application. (Carandini 2012, 508)
Elsewhere, Carandini and Heeger echo the same point:
[T]here seem to be many circuits and mechanisms underlying normalization and they are not necessarily the same
across species and systems. Consequently, we propose that the answer has to do with computation, not mechanism.
(Carandini and Heeger 2012, 60)
Given the apparent boldness of these claims, it is unsurprising that philosophers have started to take notice. For
example, Mazviita Chirimuuta (2014), seeking to build on considerations of this kind, develops a full-blown argument
against mechanistic analyses for these kinds of neural computations. Canonical neural computation, she says, “creates
difficulties for the mechanist project [because these] computations can have numerous biophysical realisations, and so
straightforward examination of the mechanisms underlying these computations carries little explanatory weight.”
6
(Chirimuuta 2014, 127). CNC modeling is instead supposed to embody a “distinct explanatory style”, and “[s]uch
explanations cannot be assimilated into the mechanistic framework” (Chirimuuta 2014, 127). The claim here is clear:
CNCs cannot be accommodated by the framework of mechanistic explanation, and must therefore embody a distinctive
non-mechanistic explanatory form, because these computations are associated with multiple neural circuits or
mechanisms in the brain. In other words, CNCs are multiply realized, and therefore cannot be analyzed in mechanistic
terms.
Chirimuuta
explicitly
formulates
the
key
premise
as
follows:
[The] most important characteristic of normalization as a CNC is that there is now good evidence that normalization is
implemented by numerous different biophysical mechanisms, depending on the system in question…In other words,
normalization is multiply realized. (Chirimuuta 2014, 138-139)
According to Chirimuuta, the conclusion we are supposed to draw from this premise — that the same computations
performed across multiple neural circuits, areas, and species — is that CNCs cannot be characterized properly in
mechanistic terms and instead require a distinctive, non-mechanistic explanatory framework. Chirimuuta therefore
defends what she calls a “distinctness of computation” thesis about CNCs, according to which “ there can be principled
reasons for analyzing neural systems computationally rather than mechanistically.” (2014, 139). One clear-cut way of
defending the “distinctness” thesis would be to show that CNC explanations fail to satisfy various norms on adequate
mechanistic explanations. Indeed, this is precisely what Chirimuuta sets out to do. She argues that CNC explanations
“should not be subject to mechanistic norms of explanation” (2014, 131). In particular, she identifies Kaplan and
Craver’s (2011) Model–Mechanism–Mapping (3M) principle as the central mechanistic norm from which CNC
explanations evidently depart. As a reminder, the 3M principle states:
3M: A mechanistic model of a target phenomenon explains that phenomenon to the extent that (a) the elements in the
model correspond to identifiable components, activities, and organizational features of the target mechanism that
produces, maintains, or underlies the phenomenon, and (b) the pattern of dependencies posited among these elements in
the model correspond to causal relations among the components of the target mechanism. (Kaplan 2011; Kaplan and
Craver 2011)
The objective of 3M is to consolidate some of the major requirements on explanatory mechanistic models and highlight
the primary dimensions along with mechanistic models may be assessed. Explanatory mechanistic models reveal
aspects of the causal structure of the mechanism – the component parts, their activities, and organization – responsible
for the phenomenon of interest. Mechanistic models are explanatorily adequate to the extent that they accurately and
completely describe or represent this structure.9 Chirimuuta claims that CNC models fail to satisfy 3M, and therefore
cannot be assimilated into the mechanistic framework. She characterizes the conclusion of her critical analysis as
follows10:
Along the way we have seen that the mechanists’ 3M requirement […] sit[s] oddly (to put it mildly) with the new
interpretation and status of the normalization model. For Carandini and Heeger (2012) no longer think that we should
9
Accuracy and completeness requirements on mechanistic models are implicit in 3M. It is beyond the scope of this chapter to address how the norms
of accuracy and completeness are central to assessing mechanistic explanations. For further discussion, see Craver and Darden (2013).
10
It is important to acknowledge that Chirimuuta (2014) not only develops a negative argument against embracing a mechanistic perspective on CNC
explanation, she also develops a positive characterization of what distinct kind of non-mechanistic explanatory form CNC models supposedly
encompass — what she terms an informational minimal model (or I-minimal model). Her positive view is interesting and merits consideration on its
own terms, but is beyond the scope of this chapter.
7
interpret the normalization model mechanistically, as describing components and dynamics of a neural mechanism. […]
[A]ny digging down to more mechanistic detail would simply lead us to miss the presence of CNC’s entirely, because
of their many different realizations. (140)
Here is one plausible way of summarizing the entire argument:
P1. If a given model is a mechanistic explanation, then it satisfies 3M.
P2. If a given phenomenon is multiply realizable, then a model of that phenomenon will not satisfy 3M.
P3. Canonical neural computations such as divisive normalization are multiply realizable.
P4. Models of canonical neural computations do not satisfy 3M. (Inference from P2 and P3)
C. Models of canonical neural computations are not mechanistic explanations. (Inference from P1 and
P4)
As should be obvious now, P2 is the crucial premise in the argument. P2 identifies MR as a sufficient condition for a
given model failing to satisfy 3M. There are at least two possible interpretations of P2 that are consistent with the letter
of 3M. One way of interpreting P2 is that it outlines a sufficient condition for a given model to count as a bad (or
inadequate) mechanistic explanation. Another interpretation is that is provides a sufficient condition for not counting as
a mechanistic explanation at all (good or bad). The latter interpretation seems to be the one that Chirimuuta prefers. In
what follows, I will make a case for rejecting (both interpretations of) P2 by showing how MR considerations are
irrelevant to assessing the quality of a mechanistic 0explanation.11 Once this barrier to understanding is removed, it will
be clear how CNC explanations are properly characterized as mechanistic explanations. Although Chirimuuta develops
an interesting positive proposal centered on what she terms an informational minimal model (or I-minimal model), the
focus of the present chapter is on her negative thesis.
4. Sound localization: Birds
Specific cases of CNCs are too controversial to be useful for the purposes of illustrating how MR considerations are
irrelevant to assessing the adequacy of mechanistic explanations. Instead, I want to focus on one of the most successful
explanatory models in all of systems neuroscience, the neural circuit model of auditory sound localization in the barn
owl. For reasons that will become evident shortly, this is a paradigmatic example of a mechanistic explanation.
