1. Introduction

There is a burgeoning cottage industry in philosophy dedicated to criticizing the modularity of mind hypotheses in evolutionary psychology (see e.g. Buller Reference Buller and Hardcastle1999; Davies Reference Davies1996, Reference Davies and Hardcastle1999; Davies et al. Reference Davies, Fetzer and Foster1995; Grantham and Nicols Reference Grantham, Nichols and Hardcastle1999; Griffiths Reference Griffiths1997; Richardson Reference Richardson1996; Shapiro Reference Shapiro, Allen and Cummins1998, Reference Shapiro and Hardcastle1999; Shapiro and Epstein Reference Shapiro and Epstein1998; Sober Reference Sober1997; Sterelny Reference Sterelny1995). We know because we are part of it (cf., Buller and Hardcastle, Reference Buller and Hardcastle2001). But the criticisms mustered against evolutionary psychologists’ wanton and unwarranted methodological assumptions of mind modules also find a home in cognitive neuroscience. Cognitive neuroscientists assume that they can localize brain function; they seek discrete, physically constant brain “modules,” a material analogue for the psychologists’ set of distinct mental software packages. Indeed, the criticisms should be all the more pointed because, according to some, the brain is where we should be finding these alleged cognitive units. (Mother Nature does build psychologies out of brains, after all.) If anybody has localization data that would support modularity, it should be the neuroscientists.

But they don't have the data. Actually, it is worse than that, because—appearances to the contrary—they don't even have a good way of accessing the appropriate evidence. It is a bias in neuroscience to localize and modularize brain functions. Single cell recordings, the study of neurological deficits, and imaging studies all feed into the Gallian view that different brain areas do different things and the things being done are confined to particular processing streams. But these are just prejudices, nothing more. And the underlying assumptions could very well be wrong.

2. Localization and Single Cell Recordings

When scientists do single unit recordings from a set of neurons they assume that they are busy examining a discrete system. Why else record from this particular set, isolated in this particular fashion, in the first place? Many neuroscientists spend their entire careers delineating small functional areas in the brain, and they are wildly successful in their endeavors. They have identified at least 36 different topographical visual processing areas in cortex (DeYoe and van Essen Reference DeYoe and van Essen1988); they have differentiated the “what” from the “where” object processing streams (Mishkin et al. Reference Mishkin, Ungerleider and Macko1983); and they have distinguished motion detection from contour calculations (Livingstone and Hubel Reference Livingstone and Hubel1988). Our maps of brain function are getting more and more complicated as we learn more and more about the processing capacities of individual cells. And all these projects are founded on the belief that our brain operates in this way, that we have discrete processing streams that feed into one another.

Yet the most neurons we have ever been able to record from simultaneously are around 150; the most cells we can ever see summed local field potential activity over are a few thousand. But brain areas have hundreds of thousands of neurons, several orders of magnitude more than we can access at any given time. And these neurons are of different types, with different response properties and different interconnections with other cells, including other similar neurons, neurons with significantly different response properties, and cells of other types completely (such as glia). Any conclusions we draw about the behavior of whatever cells we are recording from are going to be limited to very basic stimulus-response and correlation analyses of whatever neuronal subtype we are currently examining. Hence, the functionality we ascribe based on these relatively meager sorts of experiments might be much more restricted than what the cells are actually doing.

We insert an electrode in or near a cell and then record what it does as we stimulate the animal in some fashion. We record from a cell in a vestibular nucleus and then move the animal's head about to see if that changes the activity of the neuron (cf., Newlands and Perachio Reference Newlands and Perachio1990a, Reference Newlands and Perachio1990b). If it does, then we move it some more or we move it differently and see how that changes the neuronal output. If it doesn't, then we either try another nearby cell or we try some other stimulus. But what we can't do is record from all the neurons in some isolated area, even if the area is very small. And what we can't do is test any given cell for all the known functional contributions of brain cells in general. So, what we conclude about any cell will only reflect the cells we or others have actually recorded from using stimuli we or others have actually used. This research strategy systematically underestimates when neurons actually respond and under what conditions.

