E X P L A N AT I O N I N

C O M P U TAT I O N A L
NEUROSCIENCE: CAUSAL
AND NON-CAUSAL
m. chirimuuta*
February 3, 2016

contents
1

2
3
4

Introduction
1.1 Efficient Coding Explanation in Computational Neuroscience
1.2 Defining Non-causal Explanation . . . . . . . . . . . . . . . . .
Case I: Hybrid Computation
Case II: the Gabor Model Revisited
Case III: A Dynamical Model of Prefrontal Cortex
4.1 A New Explanation of Context-Dependent Computation . . .
4.2 Causal or Non-causal? . . . . . . . . . . . . . . . . . . . . . . .
Causal and Non-causal: does the difference matter?

2
3
4
7
14
16
16
17
22

abstract
This paper examines three candidate cases of non-causal explanation in
computational neuroscience. I argue that there are instances of efficient
coding explanation which are strongly analogous to examples of non-causal
explanation in physics and biology, as presented by Batterman (2002), Woodward (2003) and Lange (2013). By integrating Langes and Woodwards
accounts I offer a new way to elucidate the distinction between causal and
non-causal explanation, and to address concerns about the explanatory
sufficiency of non-mechanistic models in neuroscience. I also use this
framework to shed light on the dispute over the interpretation of dynamical
models of the brain.

* [email protected]. Penultimate Version. Forthcoming in BJPS.

introduction

introduction

In recent philosophy of neuroscience and cognitive science there has been

an overriding emphasis on mechanisms and mechanistic explanation1 and
some authors have gone so far as to present the mechanistic approach as the
only game in town when it comes to explaining various phenomenafrom
sub-cellular signalling to computation in neural circuits and person-level
decision making.2 Following Woodward (forthcoming) I treat mechanistic
explanation as a sub-type of causal explanation characterized by features
such as reliance on decompositional methods and sensitivity to the details
of implementation. A working assumption of the mechanist brand of
philosophy of neuroscience has been that to explain a phenomenon is to
describe the parts of the causal nexus that give rise to it.
Not surprisingly, a number of authors have detected a greater degree
of methodological and explanatory pluralism amongst the sciences of the
mind-brain.3 On a number of occasions it has been argued that applications
of dynamical systems theory (DST) yield non-mechanistic explanatory
models.4 In particular, Ross (2015) has argued that the explanatory
patterns employed in highly abstract, dynamical models of spiking neurons
(Izhikevich, 2010) are of the same form as the minimal model explanations
first described by Batterman (2002) in the context of statistical mechanics.
The published position of the mechanists has been that dynamical models
are either not explanatory but merely predictive and descriptive (Kaplan
and Craver, 2011), or that when they do offer explanations these are
of a mechanistic sort [Bechtel (2011), Kaplan and Bechtel (2011)]. The
debate over DST in neuroscience is far from settled. Yet if one does
accept that dynamical models exemplify minimal model explanation, this
raises the interesting question of whether such explanations are also noncausal. Robert Batterman and Collin Rice have argued that minimal model
explanations in physics and biology are non-causal,5 whereas Silberstein
and Chemero (2013) hold that DST offers neuroscientists and cognitive
scientists a distinctive genre of causal explanation.6
My central claim in this paper is that there are indeed instances of noncausal explanation to be found within neuroscience. I aim to convince
you that the cases in neuroscience are at least strong as the familiar
examples from physics and biology. Along the way, I will elucidate the
distinction between causal and non-causal explanation by developing James
Woodwards proposal that non-causal explanation occurs when we are able
to answer questions about non-actual scenarios (what-if-things-had-beendifferent or w- questions), even though those scenarios are not the
consequence of hypothetical interventions.
In Sections 2 and 3, I will focus on efficient coding explanation
in computational neuroscience where, I will argue, one finds numerous
1 See e.g. Machamer et al. (2000), Craver (2007), Bechtel (2008), Kaplan and Craver (2011), Kaplan
(2011), Kaplan and Bechtel (2011), Levy (2014) amongst many others.
2 Kaplan (2011); Piccinini and Craver (2011); Piccinini and Bahar (2013).
3 E.g. Weiskopf (2011), Barberis (2013), Serban (2015).
4 See e.g. Chemero and Silberstein (2008), Stepp et al. (2011), Silberstein and Chemero (2013).
5 Batterman (2010); Rice (2012); Batterman and Rice (2014)
6 An equivalent debate is ongoing over the status of network models in neuroscience. Craver
(2014) argues, contra Huneman (2010), that examples put forward of network models
providing non-causal, topological explanations of neural function are either mechanistic or
non-explanatory. Due to limitations of space I will not discuss network models in this paper.
However, it would be an interesting exercise to apply the framework I develop below to these
kinds of examples.

introduction
instances of distinctively mathematical, non-causal explanation, of the
sort discussed by Lange (2013). Importantly, these explanations meet the
mechanists own criterion for explanatory sufficiency when they are able
to answer w-questions (Kaplan, 2011, 354). In Section 4 I will turn to
one recent example of a dynamical model of a region of the monkey
brain. Here, it is less clear that the model offers a non-causal explanation
and I will argue that the issue turns on whether or not one is willing
to give a realist interpretation of the model components resulting from
dimensionality reduction analysis employed by the model builders. In the
remainder of this section I will say more about the relevant background,
defining efficient coding explanation and presenting my preferred account
of non-causal explanation.
1.1 Efficient Coding Explanation in Computational Neuroscience
In a recent publication I argue that models in computational neuroscience
often yield a distinct, non-mechanistic, pattern of explanation which I call
efficient coding explanation (Chirimuuta, 2014). The term computational
neuroscience labels a broad research area which uses applied mathematics
and computer science to analyze and simulate neural systems. That paper
responds to the work of Kaplan (2011), which attempts to incorporate all
explanatory models of this field within the mechanistic framework. The
case turns on the particular example of the Gabor model of V1 receptive
fields, where a mechanistic criterion for explanatory success, the models
to mechanism mapping constraint (3M) [Kaplan (2011, 347), Kaplan
and Craver (2011, 611)] fails, and yet the model is still able to provide
counterfactual information, thus answering w-questions (Woodward,
2003).
How can this be?7 Well, the models in question ignore biophysical
specifics in order to describe the information processing capacity of a neuron or neuronal population. They figure in computational or informationtheoretic explanations of why the neurons should behave in ways described
by the model. So while, on the one hand, such receptive field models may
simply be thought of as phenomenological descriptions which compactly
summarise observed responses of neurons in primary visual cortex (Kaplan,
2011, 358 ff.), on the other hand, by analysis of the information theoretic
properties of the Gabor function itself, one gains an explanation of why
neurons with the properties captured by the model appear at this particular
stage of visual processing.
In short, such models figure prominently in explanations of why a
particular neural system exhibits a characteristic behaviour. Neuroscientists
formulate hypotheses as to the behaviours role in a specific informationprocessing task, and then show that the observed behaviour conforms to
(or is consistent with) a theoretically derived prediction about how that
information could efficiently be transmitted or encoded in the system,
given limited energy resources. Typically, such explanations appeal to
coding principles like redundancy reduction (Barlow, 1961)the notion that
more information can be transmitted through a cable (e.g. axon) of fixed
bandwidth if some of the correlations between signals are removed. They do
not involve decomposition of biophysical mechanisms thought to underlie
the behaviour in question; rather, they take an observed behaviour and
formulate an explanatory hypothesis about its functional utility.
7 See Chirimuuta (2014, section 5.1) for more detailed discussion of the points covered here.

introduction
It is worth saying a word about the notion of efficiency in play here. A
feature of this research is that neuroscientists draw on knowledge of manmade computational systems and attempt to reverse-engineer the brain,
looking for the principles of neural design (Sterling and Laughlin, 2015).
A basic fact is that information processing makes substantial demands on
resources, both in terms of the material required to build a computer or
nervous system, and the energetic cost of computational processing. It
is assumed, reasonably, that the explanation of many features of neural
systems can be derived from consideration of resource constraintsthe
need to achieve good computational performance in spite of a relatively
small resource budget. As Sarpeshkar (1998, 1602) writes:
The three physical resources that a machine uses to perform its
computation are time, space, and energy. Computer scientists
have traditionally treated energy as a free resource and have
focused mostly on time . . . and space . . . . However, energy
cannot be treated as a free resource when we are interested
in systems of vast complexity, such as the brain. . . . Energy
has clearly been an extremely important resource in natural
evolution. [Cf. Attwell and Laughlin (2001); Sterling and
Laughlin (2015)]
So while, in what follows, it is instructive to compare efficient coding
explanations to optimality explanations in biologybecause both are in the
business of comparing actual biological systems to theoretically optimal
solutionsthe field does not rely on the strong adaptationist assumption
that the brain of humans, or any other animal, is somehow optimal.8 One
must instead make the weaker assumption that there is some process or
mechanismeither evolutionary, developmental or occurring during the life
of the organism as adaptation through neural plasticitywhich causes the
system to tend towards the optimal solution. Efficient coding explanations
typically proceed without specifying what that process is.
1.2 Defining Non-causal Explanation
Efficient coding explanations often make fine-grained predictions about
what would occur in counterfactual scenarios. This is possible because of
the way that the efficiency of a computational procedure is sensitive to the
nature of the particular task at hand. For example, neurons with a certain
kind of receptive field structure might be the most efficient means to encode
sensory information in one kind of environment, but not for another. Thus
one can show how neural properties are counterfactually dependent on the
evolutionary or developmental environment. For this reason I have argued
that this branch of computational neuroscience employs a proprietary kind
of non-mechanistic, causal explanation. Yet if one is willing to extend the
notion of a mechanism to include the whole apparatus of natural selection
and ontogenesis one might propose that computational neuroscience is still
just in the business of discovering mechanisms. However, the assimilatory
impulses of even the most flexible-minded mechanist would have to stop at
8 Nor do optimality explanations in biology always rest on this assumption (Godfrey-Smith,
2001). Instead, the point is to show that an observed feature has similarities with a theoretically
predicted optimum, though there may be substantial departures from optimality due to
structural or other constraints. Also, Wouters (2007) gives an account of non-causal design
explanation in biology which does not depend on any claims about the optimality, or near
optimality, of biological systems.