Nevertheless, the capacity to localize airborne sounds is implemented in a diversity of neural mechanisms across
different species throughout animal kingdom. Yet, I will argue, this does not in any way compromise the explanatory
status of the model. Nor does it affect the adequacy of any other mechanistic explanations for sound localization in
different systems or species. This example helps to show how MR considerations are irrelevant to assessing mechanistic
explanations.
Many species of birds are capable of accurately localizing sounds on the basis of auditory cues alone such as
during flight in complete darkness. These animals exploit the different arrival times of a sound at the two ears (Figure
2). Although these interaural time differences (ITDs) may only be microseconds apart, birds have evolved an exquisite
11
There is another route to undermining P2 on entirely different grounds from those developed in the present chapter. This strategy involves
challenging the basic premise of multiple realization itself. In particular, following a line of argument first developed by Mundale and Bechtel (1999),
one could argue that the initial plausibility of MR claims (at the heart of the challenge from canonical neural computation) rests on a spurious
mismatch between a coarse-grained functional individuation criterion (for establishing the sameness of the computation being performed) with an
excessively fine-grained structural individuation criterion (for establishing the difference of underlying brain mechanisms). Although it will not be
pursued here, at least for the case of divisive normalization, the granularity objection appears to have teeth.
8
strategy to detect and use ITDs to localize sounds. Unraveling precisely how ITDs are computed in the brain is one of
the great success stories of modern neuroscience.
Figure 2. Sound localization in birds (A) and mammals (B). Both rely on computing the difference between the
arrival times at the two ears (ITD) to localize airborne sounds in the horizontal plane.
More than 50 years ago, the psychologist Lloyd Jeffress proposed an elegant computational model for sound
localization in the horizontal plane (Jeffress 1948). The Jeffress model involves three basic elements: (e1) converging
time- or phase-locked excitatory inputs from both ears, (e2) coincidence detectors that respond maximally when signals
arrive from each ear simultaneously, and (e3) an arrangement of delay lines with systematically varying lengths from
each ear so that different coincidence detectors encode different ITDs (Figure 3a).12 Since neural transmission delay
time is directly proportional to an axon’s length, tuning for different ITDs can be achieved by having axonal “delay
lines” of systematically varying lengths project from each ear onto different individual coincidence detectors. A final,
related detail of the model is that the set of coincidence detectors are topographically organized such that adjacent
coincidence detectors represent adjacent locations in the horizontal plane (Figure 3a, schematically represented by
grayscale coding along the sagittal plane). Strikingly, Jeffress developed the model to account for a body of human
psychophysical data on sound localization, and did so in the absence of information about the nature of the underlying
brain mechanisms.13
Remarkably, all of the major elements of the Jeffress delay line model (e1-e3) have now been confirmed in the
barn owl (Carr and Konishi 1990; Konishi 2003; Pena et al. 2001).14 Careful behavioral, anatomical, and physiological
investigations have revealed a neural circuit for computing ITDs involving delay line-based coincidence detection of
signals from the two ears. More specifically, so-called bushy cells in the left and right nucleus magnocellularis (NM)
send time-locked excitatory inputs from each ear, implementing e1 of the Jeffress model. Neurons in the nucleus
laminaris (NL), the first station of binaural processing in the avian auditory brainstem, are maximally responsive when
ipsilateral and contralateral input signals arrive simultaneously. In other words, these neurons perform coincidence
detection, implementing e2 of the model. Individual NL neurons tuned to the same characteristic frequency show
12
Temporally coincident inputs can be precisely defined in the model as occurring when the sum of acoustic and neural transmission delays
originating from one ear equals that from the other ear: Ai + Ni = Ac + Nc, where A indicates the auditory input signal, N indicates the neural
transmission delay, and subscripts i and c indicate ipsilateral and contralateral, respectively. For further discussion, see Konishi (2003).
13
As the original Jeffress model involved a set of minimally constrained mechanistic conjecture about how sound localization might be performed, it
constitutes a clear example of a how-possibly mechanistic model (Craver 2007; Kaplan 2011).
14
Although the model was originally verified in the owl, extremely similar observations have been confirmed in most other bird species (Grothe and
Pecka 2014).
9
different preferred ITDs in virtue of differences in the axonal conduction delays from each ear.15 More specifically,
since the time delay of neural conduction through an axon is directly proportional to its length, neural tuning for
different ITDs can be achieved by having axons (“delay lines”) of systematically varying lengths project from each ear
onto individual NL neurons. A matrix of NL neurons receive inputs from axons of systematically varying lengths (and
with systematically varying interaural delays) project along the length of the nucleus, implementing e3 of the Jeffress
delay line model (Figure 3a).16 Because different NL neurons have different preferred ITDs, a population code is
required to represent the entire physiologically relevant range of ITDs (Figure 3b)17, and by extension, the entirety of
horizontal auditory space.18
Figure 3. Neural computation of interaural time differences (ITDs) in birds and mammals. (A) Jeffress-type
computational mechanism observed in the nucleus laminaris (NL) of birds involving coincidence detection of
excitatory inputs from the two ears. (B) Distribution of different preferred ITDs across a population of
narrowly-tuned individual neurons in one hemispheric NL. Shaded area indicates physiologically relevant range
of ITDs. (C) Computational mechanism in mammals involving precisely timed hyperpolarizing inhibition that
adjust the timing of excitatory inputs to coincidence detector neurons in the medial superior olive (MSO). (D)
For a given frequency band, the ITD tuning of the population of MSO neurons in the left MSO is the inversion of
that of the corresponding population of neurons in the right MSO. (Source: Grothe 2003)
Critically, the neural circuit model of sound localization developed by Konishi and colleagues is a
paradigmatic example of a mechanistic explanation in computational neuroscience. Its explanatory status is beyond
reproach within the neuroscience community, and more importantly, it exhibits all the hallmarks of an adequate
15
Frequency tuning is observed in NL and in many other auditory neurons mainly because hearing begins when the cochlea mechanically filters
incoming sounds into separate frequency components. Consequently, all output signals from the cochlea are already broken down or filtered
according to their various frequencies.