This sort of unit study attempts to combine scores, hundreds, or even thousands of single-unit recordings together to try to analyze the population. Theoretically, we could, perhaps, in principle, delineate a nervous system region stereotaxically if it had reproducible correlations between afferent and efferent connections such that we could ultimately articulate the neurophysiological function of the defined region. However, the likelihood of success for this type of study decreases as the complexity of the organism increases. We can draw functional conclusions regarding the activities of neurons in the abdominal ganglia of Aplysia, or the segmental ganglia of the leech. But the architecture of these organisms’ central nervous system is so different from mammals’ that the probability of successfully using similar techniques is very low to zero.

In addition, the actual processing of information that goes on in those cells involves lots of different kinds of excitatory and inhibitory inputs from other areas in the brainstem, cerebellum, and cerebral cortex. Our dorsal horn is supposed to integrate afferent nociceptive information from the periphery and pass it onto the motor system (among other things), but it doesn't do that segregated from the rest of the brain and what the brain is trying to do. It is integrating and passing as we are trying to pursue prey or flee from an enemy. Moreover, the brain regions that perform these tasks are often connected to the very area we are recording from. The motor system feeds back down into the dorsal horn, as does the thalamus and significant parts of cortex (cf. Hardcastle Reference Hardcastle1999).

We sometimes wonder whether we can't play a variant of the parlor game Six Degrees of Kevin Bacon in the brain. (This is the game in which you can connect any actor or actress in a movie to a movie starring Kevin Bacon via common actors in no more than six moves. So, for example, Elizabeth Taylor has a Bacon number of 2; she was in Rhapsody (1954) with Vittorio Gassman, who was in Sleepers (1996) with Kevin Bacon.) We can connect any neuron in the brain to any other neuron with just a few steps, probably fewer than six.

The impact on cognitive processing of such rampant feedback connections in the brain is only just now starting to be explored in neuroscientific research, though exactly how to do this is a difficult question to answer. Of course, neurophysiologists aren't fools; they design their experiments keeping in mind the known anatomic connections between and among the relevant structures. At the same time, any actual experimental observations of all the remote influences on the dorsal horn, for example, are impossible, despite however many individual neurons we record from. We simply don't have any way of conducting such extensive, invasive tests on live animals. At best, the particular influences assumed in any particular recording series are a matter of previously accepted gospel, dogma, and faith.

Let me give you an analogy for what we are talking about. We have all heard the old joke about the four men who examine an elephant blindfolded. The one who feels the legs believes he is before a tree; the one who gets the truck thinks he has a snake on his hands, and so forth. Experiments in neurophysiology using single-cell recordings are much like that, only each investigator examines a different creature in a different pen on different days, and they only get to look at one square inch per examination period. They simply can't holler out to their companions, “See what happens at your end when I kick the tree trunk.” Moreover, they have to read about each other's conjectures and descriptions in the local newspaper, which may or may not accurately report what the others think and do, if they report such activities at all. And not only do they have to identify what it is they are examining, but they also have to figure out what the function is for each part they are allowed to touch.

Discerning brain function when limited to touching single cells is no easy feat. No wonder neuroscientists assume functional localization. It makes their task at least conceivable. If they didn't assume some sort of functional specificity in discrete brain regions, then what would the purpose of single-cell recordings be at all? In order to get anything theoretically useful out of single-cell recordings, we have to have already accepted that the data are going to tell us about how a specific brain area functions.