introduction
the idea of distinctively mathematical non-causaland non-constitutive
explanation. So a central motivation for exploring the question of whether
non-causal explanations occur in neuroscience is as a way to get clearer on
the limits of the mechanist framework.
Woodwards interventionist theory of causal explanation has been extremely influential in the philosophy of neuroscience, and Woodwards
proposal that explanatory sufficiency is tracked by the ability of a theory
or model to address w-questions is accepted by authors such as Kaplan
and Craver. The basic intuition is that, a successful explanation should
identify conditions that are explanatorily or causally relevant to the explanandum: the relevant factors are just those that make a difference to
the explanandum in the sense that changes in these factors lead to changes
in the explanandum Woodward (forthcoming, 5).
Interestingly, Woodward (2003, 221) suggests that the ability to address
w-questions may range beyond causal explanation, writing:
the common element in many forms of explanation, both causal
and noncausal, is that they must answer what-if-things-hadbeen-different questions. When a theory tells us how Y would
change under interventions on X, we have (or have the material
for constructing) a causal explanation. When a theory or
derivation answers a what-if-things-had-been-different question
but we cannot interpret this as an answer to a question about
what would happen under an intervention, we may have a
noncausal explanation of some sort.
Woodward gives the example of the hypothesis that the stability of the
planets is counterfactually dependent on the four dimensional structure
of space-time. What if space-time had been six-dimensional? There is no
intervention associated with this question;9 but the hypothesis is that if
things had been different then planetary orbits would indeed be less stable.
In this paper I follow Woodward in defining an intervention as an
idealized, unconfounded experimental manipulation of one variable which
causally affects a second variable only via the causal path running between
these two variables (Woodward, 2013, 46).10 Like various other authors,11 I
believe it is useful to de-couple the counterfactualist parts of Woodwards
account of explanation from the causal, interventionist ones and thereby
develop an account of non-causal explanation. Moreover, I propose that this
account be integrated with Langes notion of distinctively mathematical
explanation to give a clearer standard for differentiating causal from noncausal explanations than is often employed in the literature.12
9 For one thing, if God in a new act of creation were to change the dimensionality of space-time,
this could not be thought of as a possible intervention to be performed by finite beings. It
would be contentious to extend an interventionist account of causation to acts of creation by
infinite beings. More to the point, the theory of general relativity tells us that the counterfactual
dependence of planetary stability on the geometry of space-time is not a result of any causal
relationship between these two. For this reason, we cannot think of alterations in space-time
which result in changes in planetary stability as interventions in Woodwards sense. See
definition at the start of next paragraph, and see Woodward (2014, 702).
10 I take this to be uncontentious since many mechanist authors have adopted Woodwards
interventionist approach to causation, most notably Craver (2007).
11 See Bokulich (2008), Bokulich (2011), Saatsi and Pexton (2013).
12 A new paper by Baron et al. (forthcoming) independently hits upon this idea of using the
framework of counterfactual explanation to characterise distinctively mathematical explanation.
Their more formal presentation of this synthesis is a useful supplement to the examples I
present below.

introduction
Marc Lange analyses distinctively mathematical explanations of regularities and events in terms of the modal strength of mathematical facts, in
comparison to ordinary causal laws. For example, Lange (2013, 488) writes:
That Mother has three children and twenty-three strawberries,
and that twenty-three cannot be divided evenly by three, explains why Mother failed when she tried a moment ago to
distribute her strawberries evenly among her children without
cutting any.
This is conceptually different from any causal explanation that mentions, for
instance, that the attempt made Mother hungry and frustrated (hangry)
and so she ended up eating two of the strawberries herself, or that one of
the little darlings stole from the other, etc.
Thus I share Langes view that distinctively mathematical explanations
employed in science are non-causal ones (Lange, 2013, 506). Furthermore,
I propose that we bring Langes notion of modal strength in non-causal
explanation to bear on Woodwards idea that non-causal explanations
occur when we show that there are dependencies between the explananda
and explans which cannot be understood in interventionist terms. In
such cases, knowledge of the dependencies does not show us how things
would be in a range of counterfactual scenarios in which we perform
manipulations on the explananda. Instead, we are told how things would
be under certain impossible scenarios in which the laws of mathematics
are altered.13 This of course assumes that counterpossible statements
counterfactuals or subjunctive conditionals with impossible antecedents
can be non-vacuously true. I invite the reader to consider that it is intuitive
that the statement, if thirteen were evenly divisible by three, then I could
share my bakers dozen of doughnuts equally amongst my three best
friends is non-vacuously true, and that the statement if thirteen were
evenly divisible by three, then a bakers dozen of doughnuts would be a
healthy snack is non-trivially false; what is more, there is a recent literature
on the semantics of counterpossibles which underwrites these intuitions.14
In the examples I present in the following two sections, we have a noncausal explanation which is reliant on a trade-off demonstrated in the theory
of information. Such trade-offs are candidates for being brute mathematical
factsnothing could be done to make them not obtain. For this reason, we
can think of the trade-offs as modally strong mathematical facts, in Langes
sense, and as yielding information about counterfactual dependencies
which go beyond interventionist interpretation, in Woodwards sense. Like
Lange (2013), but unlike Saatsi and Pexton (2013), I intend my account
of non-causal explanation in neuroscience to cover explanations both of
particular events and regularities, since the kind of mathematical facts that
my examples depend on do yield explanations of both sorts.
In contrast to the account of non-causal explanation just outlined, Robert
Batterman and Collin Rice have focussed on the way that certain models
in physics, biology and economics idealise away from the causal processes
which lead up to a phenomenon. Such models result in representations of
the causes of phenomena which are, at best caricatures (Batterman and
Rice, 2014). So these authors argue that the relevant notion of explanation
is not a causal one since the explanation of a phenomenon does not arise
13 While it is thought not even God could affect such interventions.
14 See e.g. (Brogaard and Salerno, 2013) and (Bjerring, 2014). I am very grateful to an anonymous
reviewer for pointing me to this literature.

case i: hybrid computation

because we are provided with information about what caused it. This is
how Rice (2015, 600) makes the case:
When considered together, these idealizations entail that optimality models usually provide little, if any, accurate information
about the actual causes, or causal mechanisms, within the
models target system(s). In the end, the highly idealized
optimality model represents mathematical relationships between
constraints, tradeoffs, and the systems equilibrium point that
do not mirror any causal relationships (or processes) in the target
system. Put different, optimality models fail the kind of modelto-mechanism-mapping requirement of causal theories.
This is a less stringent notion of non-causal explanation than the one
that I will employ below. Although all of my examples of distinctively
mathematical explanations do abstract away from causal details, this feature
is not what makes them non-causal (Lange, 2013, 506). While my cases of
efficient coding explanations involving trade-offs do bear comparison with
Rices discussion of optimality models in biology, I would like to emphasise
an important difference, namely that the optimality models discussed by
Rice (e.g. the Fisher model of the 1:1 sex ratio) posit trade-offs which could be
subject to experimental intervention (e.g. through engineering of a species
or environment such that the cost of producing sons is not equal to the cost
of producing daughters), whereas the trade-offs central to my examples are
in a stronger sense fixed. This is because they are mathematical rather than
empirical facts.
Irvine (2014, 12) argues that if the equilibria discovered in optimality
modelling show stability or robustness in the face of experimental interventions, then the model cannot offer causal explanation. Again, I
suggest this is too loose a notion of non-causal explanation because she
intends it too include models whose equilibria could shift if we were to
change something fundamental about the system such as intervening
on inheritance mechanisms or introducing random fluctuations in the
environment. But these are just different kinds of interventionsimpractical
or technically infeasiblebut still within the remit of causal analysis.
Irvines proposal does not give us a clear cut way of deciding whether an
explanation is causal or non-causal. In the next two sections of the paper I
will show that it pays to uncover the use of mathematical facts in efficient
coding analyses because these offer cases of non-causal explanations in
neuroscience which are analogous to Langes examples from biology, and
which can be incorporated into Woodwards counterfactualist framework.

case i: hybrid computation

My first example is an explanation of the efficiency of the brainthought

of as a biological computational systemas being due to the advantages
of hybrid computation.15 The account was presented by Rahul Sarpeshkar
from the department of electrical engineering and computer science at MIT.
Sarpeshkars research aims to use insights gleaned from our understanding
of computation in biological systems in order to build more powerful and
15 Needless to say, my arguments do not turn on whether or not this explanation is disputed or
accepted within the scientific community; rather, it serves to illustrate a pattern of explanatory
reasoning. Recent work on this topic is reviewed by Sterling and Laughlin (2015, chap. 10).