16
To make the critical role played by delay lines and coincidence detection clear, it is helpful to consider how different arrangements of delays from
the two ears give rise to coincident inputs from different locations in auditory space. For example, axons of equal length projecting from each ear onto
a coincidence detector neuron in NL will have equal internal neural conduction delays (Figure 3a, central column positions); and consequently will
give rise to coincident inputs only when sounds are emitted from straight ahead and reach both cochlea at the same time (i.e., when ITD = 0). Because
these neurons fire maximally for ITDs of zero, they are said to be tuned or have a best or preferred ITD of zero. By contrast, an NL neuron receiving a
short axonal projection from the left (ipsilateral) ear and a long axonal projection from the right (contralateral) ear, will receive coincident inputs and
exhibit tuning only for sounds coming from the right auditory hemifield. This is because signals from the right ear must travel a longer path length
compared to those from the left and therefore have internal transmission delays that precisely offset the delay in the arrival time at the two ears.
Conversely, a NL neuron with a short axonal projection from the right ear and a long axon from the left ear, will receive coincident inputs and exhibit
tuning only for sounds coming from the left auditory hemifield.
17
What comprises the physiologically relevant range of ITDs is determined by factors such overall head size, and more specifically, the distance
between the ears.
18
It turns out that these neurons are also topographically organized in NL, such that neurons in adjacent positions in NL code for spatially adjacent
locations in contralateral auditory space, thereby implementing another detail of the Jeffress model (Figure 3a).
10
mechanistic explanation. In particular, it readily satisfies 3M. There is a clear mapping from elements of the model onto
all key aspects of the target mechanism in the avian auditory system. The parts, activities, and organization identified in
the model are implemented by corresponding structures, activities, and organizational features in the avian brainstem.
NL neurons function as coincidence detectors and axons from NM neurons serve as delay lines. Different individual
coincidence detector neurons in NL exhibit different response or tuning properties such that across the population the
full range of physiologically relevant ITDs are encoded. Finally, the temporal and spatial organization depicted in the
model is precisely reproduced in the avian brainstem circuitry. As described in detail above, the timing of excitatory
inputs received by individual NL coincidence detector neurons reflects the exquisite spatial organization of axons (i.e.,
component parts) that are systematically arranged such that their lengths generate neural transmission delays that
precisely offset specific ITDs. This organization is essential to their ITD tuning properties.
One way of elucidating why mechanistic models have explanatory force with respect to the phenomenon they
are used to explain appeals to the same sorts of considerations that one might plausibly appeal to for other kinds of
causal explanations. Along these lines, adequate mechanistic explanations allow us to answer a range of what-if-thingshad-been-different questions (or w-questions) just as causal explanations do (Kaplan 2011; Kaplan and Craver 2011;
Woodward 2003). Purely descriptive models may either fail to provide answers to w-questions and offer no explanation
at all, or answer only a very restricted range of w-questions and offer superficial explanations.. By contrast, deeper
mechanistic explanations afford answers to a broader range of w-questions concerning interventions on the target
mechanism than do more superficial ones. The model for sound localization in the barn owl is an example of a deeper
mechanistic explanation. The model allows us to answer a multitude of w-questions. For example, how would the
response profile of a given coincidence detector neuron in NL change if we intervened to vary the length of either the
contralateral or ipsilateral axonal delay line through which it receives its inputs? How would the neural circuit perform
if excitatory inputs from one of the two ears were completely or partially eliminated? Which individual coincidence
detector neuron in NL would respond if we artificially set the ITD value of a sound to x microseconds? And so on.
Mechanistic models explain because they deliver answers to these and other w-questions.
5. Sound localization: Mammals
For many years, it was thought that all birds and mammals perform sound localization by relying on the same Jeffresstype mechanism involving coincidence detection of excitatory inputs coming from the two ears. However, mounting
evidence now suggests that mammals compute ITDs and thereby localize sounds using a different underlying
mechanism. It is becoming increasingly clear that mammals (unlike birds) do not rely on a population of neurons with a
precise arrangement of axonal delay lines from the two ears in order to compute ITDs and perform auditory localization
(Ashida and Carr 2011; Brand et al. 2002; Grothe 2003; McAlpine and Grothe 2003; McAlpine, Jiang, and Palmer
2001; Myoga et al. 2014)19 Instead, the emerging picture attributes a major role for synaptic inhibition in the processing
of ITDs (Figure 3c). More specifically, precisely timed hyperpolarizing inhibition controls the timing of excitatory
inputs reaching binaural coincidence detector neurons in the mammalian auditory brainstem structure known medial
superior olive (MSO).20 Inhibition slows the transmission of excitatory inputs to the MSO in such a way as to precisely
offset the difference in arrival time at the two ears arising from the specific location of a sound source. This altered
temporal sensitivity of binaural neurons in the MSO provides the basis for ITD computation. The fact that excitatory
19
The synaptic inhibition model has primary been described in gerbils and guinea pigs. However, there is evidence that similar neural mechanisms for
ITD computation are at work in other mammals including cats and possibly even humans. For additional discussion, see Grothe (2003) and
Thompson et al. (2006).
20
MSO receives bilateral excitatory inputs from so-called spherical bushy cells (SBCs) in both ventral cochlear nuclei (VCNs). For additional
discussion of the neural circuit underlying sound localization in mammals, see Grothe (2003).
11
inputs from both ears would reach the MSO without any significant interaural conduction delays (and thus would
always coincide at ITDs of zero) in the absence of inhibition, clarifies its role in ITD computation. Mammalian sound
localization therefore reflects the convergence of bilateral excitatory and exquisitely timed inhibitory inputs onto
coincidence detector neurons in the MSO (Figure 3c).