3. Lesion Studies and the Assumption of Brain Constancy

Ideally, neuroscientists try to conjoin their single-cell studies with some sort of lesion experiment. Once scientists construct a general flowchart of the relevant structures based on anatomy experiments, and they have estimated normal unit behavior from a series of single cell studies, they then try to knockout the hypothesized functions by placing lesions in otherwise normal animals. They run their experiments based on the assumption that these lesions, placed in regions known to be important, will change the unit behavior of the cells they are studying in a consistent fashion. If they witness such a change, they use that information to explain the relative functional contributions of the lesioned region to the cells under scrutiny. In other words, they are using lesion studies to try to derive a functional boxology for the brain, just as cognitive psychologists use reaction time distributions and error measurements to find one for the mind.

Right off the bat, though, we find technical difficulties. There are many different types of lesions, including electrolytic, incisional, excisional, and chemical. The optimal lesion spares the nerve tracts themselves while eliminating populations of nerve cells. (Chemical lesions are generally the first choice for this type of lesion.) But lesion studies are notorious for highly variable functional damage. Where the lesion actually occurs and how widespread it actually is—facts over which scientists have much less control than one might assume—leads to all sorts of differently ordered remaining populations. As a result, each lesion is unique, and the actual functional deficit after lesioning an animal is highly variable. A genuine replication of a lesion study in neuroscience is a practical impossibility.

But there is a larger theoretical concern. What neuroscientists know, but generally ignore, is that any functional change in the central nervous system will lead to compensatory changes elsewhere (see discussion in Hardcastle and Stewart, Reference Hardcastle, Stewart, Machamer, Grush and McLaughlin2001). If we ablate the semi-circular canals in a rat's ear, then within hours the rat recovers static functioning in its vestibular system (Baarsma and Collewjin Reference Baarsma and Collewjin1975; Schaefer and Meyer Reference Shaefer, Meyer and Kornhuber1974; Sirkin et al. Reference Sirkin, Precht and Courjon1984). The vestibular nuclei of the brainstem recover a level of baseline activity, even though it is receiving no information from the periphery anymore. How does it do this? Frankly, we have no idea.

Here is another example closer to home for many of us. Temporomandibular joint syndrome (TMJ), a progressive degeneration of the joint that connects the lower jar to the skull, is often accompanied by tinnitus in the ipsilateral ear. Indeed, ringing in the ear is a diagnostic criterion for TMJ. But we don't know why having one's mandibular joint fall apart would provoke some change in the auditory system. In fact, we didn't even know that there was an auditory nerve that ran alongside the lower jaw until recently. It is hard to derive the functions of various brain areas if we don't even have a sketch of the correct neuroanatomy in the first place, and this task becomes even more difficult if the brain won't hold still for us.

Because it is highly plastic, poking holes in the brain in one place will provoke to it react in some fashion in some other place. Usually these other places are not components in the system or region being studied. But even if they are, neuroscientists ignore plasticity of the brain in favor of assuming a consistent functional alteration as caused by the lesion and nothing more. How are investigators supposed to evaluate some observed functional change when the difference they see might have been evoked by the brain's attempt to compensate for its loss and not by any specific deficit induced by the lesion?

The short answer is that they can't if they are restricted to single-cell recordings and lesion studies. To answer this question we need to be able to see the activity of the entire brain at once and over time. But we can't do that.

4. Functional Imaging to the Rescue?

We want to claim that neuroscientists’ assignments of function to brain regions or areas are not warranted by the data. They are not warranted because their simplifying assumptions of localization of function and functional constancy are radically false. We admit these seem ludicrous conclusions, for neuroscientists give us functional hypotheses all the time with a perfectly straight face, about brain areas both large and small. The frontal lobes allow for planning complex behaviors (cf., Goldman-Rakic Reference Goldman-Rakic1996; though see Carpenter et al. Reference Carpenter, Just and Reichle2000); the lingual gyrus is sensitive to peripheral visual information (Emiliano et al. Reference Emiliano, Frith and Driver2000); certain cells in superior colliculus are multi-modal and can process both visual and auditory inputs (cf., de Gelder Reference De Gelder2000).