case i: hybrid computation

efficient artificial computers (Sarpeshkar, 2010). His analyses must often
abstract away from the implementational details which obviously differ
tremendously from biological to man-made devices, and they consider the
problem of computation in more purely mathematical terms. At the same
time, this research is in the tradition of other efforts to apply information
theory to neural systems in order to gain understanding of why the nervous
system has the anatomy and physiology that it does.16
Since the birth of modern computing and the concurrent rise of quantitative neurophysiology in the mid twentieth century, many have compared
these two very different kinds of information processing devices and have
pondered the question of whether computation in the brain is analogue or
digital.17 It is known that brains are vastly more energy efficient than any
digital computer. One estimate for the power (energy consumption) of the
human brain is 12W (Sarpeshkar, 1998, 1601), whereas a supercomputer
such as IBMs 2007 Roadrunner, engaged in an equivalent number of
around 1015 processing events per second (i.e. 1 petaflops) runs at 2.4
MW (Komornicki et al., 2009). A question posed and discussed by various
neuroscientists is how to account for the impressive energy efficiency of
neural tissue. Sarpeshkar (1998) considers this question by examining the
relative efficiencies of analogue and digital computational systems.18
The key characteristic of a digital system is that, by definition, one of
its components (e.g. a wire) can only represent 1 bit of information at a
given time. This is because signals are all or nothing events, so the wire has
only two informationally relevant stateson or off, 1 or 0which amounts
to 1 bit of information.19 In contrast, for an analogue system the signal
varies continuously with some physical variable of its components (e.g.
voltage), so any one component has the potential to represent countless
bits of information at a given time (Sarpeshkar, 1998, 1605). For example, to
transmit 4 bits of information, one would need four separate components
(wires) if coding digitally, but the same amount of information could be
transmitted on a single wire in an analogue system, so long as 16 different
physical states of the wire (e.g. voltages) could be unambiguously associated
with 16 different signals.20 Note that these are not empirical claims about
copper wires, axons, or any other bits of hardware; they are, if you like
analytical statements about what we mean by representing or transmitting
information in these different ways.
It follows that analogue systems are far less hungry for resources
raw materials and energy needed to build and maintain the signalling
16 See Cover and Thomas (2006) for more on information theory and Rieke et al. (1999) on the
application to neuroscience.
17 E.g. MacKay (1991, 40) quoted by Husbands and Holland (2008): Later in the 1940s, when I
was doing my Ph.D. work, there was much talk of the brain as a computer and of the early
digital computers that were just making the headlines as electronic brains. As an analogue
computer man I felt strongly convinced that the brain, whatever it was, was not a digital
computer. I didnt think it was an analogue computer either in the conventional sense. See
also von Neumann (2000), McCulloch and Pitts (1943), MacKay and McCulloch (1952).
18 To be more precise, the important contrast is between analogue and asynchronous digital or
pulsatile systems. These are systems in which the signals are discrete, all or nothing events,
but they operate in a continuous time frame. Neuronal spikes are a good example. Computer
microprocessors, in contrast, send their discrete signals on a strictly clocked time schedule, and
this is what is normally meant by digital computation.
19 Where information is defined in the Shannon-Weaver sense as log2 (N), N being the number
of possible and equally probable states of the wire.
20 Of course, in a true analogue system the signal is a continuous function of a physical magnitude
such as voltage so the idea of associating just 16 different signals with 16 discrete voltages in
the wire is artificial. But it serves when making the comparison in resource consumption across
digital and analogue systems.

case i: hybrid computation

ANALOGUE'TYPE'CODE'

DIGITAL'CODE'
not'cold'

hardly'cold' not'cold'
less'cold'
cold'
very'cold'

cold'

coldest'

Figure 1: Illustration of the Trade-off between Resource Consumption and Noise

Susceptibility. Given a fixed material resource which is subject to random
fluctuations in its physical states, e.g. one persons arm, a digital code can
only transmit 1 bit of information, signalling two different possible states,
cold and not cold . In the analogue style code, the fixed material
resource can be used to transmit an indefinite amount of information, since
meaningful signals are defined across a continuous array of physical states.
The angle of an arm from 0 to 90 is a physical continuum, and in this
analogue style code, 6 states of the world are associated with 6 different
ranges of arm angles. However, it is clear that the more states one tries
to encode, the more susceptible the system is to noise. Here, any slight
wobble of the arm could lead to an error in signal transmission.

componentsthan are digital ones. However, the downside of analogue

computation is that such systems are much more susceptible to corruption due to noiserandom fluctuations in the physical states of the
componentsthan are digital ones. To take our example of a single wire
used to transmit 4 bits of information, it is easy to see that if there are
random changes in the voltage due to motion through magnetic fields, for
example, then what started out as a 2V signal could end up being received
as a 2.5V one. And the more information one attempts to transmit through
a single wire, the greater the problem because the difference between
the physical magnitudes encoding the signals must decrease. Again, this
problem is not an empirical observation of the behaviour of metal wires
but occurs for any signalling system whatever its material realisation, as
Figure 1 illustrates.
Thus there is a rigid trade-off between economy of resource consumption
and susceptibility to noise.21 This trade-off is inherent to the definitions
of noise, information and signal within the mathematical theory of information, and could not be altered through any empirical intervention.
Indeed, noise is a result of the physical makeup of the components (e.g.
random electron motion in nano wires, or random ion flux across neuronal
membranes), and one may successfully engineer less noisy components. But
the trade-off still obtains: the more information one tries to send through
21 In my exposition I concentrate on resource consumption in terms of number of material
components required to transmit a given amount of information. The trade-off also occurs with
respect to energy consumed to generate signalling states. One could increase the maximum
voltage of a wire in order to create a greater spacing between the voltages of signalling states,
and thus make the system more resistant to noise corruption, but this would obviously require
a bigger energy investment.

1622

case i: hybrid
computation
Rahul
Sarpeshkar

Figure 2: Hybrid Computation. Each hybrid link combines analogue processing

Figure with
5: Hybrid
computation. In this form(A/D/A)
of computation,
analog
an analogue-digital-analogue
converter
to processing
clean up the
is followed by restoration of the analog signal to a set of M discrete attractor
signal. Sarpeshkar (1998, 1622) writes that hybrid chains allow us to
states. As discussed in section 4.2, hybrid chains allow us to operate with the
operate with the precision and complexity characteristic of digital systems,
precision and complexity characteristic of digital systems, while doing efficient
while doing efficient analog processing. Sarpeshkar (1998, fig. 5),
analog processing.
permission needed.

precision.
Using A/D/As
probably
not abe
good
technique
for maintainones
components,
the moreisthe
signal will
affected
by noise.
Given any
anything more
4 bits
of precisionone
on tries
an analog
we shall
set ing
of components
thethan
more
economically
to useinput.
thoseAs
components,
22
in section
4.3,are
thegoing
mainto
use
A/D/Asby
is noise.
in distributed
analog
thediscuss
more ones
signals
befor
corrupted
computation,
where
it
is
unnecessary
to
maintain
too
much
precision
on
The point of Sarpeshkars analysis is to sketch out an optimally efficient
one wire.
system, given this trade-off. The 1998 paper contains a number of different
To maximize the efficiency of information processing in a hybrid chain,
calculations to show, for example, the point at which a given analogue
there is an optimal amount of analog processing that must occur before
system
useless
due tolink;
noise
signalbecomes
restoration
in a hybrid
thataccumulation.
is, hybrid linksSarpeshkar
should not proposes
be too
hybrid
computation
as
the
best
general
solution
to
the
problem
building
long or too short. If the link is too long, we expend too muchofpower
(or a
system
to noise
a digital
but economical
area, which
or both)isinresilient
each analog
stage(like
to maintain
thecomputer)
requisite precision
at the
with
resources
(like an Ifanalogue
Asexpend
Sarpeshkar
(1998,
1636)
input
of the A/D/A.
the link iscomputer).
too short, we
too much
power
puts
hybrid
computation
combines
the best
the analog
digital
(orit,area
or both)
in frequent signal
restorations.
In of
section
4.2.2, weand
analyze
the optimal
length
of a hybrid
say,
if weidea
are is
worlds
to create
a world
that is link
morequantitatively.
efficient thanNeedless
either. to
The
basic
about efficiency,
then
the link can
be as longsteps
or as using
short asdigitalwe
to unconcerned
alternate between
digital and
analogue
processing
like,
as
long
as
we
meet
the
A/D/A
constraint.
analogue converters. This way, one has the benefit of small chunks of
efficient analogue computation, in which noise accumulates, interspersed
4.2.1 The
A/D/A. To
signal,
we mustThe
have
discrete
attractor in
with digital
processing
to restore
clean aup
the signal.
idea
is illustrated
states.
In
digital
signal
restoration,
the
input
signal
is
compared
with a
Figure 2.
threshold,
and
high-gain
circuits
restore
the
output
to
an
attractor
state
Most of Sarpeshkars analysis is illustrated in terms of metal that
wires
is a function of the input attractor state. The input may deviate by a fairly
and other artificial electrical components; but he is keen to argue that
large amount from its attractor state, and the output will still be very close
his conclusions generalise to biology (Sarpeshkar, 1998, 1634 and 1636).
to its attractor state. The noise immunity of digital circuits arises because the
Furthermore,
it is in
interesting
to consider
neuronal
physiology in
terms
typical distance
voltage space
between an
input attractor-state
level
and of
theahybrid
hypothesis:
threshold
level is many times the variance of the noise or the offset in the

Action potentials are all-or-none discrete events that usually

occur at or near the soma or axon hillock. In contrast, dendritic
processing usually involves graded synaptic computation and
graded nonlinear spatiotemporal processing. The inputs to the
dendrites are caused by discrete events. Thus, in neuronal
information processing, there is a constant alternation between
spiking and nonspiking representations of information. . . . This
alternation is reminiscent of the constant alternation between
22 As a historical aside it is worth noting that back in the 1950s John von Neumann, one
of the founders of modern digital computing, observed that one could offset the effect of
noisy components in the brain by using more of them in a code with built in redundancy,
essentially highlighting this trade-off between the economy and reliability of a coding system
(von Neumann, 2000).