The mechanism underlying ITD computation and sound localization in mammals differs from the mechanism
observed in birds in several major ways. First, even though the inhibitory mechanism involves similarly functioning
parts (i.e., neurons serving as coincidence detectors), the functional activities – tuning properties – of coincidence
detector neurons in MSO are fundamentally different from those observed in the avian auditory system. Specifically,
individual MSO neurons do not have different preferred ITDs distributed across the entire relevant range of ITDs, as is
found in birds (Figure 3b). In mammals, all MSO neurons tuned to the same frequency band within each hemispheric
MSO exhibit the same ITD tuning, and horizontal sound location is read out from the population-averaged firing rate
across two broadly tuned spatial channels – one for each hemispheric MSO (Figure 3d). What this means is that a
change in the horizontal position of a sound source will induce a specific pattern of change in population activity in one
hemisphere and an corresponding change of opposite sign in the population activity in the other hemisphere. For
example, a sound moving away from the midline (where ITD = 0), might serve to increase activity in the contralateral
MSO and correspondingly decrease activity in the ipsilateral MSO, thereby indicating that the sound source has shifted
to a more lateral position. It is therefore the relative difference between the two hemispheric channels (the relative
activity across the entire population of MSO neurons) that encodes ITD information and indicates the horizontal
location of a sound source. Because individual MSO neurons within each hemisphere carry similar information, this
argues against the local coding strategy observed in birds in which each individual neuron encodes information about
different ITDs, and instead strongly implies that a population code is used to represent ITDs (Lesica, Lingner, and
Grothe 2010; McAlpine, Jiang, and Palmer 2001).
Second, the mammalian mechanism for ITD computation involves different parts doing different things. As
indicated above, in addition to the excitatory projections originating from the left and right cochlea, MSO neurons also
receive bilateral inhibitory projections from other structures in the auditory system that are highly specialized to
preserve the fine temporal structure of auditory stimuli with high precision.21 These structures generate temporally
accurate patterns of inhibition that precisely control the timing of excitatory inputs reaching the MSO.
Third, and perhaps most obviously, certain key parts and properties of the mechanism for computing ITDs in
birds are simply missing from the inhibitory mechanism found in mammals. Specifically, there are no axons serving as
delay lines. Axon lengths are roughly equivalent such that excitatory inputs from both ears would reach the MSO
without any appreciable interaural conduction delay in the absence of inhibition, and thus always coincide at ITDs of
zero. It is only through precise inhibitory control over excitatory timing that appropriate ITD tuning is achieved.
Despite these profound differences in underlying mechanisms, the inhibitory control model similarly provides
an adequate mechanistic explanation of sound localization in mammals. Like the delay-line model, it straightforwardly
satisfies 3M. The inhibitory control model describes the parts, activities, and organizational features of the target
mechanism in the mammalian nervous system underlying ITD computation and sound localization.
6 Multiple realizability, scope, and mechanistic explanation
21
Glycinergic neurons in the medial nucleus of the trapezoid body (MNTB) provide the main source of hyperpolarizing inhibition to the mammalian
MSO. For further discussion, see Grothe (2003).
12
The upshot of the preceding discussion should now be clear. Together, the cases described above provide a strong
counterexample to P2 in Chirimuuta’s argument – the claim that if a given phenomenon is multiply realizable, then
models of that phenomenon will fail to satisfy 3M. Here we have just seen that the phenomenon of ITD computation (or
sound localization) is multiply realized, and yet both the delay line model for birds and the inhibitory control model for
mammals each individually satisfy the requirements of 3M. Both provide adequate mechanistic explanations. Indeed,
both are widely regarded as paradigmatic instances of mechanistic explanations. Hence, at least for these cases, MR
appears irrelevant to assessing their explanatory status. This result should be prima facie puzzling given the strong
parallels between these models and models of CNCs. Before returning to consider CNCs, it will help to explore a bit
further how these argumentative moves play out in the context of models of ITD computation canvassed above.
Suppose that instead of being satisfied that both the delay line model and the inhibitory control model
independently provide adequate mechanistic explanations of sound localization (because each independently satisfies
3M), an additional requirement is imposed to the effect that there must be a single, unifying explanation to cover both
birds and mammals since both perform the same computation (i.e., ITD computation). To be clear, what is being
demanded here is a wide-scope explanation that unifies or subsumes all known systems in which a given computation is
performed. Although a detailed comparison goes well beyond the purview of the current chapter, this requirement
reflects deep similarities with unificationist accounts of explanation (e.g., Kitcher 1981). The general idea behind
unificationist views is that scientific explanation is a matter of providing a unified account of a range of different
phenomena. For ease of reference, we might therefore call this additional requirement or principle the unification
principle (UP). Because no sufficiently wide-scope or unifying mechanistic explanation will be forthcoming in the case
of ITD computation – the underlying mechanisms in birds and mammals are known to differ, after all – taking UP on
board entails the rejection of the mechanistic approach in favor of some other non-mechanistic form of explanation. But
is this inference justified? The short answer is “no”.
There are two things wrong with this line of argument. First, it trades on a confusion over the proper scope of
mechanistic explanations, and subsequently, neglects the important possibility of constructing “local” or narrow-scope
mechanistic explanations. Second, it problematically assumes that more unifying models – those with wider scope –
will always be explanatorily superior to those that are less unifying. I will address each in turn.
6.1 Scope and mechanistic explanation
Attempts to impose UP reflect a deep confusion over the proper scope of mechanistic explanations. Scope concerns
how many systems or how many different kinds of systems there actually are to which a given model or generalization
applies. In the present context, UP implies that models of ITD computation cannot successfully explain unless they
have relatively wide scope – i.e., unless there is a model (mechanistic or otherwise) that uniformly captures the common
computation performed across multiple neural circuits and taxa. Critically, although it is true that wide-scope
mechanistic explanations will be unavailable if MR obtains, it does not follow that the mechanistic approach must be
abandoned (and some other explanatory scheme embraced), as is implied by P2 in Chirimuuta’s argument. This is
because wide-scope explanations are not the only type of mechanistic explanation that can be provided for neural
computations. In particular, as the cases of ITD computation canvassed above clearly indicate, perfectly adequate
mechanistic explanations can be constructed that capture some but not necessarily all instances in which the same ITD
computation is being performed. The delay line model provides an adequate mechanistic explanation even though its
scope only includes the biological taxa of birds but not mammals. Similarly, the inhibitory control model provides an
adequate mechanistic explanation even though its scope only includes mammals but not birds. These models explain
how a given computation is implemented in a specific system or type of system without covering all systems in which
13
the computation is implemented. Crucially, they show how the scope of mechanistic models can be relatively narrow or
“local” without this restricting or compromising their explanatory power. They are perfectly adequate for the range of
systems they do in fact cover. Why?