Moreover, and more importantly, the excited hoopla over fMRI and other imagining techniques concerns exactly this point: we do have a way of looking at the activity of the whole brain at one time as tied to some cognitive activity or other. At least, that is what the PR claims. But magnetic resonance imagining, the best non-invasive recording device we currently have, only has a spatial resolution of about 0.1 millimeter and each scan samples about five seconds of activity (cf., Churchland and Sejnowski Reference Churchland and Sejnowski1988). This imprecision forecloses the possibility of directly connecting single cell activity—which operates three to four orders of magnitude smaller and faster—with larger brain activation patterns. What are we to do? The answer given all too often by neuroscientists is to fudge.

Methodological difficulties with current imagining techniques are now becoming well known (cf., Bechtel Reference Bechtel, Creath and Maienschein2000; Cabeza and Nyberg Reference Cabeza and Nyberg1997; Stufflebeam and Bechtel Reference Stufflebeam and Bechtel1997). Let us just add briefly to this discussion. Here is how most functional imagining studies work. The experimenter picks two experimental conditions that she believes differ along only one dimension; they differ only with respect to the cognitive or perceptual process she wants to investigate. She then compares brain activity recorded under one condition with what happens in the second condition, looking for regions whose activity levels differ significantly across the two. These areas, she believes, comprise the neural substrates of the task under scrutiny.

Let us set aside the fact that this so-called subtraction method has no way of determining whether the differences found are actually tied to the cognitive process and not to something else occurring concurrently but coincidentally (cf., Shulman Reference Shulman1996). Notice that how well the subtraction method will work depends upon the sensitivity of the measuring devices and that the worse the instrument is the better the method seems to be for localization studies. Low signal-to-noise ratios (SNR) means that we will find only a few statistically significant differences across conditions. And these are the sort of results neuroscientists need in order to bolster any claims identifying particular cognitive processes with discrete brain regions (Wandell Reference Wandell1999).

But as the imagining technology improves and the SNR increases, we see more and more sites that differ across trials. The more sites we get, the more it looks as though essentially the entire brain is involved in each cognitive computation. And the more it looks as though the entire brain is involved in each thought, the less it is we can justify any assumption of functional specificity in the brain.

We are but victims of imprecise instrumentation. If we extrapolate from what we might learn with more sensitive measures, we can easily see that there will come a time when this whole approach just won't work anymore. It simply ceases to be interestingly informative to learn that, as Brian Wandell notes, “the activity of one pixel is significant to one part in 10¹⁰ and the adjacent pixel is significant to one part in 10¹¹” (Reference Wandell1999, 170). Put in the harshest terms, brain imaging seems to support localist assumptions because we aren't very good at it yet.

It also seems to support functional specificity because there have been few meta-analyses done that look at activity in single brain regions across a wide variety of experimental conditions (though see Cabeza and Nyberg Reference Cabeza and Nyberg1997, Reference Cabeza and Nyberg2000; Lloyd Reference Lloyd2000). It doesn't take much work to see that there is something seriously wrong with trying to assign a particular function to each statistically significant area of subtracted activity. (For simplicity's sake, in the following discussion, we are ignoring the difficulties associated with summarizing the results from multiple studies, including the fact that different researchers use different standards of significance, different baseline conditions, and different statistical methods for culling their data. But even if we were to take these factors into consideration, we believe the conclusions we draw below would still stand.)