10

case i: hybrid computation

discrete and continuous representations of information in Figure
5 [Figure 2 above]. Thus, it is tempting to view a single neuron
as a D/A/D.23 (Sarpeshkar, 1998, 1630)
It is worth underlining some of the various claims that Sarpheshkar makes
in the article. In the passage just quoted, he offers an interpretation of
an often observed fact of neural physiologythe flow of electrical activity
from dendrites to axons, and then on to dendrites of another neuronas an
implementation of hybrid computation. It could be said that he explains this
fact by appealing to the efficiency of hybrid computation. But more central
to his project is the explanation of the efficiency of the brain in its entirety
the fact that a man made super computer consumes orders of magnitude
more energy than a biological brain. This fact is explained by showing that
hybrid computation is the optimal solution to the problem of maximising
resilience to noise while minimising resource investment, and by showing
that it is plausible that hybrid computation is implemented in biological
brains. As he states in the the abstract of the paper
Our results suggest that it is likely that the brain computes in a
hybrid fashion and that an underappreciated and important reason for the efficiency of the human brain, which consumes only
12 W, is the hybrid and distributed nature of its architecture.
(Sarpeshkar, 1998, 1601)
At the heart of Sarpeshkars account is the demonstration of the costs and
benefits of analogue and digital computation, and the optimality of the
hybrid combination. Thus, I argue, the efficiency of the brain is explained,
non-causally, by its implementation of hybrid computation. In support of
this claim I will now compare Sarpeshkars account with some standard
examples of non-causal explanation.
My first point is that Sarpeshkars proposal answer what-if-things-hadbeen-different questions. According to his discussion, the efficiency and
reliability of the brain are counterfactually dependent on the efficiency and
reliability of hybrid computation, which is in turn dependent on a trade-off
between resource consumption and reliability. If this trade-off did not occur,
then a purely digital or purely analogue system could in principle satisfy
both of the desiderata, and the physiology of brains would have been very
different. At the same time, any such scenario cannot be interpreted as the
result of a possible intervention because the existence of the trade-off is not
an empirical fact about the properties of actual objects but the result of our
information theoretic definitions of analogue and digital coding schemes
(see figure 1). That is to say, the point is not that some materials make
more reliable signalling systems than others, but that given any physical
system, and given a fixed resource investment, the amount of information
transmitted per by the system trades off against the susceptibility of the
signal to be corrupted from noise.
This is like Woodwards example of the non-causal explanation of the
stability of planetary orbits, except that we are clearly talking about what
would obtain if the principles of information theory had been different. In other
words, the information theoretic explanation of the efficiency of the brain
23 Sarpeshkar (1998, 1630) adds these words of caution: However, although the firing of a spike
is a discrete event, it does not imply that it encodes information about a discrete state. The
information encoded by a spike is meaningful only in relation to spikes in different neurons, or
in relation to earlier or later spikes in the same neuron. If these relationships are analog, then
all-or-none events do not imply the encoding of discrete states.

11

case i: hybrid computation

informs us about counterfactual (or counterpossible)24 scenarios in which
the laws of information theory are different. Tinkering with information
theory and working out the implications for coding systems cannot be
thought of as a causal intervention. One may rightly worry that such
information is not as useful as counterfactual information about the effects
of causal interventions.25 Nevertheless, it is worth pointing out that all
these kinds of explanations can be accommodated by Woodwards scheme
for non-causal explanation.
We can also ask if this example falls within the framework of Langes
distinctively mathematical explanation. To this end, it is worth comparing
it with the famous bridges of Knigsberg example. The explanation for
why no-one has ever managed to cross all of the bridges in the city of
Knigsberg (in their 1735 configuration) just once, without ever doubling
back over a bridge, or resorting to flight or swimming, depends on the
characterisation of the set of bridges as a network of nodes and edges
in graph theory. As it happens, each of the four nodes (rather than the
requisite number of two nodes) is touched by an odd number of edges,
and it is a modally strong mathematical fact that such a network cannot
allow for the desired sequential crossings (Lange, 2013, 489). Likewise, one
may interpret Sarpeshkars analysis as showing that it is a modally strong
mathematical fact that hybrid computation is the optimal way to satisfy the
twin constraints of efficiency and robustness in the face of noise. So just
as the graph theoretic properties of the bridge layout explain the observed
limitations on crossing routes, the hybrid computational properties of the
biological brain explain its efficiency.
It is also important to note that non-causal explanations of biological
systems are focussed on the problem of specifying why particular strategies
are highly efficient or theoretically optimal; that is, they have a specific kind
of explanandum. We can, in turn, then explain why natural selection would
have settled upon these strategies. But it is important to keep the causal
explanation of how did this trait evolve? separate from the non-causal
explanation of why a certain trait may be optimal. For example, if one
asks simply, why should the brain operate like a hybrid computer?, the
response could be, because during the evolution of the brain there was
strong selective pressure for the brain to be efficient and reliable, and hybrid
computation is demonstrated theoretically to be the best way to satisfy
the conflicting demands of efficiency and reliability. Here, a causal and
a non-causal explanation rub shoulders; but when arguing that distinctively
mathematical, non-causal explanations take place it is important to be
specific about the explanandum so that one can isolate the non-causal
component of broader explanations.
Lange makes a similar point in his discussion of the explanation of the
hexagonal shape of honeycombs.
The explanation is that it is selectively advantageous for honeybees to minimize the wax they use to build their combs
together with the mathematical fact that a hexagonal grid uses
24 Depending on how one understands the modal strength of information theory. It is conceivable
that key concepts in information theory could have been defined differently and the theory
still be consistent. So the Shannon-Weaver definitions do not have the obvious modal strength
of 13 is a prime number. That said, the important point here is that any counterfactual
dependency between the properties of coding schemes and information theory cannot be
undertsood as a causal one, just as the dependency of planetary stability on the dimensionality
of space-time is not causal.
25 This worry is the topic of another paper (see AUTHOR).

12

case i: hybrid computation

the least total perimeter in dividing a planar region into regions
of equal area .... This explanation works by describing the
relevant features of the selection pressures that have historically
been felt by honeybees, so it is an ordinary, causal explanation,
not distinctively mathematical. But suppose we narrow the
explanandum to the fact that in any scheme to divide their
combs into regions of equal area, honeybees would use at least
the amount of wax they would use in dividing their combs
into hexagons of equal area.... This fact has a distinctively
mathematical explanation. (Lange, 2013, 499-500)
However, in order to support the point that Sarpeshkars explanation is
distinctively mathematical, we would need to say more about the details
of his quantitative analysis. The description above of the trade-off and
the optimality of hybrid computation is based on Sarpeshkars qualitative
(intuitive) account of these coding schemes. He also gives a mathematical
description of the pros and cons of analogue and digital computation,
presented as a comparison between the resource precision curves for
analogue and digital systems (Sarpeshkar, 1998, figure 3). What these show,
on the one hand, is that the relationship between resource consumption
and signal-to-noise ratio for digital systems is compressive, such that digital
systems can increase precision by many orders of magnitude and only incur
minor increases to an already high baseline level of cost. On the other hand,
we see that for analogue computation there is a an expansive relationship
between resource consumption and signal-to-noise ratio, and also that
the maximum precision of the system is strictly bounded by thermal
noise. When precision is low, costs are extremely low, but costs increase
dramatically when the analogue system is operating in a more precise
regime. At a certain signal-to-noise ratio (the crossover point), the costs
of analogue computation begin to outpace those of digital computation.
The overall point is that analogue systems offer very cheap computation,
but only if one is willing to tolerate a poor signal to noise ratio. The
question now is, why should we think of this as a mathematical fact? One
reason to think that the facts summarised in the resource precision curves
are straightforwardly empirical is that the points plotted in such graphs are
the outputs of equations which have empirically measurable parameters,
such as the length and width of transistors. And as Sarpeshkar (1998, 1615)
writes, [t]he exact location of the crossover point will depend on the task,
technology, and ingenuity of the analog and digital designers.
The reason why there is this contrast between cheap analogue computation and costly but precise digital computation boils down to the fact
that the resource precision curves follow a logarithmic relationship for
digital computation, while for analogue computation power consumed is
proportional to signal-to-noise ratio, or to S/N2 , depending on the kinds
of transistors used. These equations are derived from examples which
cite specific parameter values, kinds of transistors, and make concrete
assumptions about the noise distribution. The key question is whether
the resource precision curves would take these forms given any parameter
values and reasonable assumptions for analogue and digital systems, or
if the results could be qualitatively different with different values and
assumptions. I suspect that the qualitative form of the results would be
unchanged but since I am not in a position to offer a proof of this assertion,
I concede that we have an open question about whether Sarpeshkars