Although scope is a dimension along which mechanistic models can and do vary (Craver and Darden 2013), it
is not an appropriate norm for evaluating mechanistic explanations. This is implicit in the 3M principle. 3M clarifies
how mechanistic models are to be judged as explanatorily adequate based on the extent to which they accurately and
completely describe the causal structure of the mechanism — the component parts, their activities, and organization —
responsible for the phenomenon of interest (Kaplan 2011; Kaplan and Craver 2011). The quality of the explanation
provided does not in any way depend on how widely the model applies. Consequently, our assessments of mechanistic
explanations should be insensitive to scope.
Despite the fact that adequate mechanistic explanations can and frequently do describe highly conserved
biological or neural mechanisms (e.g., DNA or the selectivity filter mechanism in potassium channels; Doyle et al.
1998; MacKinnon et al. 1998), resulting in wide-scope mechanistic explanations, this is not required. Instead, just like
the mechanistic explanations of ITD computation described above, the scope of explanatory mechanistic models is
often highly restricted in biological and neurobiological contexts. Along these lines, Bechtel (2009) maintains that
mechanistic accounts are often “highly particularized” in that they “characterize the mechanism responsible for a
particular phenomenon in a specific biological system” (Bechtel 2009, 762). At the limit, a mechanistic model might
aim to capture only a single case (e.g., the unique mechanism responsible for producing the specific pattern of cognitive
deficits exhibited by patient H.M.; Annese et al. 2014) or a recurring mechanism found in only one species (e.g., the
exotic mechanism underlying color vision in the mantis shrimp; Thoen et al. 2014). These “highly particularized” or
scope-restricted models can nonetheless provide perfectly adequate mechanistic explanations – by satisfying 3M – even
though their application may be restricted to extremely narrow domains.
Of course, things could have turned out differently in the cases of ITD computation surveyed above due to
differences in patterns of evolution, drift, or both. Under some counterfactual evolutionary scenarios, a common circuit
design across these taxa might have resulted instead of the actual divergent mechanisms observed today. Although these
differences would invariably change the scope of our computational models, such alterations in scope are (or at any rate
should be) inconsequential to how we evaluate these models as mechanistic explanations.
6.2 Unification and explanation
This deflationary view about the relationship between model scope and explanatory power is by no means unique or
idiosyncratic to the mechanistic framework. Although the view that wide scope is not an important property of
explanatory models places the mechanistic approach at odds with traditional covering law and unificationist accounts of
explanation, interventionist approaches to causal explanation similarly reject scope as a reliable indicator of explanatory
depth or power (Hitchcock and Woodward 2003; Woodward 2003). According to the interventionist perspective,
explanatory depth reflects the extent to which a given generalization or model provides resources for answering a
greater range of what-if-things-had-been-different questions about the phenomenon of interest. This in turn is argued to
depend on how wide the range of interventions is under which a given generalization is invariant. One generalization or
model thus provides a better or “deeper” explanation than another because it is invariant under a wider range of
interventions than the other, not because it has wider scope. Hence, both mechanistic and interventionist perspectives
provide arguments for thinking that scope and explanatory power can and often do vary independently of one another.
Along similar lines, Sober (1999) rejects the assumption underlying UP, namely, that relatively abstract,
unifying explanations are always better and are to be preferred over more detailed, disunified explanations. In
14
particular, he contends that there is “no objective reason ” (p.551) to prefer unified over disunified explanations.
According to unificationists like Kitcher, explanations involving an abundance of “micro-details” (or mechanistic
details) are “less unified” than explanations that abstract from such detail because the latter and not the former can more
easily apply across a wide range of systems with different underlying realizations. Sober argues that it is the nature of
our explanatory interests that determines which type we will prefer in any given situation. In some explanatory contexts,
highly detailed, narrow-scope explanations may be better or more appropriate. In other contexts, more abstract, widescope explanations may be preferable. Crucially, there is no automatic preference or requirement for explanations
involving abstraction away from micro-detail, as is implied by UP.
7. Canonical neural computation explanations reconsidered
We now have the appropriate resources to reconsider the claim that CNCs cannot be analyzed in mechanistic terms and
therefore require a distinct explanatory framework. In particular, we can now see how arguments against the
mechanistic analyses of CNCs trade on precisely the same confusion as the one identified above concerning the proper
scope of mechanistic explanations. Specifically, these arguments start by pointing out that wide-scope mechanistic
explanations of CNCs will invariably fail in virtue of the diversity of mechanisms across different neural circuits and
species that implement the computation. From this, together with an assumed premise embodying something highly
similar to UP, a conclusion is drawn to the effect that CNCs cannot be assimilated into the mechanistic framework
(Chirimuuta 2014). However, for reasons canvassed above, this argumentative strategy assumes that only wide-scope
explanations are appropriate, and it consequently neglects the possibility of developing narrow-scope mechanistic
explanations of computational phenomena. Yet, as the discussion of ITD computation shows, narrow-scope mechanistic
explanations can in principle be provided, even in contexts where MR is operative.
After reviewing evidence that divisive normalization computations are implemented differently across
different systems, Carandini and Heeger (2012) claim that: “it is unlikely that a single mechanistic explanation will hold
across all systems and species: what seems to be common is not necessarily the biophysical mechanism but rather the
computation” (2012, 58). On a strong reading, Carandini and Heeger (2012) appear to embrace UP. Chirimuuta
reinforces this interpretation when she claims that they “no longer think that we should interpret the normalization
model mechanistically, as describing components and dynamics of a neural mechanism” (Chirimuuta 2014, 140). She
interprets them as embracing the view that unless a wide-scope explanation of divisive normalization is forthcoming, all
manner of mechanistic explanation will be out of reach. Independently of whether or not this is the correct interpretation
of Carandini and Heeger’s view22, Chirimuuta explicitly endorses this position as her own. She takes the fact that no
single mechanistic explanation can be provided that subsumes all divisive normalization phenomena – i.e., that
describes the full range of circuits and systems that perform the same (divisive normalization) computation – as
evidence that “[s]uch explanations cannot be assimilated into the mechanistic framework” (2014, 127). But this
assumption of UP is just as problematic in the context of models of CNCs as it was in the context of models of sound
localization.