13

case ii: the gabor model revisited

explanation is distinctively mathematical in this strict sense. In the next
section I present an example which is more clear cut on this issue.
Another potential worry here is that the existence of the trade-off
illustrated in figure 1, and which I have argued is a result of our information
theoretic definitions of the coding schemes, is due instead to the empirical
fact that the states of all physical systems are liable to random fluctuations
and hence introduce noise. In a noiseless world, we would assume,
analogue computers would outperform hybrid and digital ones. Moreover,
the objection goes, the trade-off would not be there in the noiseless world
and so its existence is not a fact about information transmission that goes
beyond any empirical facts. In response I urge that even if the laws of
nature were altered such that we have perfectly noiseless physical systems
for building brains and computers, we should think of the trade-off itself as
being thereeven though natural selection and human engineers would not
be aware of itbecause of they ways that noise and information are defined
theoretically. Consider an analogous scenario: in contemporary society time
spent at leisure trades off against time spent at work; the optimal solution
to this trade-off is known as the work-life balance. In a future Utopia
where robots do all the work, nobody needs to spend time earning a wage
so everyone spends all their time at leisure. People will not even be aware
of the trade-off. But that does not alter the fact that by our definitions, time
spent at leisure is not time spent earning a wage, and vice versa. In some
sense the trade-off is still there, but empirical conditions do not make it
apparent.
That said, it must be admitted that the empirical fact that actual physical
systems are noisy is bearing substantial explanatory weight herethe tradeoff is only relevant to brain physiology and computer design because of that
fact. More generally, the notion of a trade-off only makes sense if we assume
the presence of certain constraints. So in the final section of this paper I will
consider the question of whether Sarpeshkars explanation is ultimately a
causal one because it must refer to such empirical facts. Before moving on,
it is worth noting that nothing turns on whether one uses Sarpeshkars
analysis to explain a feature of one particular brain or a trait of brains
in general. Thus, I would argue, this kind of non-causal explanation is
applicable to singular states of affairs as well as regularities.

case ii: the gabor model revisited

In the previous case we can observe a close proximity between causal,

evolutionary and non-causal, mathematical explanations in biology and
neuroscience. We see this again when we compare different efficient coding
explanations of the response properties of neurons in primary visual cortex
(V1). As discussed in detail in Chirimuuta (2014, 5.2), the 2D Gabor
function (product of a sinusoid with a Gaussian envelope) is part of the
standard model of V1. It has long been observed that neurons in this area
respond selectively to the visual presentation of short, bar like stimuli of a
specific width and orientation. The parameters of the Gabor function can be
adjusted to capture this selectivity pattern (the receptive field, RF), and this
raises the question of why V1 neurons have receptive fields that can be fit
by this equation.
I previously argued that the Gabor function is central to causal, but nonmechanistic explanations of V1 response properties. Here, I contend that

14

case ii: the gabor model revisited

there is also a non-causal part of the picture. When Gabor (1946) first
introduced the 1D version of the function it was in the context of signal
engineering. Given the problem of analysing information contained in
signal waves, he noted that there is a trade-off between ones ability to
accurately decode temporal and spectral (frequency) information, which is
known as Heisenberg-Weyl uncertainty. The more reliably one decodes
the arrival time of a signal, the less reliably one decodes its frequency, and
vice versa. One cannot, at the same time, be maximally certain about both
of these parameters. For example, Fourier analysis provides a very accurate
analysis of the spectral composition of a wave, but temporal information is
lost. Gabor proved that his function, which in effect performs a temporally
localised Fourier analysis, provides the optimal balance between recovering
temporal and spectral information. This, we can say, is a modally strong
mathematical fact.
In one of the first papers to model V1 responses with the Gabor function,
Daugman (1985, 1160) pointed out that the Gabor-like properties of V1
receptive fields could be the biological solution to the problem of jointly
resolving both spectral and spatial (rather than temporal) information:
the 2D receptive-field profiles of simple cells in mammalian
visual cortex are well described by members of this optimal 2D
[Gabor] filter family, and thus such visual neurons could be said
to optimize the general uncertainty relations for joint 2D-spatial2D-spectral information resolution.
So there is a distinctively mathematical, non-causal explanation of why
the Gabor function is optimal, which can form part of a bigger causal
explanation of why V1 neurons should have evolved (or developed) their
observed properties.
Other neuroscientists have focussed on the causal factors leading to the
development of Gabor-like RFs, and such enquiry does yield experimental
interventions. For example, Hyvrinen and Hoyer (2001, 2413) suggest
that, [t]he reason why the CRFs [classical receptive fields] have Gabor-like
shapes might thus be that these kind of CRFs are optimal for analyzing
the input that the visual system typically receives. In other words, the
hypothesis is that in evolutionary or developmental time, V1 RFs have
been adjusted to be most sensitive to the kind of visual information usually
prevalent in the natural environment. As it happens, short edges and bar
like structures are very common in images of natural scenes, and so it seems
plausible that the Gabor-like receptive fields are a good way to recover
these kinds of stimuli. In order to test the hypothesis that in the course
of development, neurons in V1 adjust their RF shapes to be most tuned
to the common kinds of structures in the environment it is possible to do
causal experiments in which animals are reared in particular environments
(e.g. one which is full of vertical bars) and see if neurons with a preference
for vertical stimuli then dominate V1. Such experiments were quite popular
in the 1960s and 70s, and significant developmental effects could be shown
[e.g. Blakemore and Cooper (1970)].
It should be noted that causal and non-causal explanations can be
mutually supportive. One may demonstrate that a neural system bears
a similarity to a theoretically optimal system but this could just be a
coincidence. If by changing the circumstances in which the system
lives, one changes what counts as the optimum for the system, and the
system diverges from the previously observed behaviour in a predictable

15

case iii: a dynamical model of prefrontal cortex

way, that is good evidence that the previous behaviour was functionally
significant and did not just bear a coincidental resemblance to the theoretical
optimum. So any alternation between causal and non-causal explanation in
neuroscience should not be taken as a troubling sign of theoretical doublethink, but as a scientifically responsible strategy which profits from the
insights of mathematical theorising, while keeping theoretical speculation
grounded in empirical facts.

case iii: a dynamical model of prefrontal

cortex

When addressing the challenge of understanding actual behaviours, rather

than just single neurons, or general computational features of the brain,
neuroscientists must reckon with the daunting complexity of neural systems.
Multi-electrode neurophysiology allows researchers to listen in on the
activity of 100s of neurons at a time, but even such small samples of
activity from one brain area are extremely difficult to interpret functionally.
Techniques of dimensionality reduction and dynamical systems analysis,
imported from other branches of science, have become popular in the quest
to simplify the brain. One recent study from Bill Newsome and Krishna
Shenoys labs at Stanford employs these techniques in order to shed light
on the question of how the brain makes behavioural decisions based on
complex sensory stimuli. It is worth considering whether their explanation
of context-dependent decision making is causal, non-causal, or both. In
order to address this question, it is first necessary to describe the study in a
fair amount of detail.
4.1 A New Explanation of Context-Dependent Computation
The focus of enquiry was the flexibility of humans and other primates
in responding to relevant sensory stimuli, depending on contextual cues.
The behavioural task targeted this ability by presenting monkeys with a
visual display of hundreds of red and green moving dots. The proportion
of different coloured dots would change from trial to trial, as would the
predominant direction of motion. On some trials, a cue indicated to the
monkeys that they should give their response based on the majority colour,
and on the other trials they were cued to respond based on the predominant
direction of motion. Thus the same stimulus required different kinds of
responses depending on a contextual cue.
During this task, recordings were taken from populations of neurons
in prefrontal cortex (PFC), an area believed to be involved in controlling
flexible, context-dependent behaviour. A number of models have been
proposed to account for the neural basis of this behaviour, and one of the
aims of the study was to test those models against the behavioural and
neural data. The crucial step in the assessment of the models was the use of
principle components analysis (PCA) as a means to simplify the neuronal
population dataset so that it could be represented in a three dimensional
space. Mante et al. (2013a, 79) interpret their three axes as corresponding
to choice (which out of two responses the monkey will give), motion (the
representation of information about the predominant direction of movement
of the dots), and colour (the representation of information about the

16

case iii: a dynamical model of prefrontal cortex

predominant colour of the dots). The dynamical response of the population
to the stimulus on any one trial is plotted as a state space trajectory in these
three dimensions. One key finding was that whether or not the contextual
cue is for a colour or motion response, the irrelevant sensory information
is still represented by the neuronal population. This rules out one popular,
early selection model of context-dependent computation, which assumes that
the irrelevant information is filtered out before it reaches PFC. Another key
finding was that the orientation of the axes relative to the others remains
fixed, regardless of which context is used. This rules out two other models,
which assume that the choice axis lines up with the axis of relevant sensory
information, leading to the contextually appropriate response (Mante et al.,
2013a, 81).
Given the failure of these existing models to account for the data,
the authors go on to present their own model, which they describe as,
a previously unknown mechanism for selection and integration of taskrelevant inputs (Mante et al., 2013a, 78). The first step is the training
of an artificial network of recurrently connected, nonlinear neurons
(Mante et al., 2013a, 81) to perform a simulated version of the behavioural
experiments, and see if the artificial data have the same qualitative
features of the real neuronal data. They then analyse the dynamics of the
trained, artificial network to see if they can account for these features.
In the three dimensional state space representation, the model population
did indeed have the same qualitative features as the neuronal population:
irrelevant sensory information was still represented at the population level,
and the position of the choice axis remained fixed relative to the other axes
(Mante et al., 2013a, 82). The use of the artificial network allowed them
to perform a simulated experiment not possible with the real neuronal
population, which is to perturb the network into a particular state, and
see what state the network relaxes into. This reveals the fixed points of the
network dynamicsthe states which the network always relaxes back to.
The analysis showed a series of approximate fixed pointsknown as a line
attractoralong the choice axis. Another feature of this dynamical system,
referred to as the selection vector, was that perturbation of the network in
the direction of the relevant stimulus dimension caused the network to
relax to a point along the line attractor which is closer to a final choice
point, whereas perturbation of the network in the direction of the irrelevant
stimulus dimension resulted in the network just relaxing back to its initial
point on the line attractor. The authors interpret this feature of the system as
a mechanism for the integration of relevant sensory information (Mante
et al., 2013a, 83).
4.2 Causal or Non-causal?
It is interesting that the authors of the study repeatedly refer to these features of the network dynamicsthe line attractor and the selection vector
as indicating a possible mechanism for context dependent computation in
PFC. The important point is, however, that the components of the posited
mechanism are features revealed only by the dynamical analysis.26 This
26 Note that because the fixed points, etc., are only revealed by reverse engineering the model
through perturbation and relaxation, rather than through an equivalent experiment on the
real neural network (which would not be possible with current techniques), the authors are
careful to present it as a possible or plausible explanation of the behaviour of PFC, rather
than a how-actually model of this brain area. However, it can be understood as an actual
explanation of the artificial neural network, so for the purpose of this section I will just focus on

17

case iii: a dynamical model of prefrontal cortex

means that the notion of mechanism in play here is rather different from the
one usually employed by philosophers of neuroscience and it is not at all a
foregone conclusion that the kind of explanation offered by this model is a
causal-mechanistic one.
In their discussion of the use of dynamical models to give causalmechanistic explanations, Kaplan and Bechtel (2011, 439 and 443) claim that
this is possible if the equations of the model can be interpreted as describing
the activities of parts of actual neurons or circuits. This is consistent with
the models to mechanism mapping (3M) criterion presented by Kaplan
(2011, 347) and Kaplan and Craver (2011, 611). However, the fixed points
which Mante et al. (2013a) characterise as explaining the context dependent
computations cannot be thought of as directly mapping onto any actual
parts of a real or artificial neural network; rather, they are attractors in an
abstract, low dimensional state-space.
It is worth comparing the PFC model with a dynamical model more
obviously amenable to the mechanistic account. Bechtel (2011, 553)
presents the example of the dynamical model of circadian rhythms in
Drosophila, writing that, the equations are advanced . . . as descriptions of
the operations of specific parts of a mechanism and that, an important
part of evaluating the adequacy of a computational model is that the
parts and operations it describes are those that can be discovered through
traditional techniques for decomposing mechanisms. In this case it is
straightforward to advance a model-to-mechanism mapping: the equations
in the model represent such things as the rate of change in concentration of
mRNA of a particular gene (per) in terms of a rate of transcription and rate of
decay. A computational simulation then reveals the dynamics of the system
described by the equations, and the fact that both simulation and biological
system manifest the same oscillatory pattern is reason for thinking that the
dynamical model of the components and operations of the cell (the set of
differential equations) explains circadian rhythms. Here, the issue of how to
interpret highly abstract mathematical objects, such as line attractors, does
not arise.
In order to advance a conventional mechanistic interpretation of the
PFC model, one would have to read organizational features of the target
mechanism very liberally, such that the term refers to some non-localisable
and non-decomposable features of the network, ones which can only be
revealed by DST analysis, rather than components and activities which
are observable without such techniques. Only in this way could the 3M
criterion be satisfied. Yet this move invites the objection that this very liberal
application of 3M stretches the notion of mechanistic explanation beyond
recognition, and indeed usefulness.
On the view of Silberstein and Chemero (2013, 967), global organizational principles or features of complex systems which are revealed
by DST analysis, are not explicable in principle via localization and
decomposition, and thus should not be interpreted mechanistically. The
failure of the mechanistic heuristics of localisation and decomposition
occurs, they argue, because many structurally diverse networks which
differ in their implementational details still exhibit the same global features.
Note that the same independence from implementional details is true for
the artificial neural networks described in the PFC studythe modellers
the explanation of the behaviour of the model itself. As Sussillo and Barak (2013, 627) note, the
operation of these recurrent neural networks is often opaque to modellers because the training
of the networks does not specify how the task should be performed. Thus it is necessary to
open the black box using analytical techniques of DST.

18

case iii: a dynamical model of prefrontal cortex

trained 100 different networks, with different initial conditions and therefore
different synaptic connectivity patterns, and yet the global dynamical
features were the same in each case (Mante et al., 2013b, 23).
So is the model explanatory in another, non-mechanistic, or even noncausal sense? One way to address this question is to see how closely
the study resembles the kind of modelling practices used in physics and
engineering (e.g. nonlinear fluid dynamics), which are often taken to be
a exemplary of non-causal explanation. After all, many of the analytical
tools employed in the PFC study were first developed in that context
(Ott, 2002). Batterman (2002, 23) introduced the notion of type-ii why
questions as those which require an explanation as to why a collection
of micro-physically very different substances all exhibit the same macrobehaviour, e.g. during phase transitions. Such explanations are provided
by mathematical abstraction techniques, such as the renormalisation group,
which demonstrate that all of the different substances are members of the
same universality class, despite their low-level differences.
Ross (2015) shows that an analogous pattern of explanation occurs
when dynamical models and abstraction techniques are used in cellular
neuroscience, to explain why anatomically very different neurons all exhibit
the same spiking behaviour, described by a canonical model. Now one
could argue that a similar explanation is provided by the PFC study because
the authors demonstrated that 100 different networks trained to perform
the context-dependent decision, all with different patterns of synaptic
connections, converged on the same high level dynamical properties (line
attractor and selection vector), which were revealed by the abstraction
techniques of PCA and the perturbation-relaxation investigation.
However, one must also consider that there is a functional explanation for
the different networks convergence on the same dynamics. Mante et al.
(2013b, 23) write that,
We trained many networks (around 100) from different initializations of the weights and biases, and each time the network solved
the problem in the same qualitative way. This points to the fact
that the selective integration task . . . placed strong constraints on
the optimization process.
The implication is that the demands of the task causally shaped the
development of the network in such a way, through the optimization
process, that all the different networks inevitably converged on a pattern of
connections which could implement the high-level dynamics and thus solve
the context-dependency problem. It is an interesting question whether the
non-causal explanation of universality, or the causal-functional explanation
of convergence onto the same behaviour is deeper, or more scientifically
significant in this case; or indeed whether the causal and non-causal
explanations complement each other, as I argued they do in the case of
the Gabor model.
Lacking firm grounds to answer this question, or even clear intuitions
either way, I will bracket the issue for now. As I see it, the more important
explanandum which the PFC study seeks to address is how it is that any
one network can perform the context-dependency taskrather than the type-ii
why question over why different networks perform the task in the same
way. Is the explanation here a non-causal one? Given that we still have
open questions about the interpretation of dynamical models, one thing
we can safely say is that the answer to this question turns on how we

19

case iii: a dynamical model of prefrontal cortex

choose to interpret the working parts of the explanation (i.e. the line
attractor and selection vector). If one regards these as somehow capturing
or summarising the actual causally-efficacious parts of the network (the
properties of real and artificial neurons), then one would be inclined to
say that the explanation is a causal one. For instance, one could say that
causal properties of the network which allow it to integrate information are
modelled and represented by the mathematical artifice of the line attractor,
where the notion of mathematical representation without explanation is
invoked (Saatsi, 2011). On the other hand, if one finds such an interpretation
far fetched, then it is natural to say that the explanation is a non-causal one.
This brings us to the point that arguably the PFC model is not explanatory
at all. The neuroscientists account of context dependent computation
is entirely reliant on the three principle axes revealed by PCA, and the
interpretation of these as indicating the choice of the network, alongside
motion and colour information. This raises the thorny question of how
we should interpret the results of factor analysis. Some might accuse
Mante and colleagues of committing the sin of reifying their essentially
meaningless primary factorsthat is, of awarding physical meaning to all
strong principal components (Gould, 1981, 250).
In sum, we have three options on the table: the first is to say that the
model offers a causal-mechanistic explanation; the second that it provides
some sort of non-causal explanation; the third position is that it is not an
explanatory model. What reason is there to favour the second position? We
can begin by highlighting problems with the first and third options.
I am not moved by the view that the PFC model is not at all explanatory;
this attitude strikes me as being tone deaf to the scientifically compelling
features of the analysis. Because of the way that their network moves
through a low dimensional space in response to sensory input, the
researchers are able to show that it is able to integrate relevant information
and ignore irrelevant information, while retaining information about both
kinds of stimuli, as observed in the actual neural data. The explanandum
is, how is this brain area able to integrate relevant sensory information?,
and by being shown specific features of the global dynamics we are indeed
provided with an explanans. If one is going to be skeptical about the
meaningfulness of principal components in this case, one ought to be
unpersuaded by any explanations which call upon statistical constructs,
such as the greater profitability of Sunnylake Farm, compared to the
otherwise identical Sunnybrook Farm, being due to the higher mean
number of eggs laid per day.27
One strike against the mechanistic interpretation is that the explanation
offered by the dynamical model of PFC is not a constitutive one. That is to
say, it is not an explanation of how a global phenomenoncomputation
comes about because of the activities of some network components, whether
spatially localised or not. Instead, the dynamical properties of the network,
which do the job of explaining the computational phenomena, are themselves global or population level properties of the network.28
27 As Gould (1981, 250) notes, the warning against reification of principal components is not
intended as a blanket ban. Sometimes interpretation is justified by our knowledge of how the
system behaves and what we are likely to be measuring.
28 See Mante et al. (2013a, 79): To study how the PFC population as a whole dynamically encodes
the task variables underlying the monkeys behaviour, we represent population responses as
trajectories in neural state space. Each point in state space corresponds to a unique pattern
of neural activations across the population. Also Mante et al. (2013a, 78): The mechanism
reflects just two learned features of a dynamical system: an approximate line attractor and a
selection vector, which are only defined at the level of the population (emphasis added). So my