As was true in the case of ITD computation, scope-restricted explanations of divisive normalization are either
22
Carandini and Heeger’s (2012) claim is compatible with both a weaker and stronger reading, only the latter of which implies a commitment to UP.
According to the weaker reading, although a single mechanistic explanation is unavailable, multiple individual mechanistic explanations may still be
available or forthcoming. On the stronger reading, the fact that no single mechanistic explanation can be given implies that no mechanistic
explanation of any kind is available or forthcoming. Although determining which reading is more closely aligned with their actual view goes beyond
the scope of this chapter, based on the broader context, they appear to endorse the weaker reading. Specifically, they repeatedly emphasize how
research into canonical neural computations such as divisive normalization involves the discovery and articulation of underlying mechanisms, stating
that “[a] key set of questions concerns the circuits and mechanisms that result in normalization” and how “[u]nderstanding these circuits and
mechanisms is fundamental” (2012, 60).
15
currently available or forthcoming. 23 For example, neuroscientists have already developed a robust narrow-scope
mechanistic explanation for how divisive normalisation computations are performed in the fruitfly (Drosophila)
olfactory system There is strong evidence that, in this specific system, normalization is implemented by GABAmediated inhibition of presynaptic connections between neurons in the fly antennal lobe (Olsen and Wilson 2008;
Olsen, Bhandawat, and Wilson 2010). Similarly robust narrow-scope explanations have been developed for divisive
normalization in the fly visual system and the mammalian retina, and the mechanisms thought to be involved differ in
important respects. Although uncertainty remains about the specific mechanism underlying divisive normalization in
mammalian primary visual cortex (V1) – the original phenomenon for which the normalization model was developed –
candidate narrow-scope mechanistic explanations of this phenomenon continue to be the subject of intense experimental
investigation in contemporary neuroscience. One plausible narrow-scope explanation posits a form of shunting
inhibition – increases in membrane conductance without the introduction of depolarising or hyperpolarizing synaptic
currents – as the mechanism underlying divisive normalization in V1 (Chance, Abbott, and Reyes 2002). Other narrowscope explanations posit mechanisms of synaptic depression (Abbott et al. 1997) or intrinsic neural noise (Carandini
2007; Finn, Priebe, and Ferster 2007). The fact that we do not at present have a complete mechanistic explanation for
this phenomenon, or that the evidence for one plausible mechanism over another remains mixed, is inconsequential.
What matters is that progress is being made, or can in principle be made, towards suitable narrow-scope mechanistic
explanations of divisive normalization phenomena and other kinds of CNCs more generally. And we should have every
reason to think that the question of how each of these neural systems performs computations will ultimately succumb to
this kind of mechanistic analysis.
Assimilating CNCs into the mechanistic framework does not require, as Chirimuuta and others have argued,
the provision of a single mechanistic explanation to cover all instances in which one and the same computation is
performed – a wide-scope mechanistic explanation – but rather only requires the provision of multiple suitable narrowscope explanations. Once recognized, the irrelevance of MR considerations becomes clear. For narrow-scope
mechanistic explanations of computational phenomena, the objective is to reveal the particular implementing
mechanism underlying a given neural computation in a specific system (or type of system), independently of whether
multiple other possible mechanisms also might have performed the computation, or whether other actual mechanisms
might be performing the same computation in other neural systems or in other species.
8. Conclusion
As the mechanistic perspective comes to occupy a dominant position in philosophical thinking about explanation across
the biological sciences including all areas of neuroscience, questions will naturally arise about its scope and limits.
Along these lines, some authors have recently appealed to modeling work involving canonical neural computations to
argue that some explanations in neuroscience fall outside the bounds of the mechanistic approach. Because these neural
computations can rely on diverse circuits and mechanisms, modeling the underlying mechanisms is supposed to be of
limited explanatory value. At its core, this is a challenge from multiple realizability. In this chapter, I argue that these
anti-mechanistic conclusions about canonical neural computation explanations are mistaken, and rest upon confusions
about the proper scope of mechanistic explanation and the relevance of multiple realizability considerations. More
specifically, I maintain that this confusion stems from a failure to appreciate how scope is not an appropriate norm on
23
For reasons that go beyond the scope of this chapter to address, the underlying mechanisms cited in these explanations of normalization are highly
similar in many respects and bear a family resemblance to one another. It seems entirely plausible to think that even though there is no single common
normalization mechanism, the fact that all these systems perform a common computational operation nevertheless places some important constraints
on the nature of the underlying mechanisms. Consequently, we might reasonably expect a family of highly similar mechanisms (and in turn,
mechanistic explanations) whenever a common computational operation is being performed across systems.
16
mechanistic explanation and therefore even mechanistic models with highly restricted scope can in principle be
explanatorily adequate. Once appreciated, it becomes clear how mechanistic models of computational phenomena can
provide legitimate explanations, even in the presence of multiple realizability.
Although there is a long history of wielding multiple realizability considerations to successfully challenge
dominant philosophical positions including mind-brain type identity theory and reductionism in biology, throughout this
chapter I have argued that such considerations does not create any insuperable obstacles for adopting a mechanistic
perspective about computational explanation in neuroscience.
References
Abbott, L. F., J. A. Varela, K. Sen, and S. B. Nelson. 1997. “Synaptic Depression and Cortical Gain Control.” Science
(New York, N.Y.) 275 (5297): 220–24.
Annese, J., Schenker-Ahmed, N. M., Bartsch, H., Maechler, P., Sheh, C., Thomas, N., … Corkin, S. (2014).
examination of patient H.M.’s brain based on histological sectioning and digital 3D reconstruction. Nature
Communications, 5, 3122.
Ashida, Go, and Catherine E. Carr. 2011. “Sound Localization: Jeffress and beyond.” Current Opinion in Neurobiology
21 (5): 745–51.
Ayala, Francisco J. 1968. “Biology as an Autonomous Science.” American Scientist, 207–21.
Batterman, Robert W. 2002. The Devil in the Details: Asymptotic Reasoning in Explanation, Reduction, and
Emergence. Oxford University Press Oxford.
Bechtel, W. 2008. Mental Mechanisms: Philosophical Perspectives on Cognitive Neuroscience. Lawrence Erlbaum.
Bechtel, William. 2009. “Generalization and Discovery by Assuming Conserved Mechanisms: Cross Species Research
on Circadian Oscillators.” Philosophy of Science, 762–73.
Bechtel, William, and Jennifer Mundale. 1999. “Multiple Realizability Revisited: Linking Cognitive and Neural
States.” Philosophy of Science, 175–207.
Bonin, Vincent, Valerio Mante, and Matteo Carandini. 2005. “The Suppressive Field of Neurons in Lateral Geniculate
Nucleus.” The Journal of Neuroscience 25 (47): 10844–56.
Brand, Antje, Oliver Behrend, Torsten Marquardt, David McAlpine, and Benedikt Grothe. 2002. “Precise Inhibition Is
Essential for Microsecond Interaural Time Difference Coding.” Nature 417 (6888): 543–47.
Carandini, Matteo. 2007. “Melting the Iceberg: Contrast Invariance in Visual Cortex.” Neuron 54 (1): 11–13.
doi:10.1016/j.neuron.2007.03.019.
———. 2012. “From Circuits to Behavior: A Bridge Too Far?” Nature Neuroscience 15 (4): 507–9.
doi:10.1038/nn.3043.
Carandini, Matteo, and David J. Heeger. 1994. “Summation and Division by Neurons in Primate Visual Cortex.”
Science 264 (5163): 1333–36.
———. 2012. “Normalization as a Canonical Neural Computation.” Nature Reviews Neuroscience 13 (1): 51–62.
Carandini, Matteo, David J. Heeger, and J. Anthony Movshon. 1997. “Linearity and Normalization in Simple Cells of
the Macaque Primary Visual Cortex.” The Journal of Neuroscience 17 (21): 8621–44.
Carr, C. E., and M. Konishi. 1990. “A Circuit for Detection of Interaural Time Differences in the Brain Stem of the
Barn Owl.” The Journal of Neuroscience 10 (10): 3227–46.
Chance, Frances S., L. F. Abbott, and Alex D. Reyes. 2002. “Gain Modulation from Background Synaptic Input.”
Neuron 35 (4): 773–82.
Chirimuuta, M. 2014. “Minimal Models and Canonical Neural Computations: The Distinctness of Computational
Explanation in Neuroscience.” Synthese 191 (2): 127–53.
Craver, Carl F., and Lindley Darden. 2013. In Search of Mechanisms: Discoveries across the Life Sciences. Chicago  ;
London: University Of Chicago Press.
Craver, C. F. Explaining the Brain. Mechanisms and the Mosaic Unity of Neuroscience 2007. New York: Oxford
University Press.
Cummins, Robert. 2000. “How Does It Work?‘ Versus’ What Are the Laws?": Two Conceptions of Psychological
Explanation.” Explanation and Cognition, 117–44.
Feest, Uljana. 2003. “Functional Analysis and the Autonomy of Psychology.” Philosophy of Science 70 (5): 937–48.
Finn, Ian M., Nicholas J. Priebe, and David Ferster. 2007. “The Emergence of Contrast-Invariant Orientation Tuning in
Simple Cells of Cat Visual Cortex.” Neuron 54 (1): 137–52. doi:10.1016/j.neuron.2007.02.029.
Fodor, Jerry. 1997. “Special Sciences: Still Autonomous after All These Years.” Noûs 31 (s11): 149–63.
Fodor, Jerry A. 1974. “Special Sciences (or: The Disunity of Science as a Working Hypothesis).” Synthese 28 (2): 97–
115.
Freeman, Tobe CB, Séverine Durand, Daniel C. Kiper, and Matteo Carandini. 2002. “Suppression without Inhibition in
Visual Cortex.” Neuron 35 (4): 759–71.
Glymour, Clark. 1970. “On Some Patterns of Reduction.” Philosophy of Science 37 (3): 340–53.
17
Grothe, Benedikt. 2003. “New Roles for Synaptic Inhibition in Sound Localization.” Nature Reviews Neuroscience 4
(7): 540–50.
Grothe, Benedikt, and Michael Pecka. 2014. “The Natural History of Sound Localization in Mammals–a Story of
Neuronal Inhibition.” Frontiers in Neural Circuits 8.
http://www.ncbi.nlm.nih.gov/pmc/articles/PMC4181121/.
Heeger, David J. 1992. “Normalization of Cell Responses in Cat Striate Cortex.” Visual Neuroscience 9 (02): 181–97.
Hitchcock, Christopher, and James Woodward. 2003. “Explanatory Generalizations, Part II: Plumbing Explanatory
Depth.” Noûs 37 (2): 181–99.
Hoyningen-Huene, Paul. 1989. “Epistemological Reductionism in Biology: Intuitions, Explications, and Objections.” In
Reductionism and Systems Theory in the Life Sciences, 29–44. Springer.
http://link.springer.com/chapter/10.1007/978-94-009-1003-4_3.
Hubel, David H., and Torsten N. Wiesel. 1968. “Receptive Fields and Functional Architecture of Monkey Striate
Cortex.” The Journal of Physiology 195 (1): 215–43.
Jeffress, Lloyd A. 1948. “A Place Theory of Sound Localization.” Journal of Comparative and Physiological
Psychology 41 (1): 35.
Johnson-Laird, Philip N. 1983. Mental Models: Towards a Cognitive Science of Language, Inference, and
Consciousness. 6. Harvard University Press.
Kaplan, David Michael. 2015. “Explanation and Levels in Cognitive Neuroscience.” In Handbook of Neuroethics,
edited by Jens Clausen and Neil Levy, 9–29. Springer Netherlands.
http://link.springer.com/referenceworkentry/10.1007/978-94-007-4707-4_4.
Kaplan, David Michael, and Carl F. Craver. 2011. “The Explanatory Force of Dynamical and Mathematical Models in
Neuroscience: A Mechanistic Perspective*.” Philosophy of Science 78 (4): 601–27.
Kaplan, D.M. 2011. “Explanation and Description in Computational Neuroscience.” Synthese, 1–35.