20

case iii: a dynamical model of prefrontal cortex

Another important point is that the model gives us no information about
how we might make changes to the network in order to affect changes to
its information processing properties. It does not tell us which connections
would have to be rearranged in order to make the computation no longer
context dependent, for example. However, one might ask, doesnt the model
tell us how we could intervene on the global dynamical properties of the
network (rather than local connectivity patterns) in order to bring about
changes in its computational properties? Isnt it therefore a pared down
representation of some causal relationships?29 I believe that we should
not interpret the counterfactual dependency between global dynamical
features and computational properties as a causal relationship, even in a
very minimal sense. This is because the dynamical description, and the
descriptions of the networks computational properties are just two different
ways of describing the same thingthe network. There is no spatiotemporal separation between putative cause and putative effect. To change
the dynamical properties is just to change the information processing
capacities of the network.
This brings me to say more about the second option which I endorse, i.e.
that the model offers some kind of non-causal explanation. We can think of
the dynamical model as offering an illuminating perspective on the network.
It makes transparent certain counterfactual relationships holding between
the networks global dynamical features and the computations which it
performsfor instance, that the ability of the system to integrate sensory
information for the duration of a trial is couterfactually dependent on it
having a line attractor rather than a point attractor. Thus the model answers
w-questions. We can say, for example, that if the selection vector were never
orthogonal to the irrelevant stimulus dimension, then that information
would not be ignored. This is another instance of an explanatory model
providing counterfactual information which should not be interpreted as
describing the outcomes of possible interventions.
So if my interpretation is correct, this example fits into Woodwards
schema for non-causal explanation. Is the explanation also distinctively
mathematical in Langes sense? Well, if it can be shown that there is
a modally strong connection between the having of certain dynamical
properties (one kind of mathematical description) and the ability to perform
certain computations (another kind of mathematical description), then this
example would also satisfy Langes criteria. But I leave this matter for
readers with greater technical expertise than I myself have. One final
point is that thinking of the model as providing explanations by offering
illuminating global perspectives on a system helps shed light on why
Batterman and Rices minimal model explanations are correctly described
as non-causal ones. The point is not just, as they suggest, that minimal
models like the DST one only offer caricatures of the causal dependencies
[Batterman and Rice (2014), cf. Rice (2015)]. More strongly, when the
features captured by the model are truly global, population-level ones, then
no causal dependencies are being represented at alleither in an accurate
or caricatured way. Instead, what the model does is reveal new, global,
features of the system which indicate why the system also has the global,
observable features which have served as our explananda.

point is that what these authors call a mechanism for context dependent computation cannot
be thought of as providing a constitutive explanation.
29 I.e. a very roughly sketched description of difference makers [Weisberg (2007), Strevens (2008)].

21

causal and non-causal: does the difference matter?

causal and non-causal: does the difference matter?

In this paper I have shown that there are instances in computational

neuroscience of explanations which are precisely analogous to familiar
examples of non-causal explanations in physics and biology. So if one
is persuaded by Langes account of non-causal, distinctively mathematical
explanation, or by the non-causal extension of Woodwards counterfactualist
framework, then one should agree that computational neuroscience is just
another domain in which such patterns of explanatory reasoning occur.
What is interesting about this result is that it stands against the dominant
mechanistic current of recent work in the philosophy of neuroscience, which
seeks to interpret explanatory practice as the mapping of mechanisms:
working out what cause produces which effect, in order to show how
neuronal components are orchestrated to govern complex behaviours,
such as navigation. Moreover, by extending Woodwards framework to
incorporate distinctively mathematical explanation I am adapting materials
frequently used by mechanists to the rather different task of analysing noncausal explanation.
The clearest cases of non-causal explanation in neuroscience are efficient
coding explanations which refer to information theoretic trade-offs in order to
show why it is that neural systems should employ particular computational
solutions, such as hybrid computation (Section 2), or Gabor filtering
(Section 3). So far most of the discussion of non-mechanistic explanation
in neuroscience has focussed on dynamical systems theory. Yet as I show
in Section 4.2, the argument that explanatory practice goes beyond the
detailing of mechanisms is less clear cut in this area. The controversy turns
on the question of whether or not we should interpret the mathematical
structures revealed by dynamical modelling (e.g. fixed points and line
attractors) as representing anything in the brain which has causal powers
to affect computations performed.
So does the difference between causal and non-causal explanation matter
to neuroscientists and neuroengineers themselves? Perhaps not. As noted
at the end of the discussion of the Gabor model, these different kinds of
explanation should be seen as complementary rather than in competition
with one another [Cf. Andersen (forthcoming)].
Still, one might point out that if what we are interested in explaining
is why some empirically realized system (like the brain) exhibits some
feature, then this explanation must always have an empirical component
(e.g. that some optimizing selective process is operative). So one objection
to my account is that the purely mathematical part of the explanation is not
really by itself an explanation of anything empirical and that one needs the
component about natural selection (or something similar) in order to arrive
at an explanation of the empirically observed facts.
In response I would agree that the non-causal component is arguably not
a stand-alone explanation of the efficiency of neural computation or the
Gabor shaped receptive fields of V1, even though researchers in theoretical
neuroscience do talk of these mathematical results as explaining brain
features. Such talk assumes that certain empirical background conditions
obtain, such that the trade-offs are relevant to the neural systems, and
that there is some developmental, evolutionary or realtime adaptational
process by which the near-optimal solution is arrived at; but such a story
is bracketed in order to focus on the relevant topic of enquiry, which is the

22

causal and non-causal: does the difference matter?

interpretation of a particular feature in terms of its utility and efficiency.
So what is interesting here is that there is a division of explanatory labour:
some neuroscientists will focus on the non-causal, mathematical explanation
of the efficiency of a feature while it is the job of others to find out about
the aetiology of that feature.
For the purposes of this paper I need not take a stand in the debate
about the possibility of stand-alone non-causal explanation of empirical
facts, or answer the question of whether we can call an explanation noncausal if there is some non-causal part of the story (or causal if there is
some empirical part of the explanation). But it is worth pointing out that
Langes examples, such as the honeycomb one, invite the same objection.
The task of this paper is to show that there are examples of explanation in
neuroscience that are of the same type as the central examples of non-causal
explanation as presented in the work of Lange, Batterman and Woodward.
I hope by now to have made that case.
Even if I concede that my three cases do not give us stand-alone
non-causal explanations, my examples are still relevant to the ongoing
debate over whether there are distinctly computational and non-mechanistic
patterns of explanation in neuroscience. While authors such as (Kaplan,
2011) echo Saatsi (2011) and argue that mathematics only has the role in neuroscience of representing physical systems via a mapping from equations to
parts and processes of a mechanism, I have argued that mathematics has a
stronger explanatory role which is independent of mechanism description.
We arrive at a picture in which theoretical neuroscience is often focussed on
the non-causal parts of explanations and therefore has explanatory norms
appealing to efficient coding and mathematical trade-offsthat are distinct
from the evolutionary and mechanistic branches of neuroscience. This,
I believe, is sufficient for me to deflect the worry that efficient coding
explanations are just another kind of evolutionary explanation, which are
themselves elliptical mechanistic explanations.
One issue which I have not addressed is the question of how we evaluate
the quality of non-causal explanations. One of the virtues of the mechanistic
account is that it lays down explicit normative criteria (Craver, 2007). There
is not space here to venture into this topic, but one obvious possibility is
to evaluate non-causal explanations according to the precision and range of
counterfactual or counterpossible information provided (i.e. how precisely
and extensively they are able to answer w-questions), in the same way that
one can evaluate causal explanations.
A novel possibility is to consider that non-causal explanation tends
to exemplify a different explanatory virtue from the causal sort, that of
unifying diverse phenomena under one general principle (Kitcher, 1989).
Though the unificatory account of explanation is less popular than it once
was, it should not be ignored that all the cases I have presented do a
good job of explaining why mechanistically quite diverse systems will
converge on similar computational solutionsin other words, answering
Battermans type-ii why questions. In this era of big data neuroscience and
relative scarcity of theoretical insight, unificatory explanatory knowledge is
a precious commodity. It will be instructive to see whether future developments in neuroscience will place non-causal explanations in increasingly
prominent roles.