Keeley, Brian L. 2000. “Shocking Lessons from Electric Fish: The Theory and Practice of Multiple Realization.”
Philosophy of Science, 444–65.
Kitcher, P. (1981). Explanatory unification. Philosophy of Science, 507–531.
Kitcher, Philip. 1984. “1953 and All That. A Tale of Two Sciences.” The Philosophical Review, 335–73.
Konishi, Masakazu. 2003. “Coding of Auditory Space.” Annual Review of Neuroscience 26 (1): 31–55.
Lesica, Nicholas A., Andrea Lingner, and Benedikt Grothe. 2010. “Population Coding of Interaural Time Differences in
Gerbils and Barn Owls.” The Journal of Neuroscience: The Official Journal of the Society for Neuroscience 30
(35): 11696–702. doi:10.1523/JNEUROSCI.0846-10.2010.
Machamer, Peter, Lindley Darden, and Carl F. Craver. 2000. “Thinking about Mechanisms.” Philosophy of Science, 1–
25.
Mante, Valerio, Robert A. Frazor, Vincent Bonin, Wilson S. Geisler, and Matteo Carandini. 2005. “Independence of
Luminance and Contrast in Natural Scenes and in the Early Visual System.” Nature Neuroscience 8 (12):
1690–97.
McAlpine, David, and Benedikt Grothe. 2003. “Sound Localization and Delay Lines–do Mammals Fit the Model?”
Trends in Neurosciences 26 (7): 347–50.
McAlpine, David, Dan Jiang, and Alan R. Palmer. 2001. “A Neural Code for Low-Frequency Sound Localization in
Mammals.” Nature Neuroscience 4 (4): 396–401.
Myoga, Michael H., Simon Lehnert, Christian Leibold, Felix Felmy, and Benedikt Grothe. 2014. “Glycinergic
Inhibition Tunes Coincidence Detection in the Auditory Brainstem.” Nature Communications 5: 3790.
doi:10.1038/ncomms4790.
Nagel, Ernest. 1961. The Structure of Science: Problems in the Logic of Scientific Explanation. Vol. 1. Harcourt, Brace
& World New York.
Olsen, Shawn R., Vikas Bhandawat, and Rachel I. Wilson. 2010. “Divisive Normalization in Olfactory Population
Codes.” Neuron 66 (2): 287–99. doi:10.1016/j.neuron.2010.04.009.
Pena, Jose Luis, Svenja Viete, Kazuo Funabiki, Kourosh Saberi, and Masakazu Konishi. 2001. “Cochlear and Neural
Delays for Coincidence Detection in Owls.” The Journal of Neuroscience 21 (23): 9455–59.
Piccinini, Gualtiero, and Oron Shagrir. 2014. “Foundations of Computational Neuroscience.” Current Opinion in
Neurobiology 25: 25–30.
Polger, Thomas W. 2009. “Evaluating the Evidence for Multiple Realization.” Synthese 167 (3): 457–72.
Putnam, Hilary. 1975. “Mind, Language and Reality (Vol. 2).” Philosophical Papers.
Rabinowitz, Neil C., Ben DB Willmore, Jan WH Schnupp, and Andrew J. King. 2011. “Contrast Gain Control in
Auditory Cortex.” Neuron 70 (6): 1178–91.
Reynolds, John H., and David J. Heeger. 2009. “The Normalization Model of Attention.” Neuron 61 (2): 168–85.
Reynolds, John H., Tatiana Pasternak, and Robert Desimone. 2000. “Attention Increases Sensitivity of V4 Neurons.”
Neuron 26 (3): 703–14.
Rosenberg, Alex. 2001. “Reductionism in a Historical Science.” Philosophy of Science, 135–63.
Rust, Nicole C., Valerio Mante, Eero P. Simoncelli, and J. Anthony Movshon. 2006. “How MT Cells Analyze the
Motion of Visual Patterns.” Nature Neuroscience 9 (11): 1421–31.
Sarkar, Sahotra. 1992. “Models of Reduction and Categories of Reductionism.” Synthese 91 (3): 167–94.
Schaffner, Kenneth F. 1967. “Approaches to Reduction.” Philosophy of Science, 137–47.
18
Shapiro, Lawrence A. 2000. “Multiple Realizations.” The Journal of Philosophy 97 (12): 635–54.
———. 2008. “How to Test for Multiple Realization.” Philosophy of Science 75 (5): 514–25.
Simoncelli, Eero P., and David J. Heeger. 1998. “A Model of Neuronal Responses in Visual Area MT.” Vision
Research 38 (5): 743–61.
Sober, Elliott. 1999. “The Multiple Realizability Argument against Reductionism.” Philosophy of Science, 542–64.
Thoen, H. H., How, M. J., Chiou, T.-H., & Marshall, J. (2014). A different form of color vision in mantis shrimp.
Science (New York, N.Y.), 343(6169), 411–413.
Thompson, Sarah K., Katharina von Kriegstein, Adenike Deane-Pratt, Torsten Marquardt, Ralf Deichmann, Timothy D.
Griffiths, and David McAlpine. 2006. “Representation of Interaural Time Delay in the Human Auditory
Midbrain.” Nature Neuroscience 9 (9): 1096–98. doi:10.1038/nn1755.
Waters, C. Kenneth. 1990. “Why the Anti-Reductionist Consensus Won’t Survive: The Case of Classical Mendelian
Genetics.” In PSA: Proceedings of the Biennial Meeting of the Philosophy of Science Association, 125–39.
JSTOR. http://www.jstor.org/stable/192698.
Woodward, J. 2000. “Explanation and Invariance in the Special Sciences.” The British Journal for the Philosophy of
Science 51 (2): 197–254.
———. 2005. Making Things Happen: A Theory of Causal Explanation. Oxford University Press, USA.
Wright, Cory D., and William Bechtel. 2006. “Mechanisms and Psychological Explanation.” Handbook of Philosophy
of Psychology and Cognitive Science, 31–79.
Zoccolan, Davide, David D. Cox, and James J. DiCarlo. 2005. “Multiple Object Response Normalization in Monkey
Inferotemporal Cortex.” The Journal of Neuroscience 25 (36): 8150–64.
19