23

References

references
Andersen, H. (forthcoming). Complements, not competitors: causal and
mathematical explanations. British Journal for the Philosophy of Science.
Attwell, D. and S. B. Laughlin (2001). An energy budget for signalling in the
grey matter of the brain. Journal of Cerebral Blood Flow and Metabolism 21,
11331145.
Barberis, S. D. (2013). Functional analyses, mechanistic explanations, and
explanatory tradeoffs. Journal of Cognitive Science 14, 229251.
Barlow, H. B. (1961). Possible principles underlying the transformation
of sensory messages. In W. A. Rosenblith (Ed.), Sensory Communication.
Cambridge, MA: MIT Press.
Baron, S., M. Colyvan, and D. Ripley (forthcoming). How mathematics can
make a difference. Philosophers Imprint.
Batterman, R. (2002). The Devil in the Details. Oxford: Oxford University
Press.
Batterman, R. (2010). On the explanatory role of mathematics in empirical
science. British Journal for the Philosophy of Science 61, 125.
Batterman, R. and C. Rice (2014). Minimal model explanations. Philosophy
of Science 81(3), 349376.
Bechtel, W. (2008). Mental Mechanisms: Philosophical Perspectives on Cognitive
Neuroscience. London: Routledge.
Bechtel, W. (2011). Mechanism and biological explanation. Philosophy of
Science 78(4), 533557.
Bjerring, J. C. (2014). On counterpossibles. Philos Stud 168, 327353.
Blakemore, C. and G. F. Cooper (1970). Development of the brain depends
on the visual environment. Nature 228, 477478.
Bokulich, A. (2008). Can classical structures explain quantum phenomena?
British Journal for the Philosophy of Science 59, 21735.
Bokulich, A. (2011). How scientific models can explain. Synthese 180, 3345.
Brogaard, B. and J. Salerno (2013). Remarks on counterpossibles. Synthese 190, 639660.
Chemero, A. and M. Silberstein (2008). After the philosophy of mind:
Replacing scholasticism with science. Philosophy of Science 75, 127.
Chirimuuta, M. (2014). Minimal models and canonical neural computations: the distinctness of computational explanation in neuroscience.
Synthese 191, 127153.
Cover, T. M. and J. A. Thomas (2006). Elements of Information Theory (2nd
ed.). Hoboken, NJ: Wiley Interscience.
Craver, C. F. (2007). Explaining the Brain. Oxford: Oxford University Press.
Craver, C. F. (2014).
presentation.

The explanatory power of network models.

PSA

24

References
Daugman, J. G. (1985). Uncertainty relation for resolution in space, spatial
frequency, and orientation optimized by two-dimensional visual cortical
filters. J. Opt. Soc. Am. A 2(7), 11601169.
Gabor, D. (1946). Theory of communication. Journal of the Institution of
Electrical Engineers 93, 429459.
Godfrey-Smith, P. (2001).
Three Kinds of Adaptationism, pp. 335357.
Cambridge: Cambridge University Press.
Gould, S. J. (1981). The Mismeasure of Man. London: Penguin.
Huneman, P. (2010). Topological explanations and robustness in biological
sciences. Synthese 177, 213245.
Husbands, P. and O. Holland (2008). The ratio club: A hub of british
cybernetics. In P. Husbands, O. Holland, and M. Wheeler (Eds.), The
Mechanical Mind in History, pp. 91148. MIT Press.
Hyvrinen, A. and P. O. Hoyer (2001). A two-layer sparse coding model
learns simple and complex cell receptive fields and topography from
natural images. Vision Research 41, 24132423.
Irvine, E. (2014). Models, robustness, and non-causal explanation: a foray
into cognitive science and biology. Synthese DOI 10.1007/s11229-014-05240.
Izhikevich, E. M. (2010). Dynamical Systems in Neuroscience: The Geometry of
Excitability and Bursting. Cambridge, MA: MIT Press.
Kaplan, D. M. (2011). Explanation and description in computational
neuroscience. Synthese 183, 339373.
Kaplan, D. M. and W. Bechtel (2011). Dynamical models: An alternative
or complement to mechanistic explanations? Topics in Cognitive Science 3,
438444.
Kaplan, D. M. and C. F. Craver (2011). The explanatory force of dynamical
and mathematical models in neuroscience: A mechanistic perspective.
Philosophy of Science 78, 601627.
Kitcher, P. (1989). Explanatory unification and the causal structure of the
world. In P. Kitcher and W. Salmon (Eds.), Scientific Explanation, pp. 410
505. Minneapolis: University of Minnesota Press.
Komornicki,
A.,
G.
Mullen-Schulz,
and
D.
Landon
(2009).
Roadrunner:
Hardware
and
Software
Overview.
http://www.redbooks.ibm.com/redpapers/pdfs/redp4477.pdf.
Lange, M. (2013). What makes a scientific explanation distinctively
mathematical? British Journal for the Philosophy of Science 64, 485511.
Levy, A. (2014). What was hodgkin and huxleys achievement?
Journal for Philosophy of Science 65, 469492.
Machamer, P., L. Darden, and C. F. Craver (2000).
mechanisms. Philosophy of Science 67, 125.
MacKay, D. (1991). Behind the Eye. Oxford: Blackwells.

British

Thinking about

25

References
MacKay, D. and W. McCulloch (1952). The limiting information capacity of
a neuronal link. Bulletin of Mathematical Biophysics 14, 127135.
Mante, V., D. Sussillo, K. V. Shenoy, and W. T. Newsome (2013a). Contextdependent computation by recurrent dynamics in prefrontal cortex.
Nature 503, 7884.
Mante, V., D. Sussillo, K. V. Shenoy, and W. T. Newsome (2013b).
Supplementary information.
McCulloch, W. and W. Pitts (1943). A logical calculus of the ideas immanent
in nervous activity. Bulletin of Mathematical Biophysics 7, 115 133.
Ott, E. (2002). Chaos in dynamical systems. Cambridge: Cambridge University
Press.
Piccinini, G. and S. Bahar (2013). Neural computation and the computational
theory of cognition. Cognitive Science 34, 453488.
Piccinini, G. and C. Craver (2011). Integrating psychology and neuroscience:
Functional analyses as mechanism sketches. Synthese 183(3), 283311.
Rice, C. (2012). Optimality explanations: a plea for an alternative approach.
Biology and Philosophy 27, 685703.
Rice, C. (2015). Moving beyond causes: Optimality models and scientific
explanation. Nos 49(3), 589615.
Rieke, F., D. Warland, R. d. R. V. Steveninck, and W. Bialek (1999). Spikes:
Exploring the Neural Code. Cambridge, MA: MIT Press.
Ross, L. N. (2015). Dynamical models and explanation in neuroscience.
Philosophy of Science 82(1), 3254.
Saatsi, J. (2011). The enhanced indispensability argument: Representational
versus explanatory role of mathematics in science. British Journal for
Philosophy of Science 62, 143154.
Saatsi, J. and M. Pexton (2013). Reassessing woodwards account of
explanation: Regularities, counterfactuals, and noncausal explanations.
Philosophy of Science 80(5), 613624.
Sarpeshkar, R. (1998). Analog versus digital: Extrapolating from electronics
to neurobiology. Neural Computation 10, 16011638.
Sarpeshkar, R. (2010). Ultra Low Power Bioelectronics. Cambridge: Cambridge
University Press.
Serban, M. (2015). The scope and limits of a mechanistic view of
computational explanation. Synthese 192(10), 33713396.
Silberstein, M. and A. Chemero (2013).
Constraints on localization
and decomposition as explanatory strategies in the biological sciences.
Philosophy of Science 80(5), 958970.
Stepp, N. A., Chemero, and M. T. Turvey (2011). Philosophy for the rest of
cognitive science. Topics in Cognitive Science 3(2), 425437.
Sterling, P. and S. B. Laughlin (2015). Principles of Neural Design. Cambridge,
MA: MIT Press.

26

References
Strevens, M. (2008). Depth: An account of scientific explanation. Cambridge,
MA: Harvard University Press.
Sussillo, D. and O. Barak (2013).
Opening the black box: Lowdimensional dynamics in high-dimenional recurrent neural networks.
Neural Computation 25, 626649.
von Neumann, J. (2000). The Computer and the Brain. New Haven: Yale
University Press.
Weisberg, M. (2007). Three kinds of idealization. Journal of Philosophy 104(12),
639659.
Weiskopf, D. A. (2011).
Models and mechanisms in psychological
explanation. Synthese 183, 313338.
Woodward, J. F. (2003). Making Things Happen. New York: Oxford University
Press.
Woodward, J. F. (2013). Mechanistic eplanation: its scope and limits.
Proceedings of the Aristotelian Society Supplementary Volume 87, 3965.
Woodward, J. F. (2014). A functional account of causation. Philosophy of
Science 81(5), 691713.
Woodward, J. F. (forthcoming).
Explanation in neurobiology: An
interventionist perspective. In D. M. Kaplan (Ed.), Integrating Psychology
and Neuroscience: Prospects and Problems. Oxford: Oxford University Press.
Wouters, A. G. (2007). Design explanation: determining the constraints on
what can be alive. Erkenntnis 67, 6580.

27