Brainmap is a wonderful database housed at the University of Texas Health Science Center in San Antonio (cf. http://ric.uthscsa.edu/projects/brainmap.html). It currently has archived 225 imaging articles, for a total of 771 experiments with 7,683 significant areas of difference being discussed. A quick search for Brodmann area 6 returns 36 articles that found area 6 significantly active in at least one experimental trial. This in and of itself might not be so interesting, but what is interesting is the wide variety of cognitive tasks the area apparently underlies. Area 6 appears significantly active after subtraction in studies of phonetic speech processing, voluntary hand and arm movements, sight-reading music, spatial working memory, recognizing facial emotions, binocular disparity, sequence learning, idiopathic dystonia, pain, itch, delayed response alternation, and category-specific knowledge, to name but a few (Ceballos-Baumann et al. Reference Ceballos-Baumann, Passingham, Warner, Playford, Marsden and Brooks1995; Colebatch et al. Reference Colebatch, Deiber, Passingham, Friston and Frackowiak1991; Derbyshire et al. Reference Derbyshire, Jones, Devani, Friston, Feinmann, Harris, Pearce, Watson and Frackowiak1994; George et al. Reference George, Ketter, Gill, Haxby, Ungerleider, Herscovitch and Post1993; Gold Reference Gold1995; Gulyas and Roland Reference Gulyas and Roland1994; Hsieh et al. Reference Hsieh, Hagermark, Stahle-Backdahl, Ericson, Eriksson, Stone-Elander and Ingvar1994; Jonides et al. Reference Jonides, Smith, Koeppe, Awh, Minoshima and Mintun1993; Martin et al. Reference Martin, Wiggs, Ungerleider and Haxby1996; Rauch Reference Rauch2001; Sergent et al. Reference Sergent, Ohta and MacDonald1992a; Sergent et al. Reference Sergent, Zuck, Terriah and MacDonald1992b; Zatorre et al. Reference Zatorre, Evans, Meyer and Gjedde1992).

Keep in mind that area 6's actual activity is systematically underreported in published articles because it would only be mentioned in studies in which it was not differentially active during baseline scanning. That is, any study in which both the control condition and the test condition prodded area 6 to light up would subtract out this activity as uninteresting for the task at hand, even though that area might be absolutely crucial for carrying out the cognitive process. (Our quick scan of the database would also miss studies that discussed area 6 under a different rubric, such as premotor cortex.) Now, what should we say the function of area 6 is? It isn't at all clear anymore how the brain is using this structure or for what purpose.

It could be the case that we simply haven't identified the function of area 6 and if we keep on doing the sort of subtraction studies that we are currently, then eventually we will find a unifying and pithy way to describe what premotor cortex is doing for us. In this instance, neuroscience would be on the right track to determining brain function, but we still have a long ways to go yet. But, it could also be the case that how a region functions depends heavily on the “neural context” (McIntosh Reference McIntosh1999, Reference McIntosh2000). Its functional role in our cognitive economy depends on how it is connected to other areas and how those other areas are responding. (The function of these areas would also be dependent on their particular connectivity and the current patterns of activation. And so it would go.) If this is correct, then searching for “the” function of particular areas is misguided, for different brain regions play different roles depending upon the cognitive tasks at hand.

In addition, as we all know, the underlying scientific principle of things like fMRI and PET is that inferences can be made due to indirect observations of metabolic changes. Consequently, they completely overlook all the components of cognition that do not have a corresponding change in metabolism, such as those due to small changes in small areas or volumes. For example, rapid (seconds to minutes) changes in membrane channel protein populations, which can lead to fundamental and distinct changes in the type and quantity of information a synaptic membrane sends or receives, are invisible in fMRI and PET. It is likely that the majority of cognitive processing occurs “below the radar” of inducing a metabolic change, and having the change be big enough that it can be detected indirectly.

We liken voxel and pixel analysis to having 1000 people in a large room. If you measure the oxygen consumed in this room you will not differentiate the following states, no matter how accurately you can measure the oxygen consumed: 500 people with red hats sleeping while 500 people with blue hats walk on a treadmill, 500 red hats on the treadmill while 500 blue hats sleep, or 250 red hats and 250 blue hats on the treadmill while 250 red hats and 250 blue hats sleep, and so on. But in order to understand what is actually going on in the room, we need to know which people are sleeping or exercising, and why. Measuring gross oxygen consumption won't tell us that.

The long and the short of it is that our experiments are not designed such that they give us much definitive functional information about any particular brain structure. Neuroscientists’ simplifying assumptions of discreteness and constancy of function are simply not justified, nor is it clear that they will be theoretically appropriate any time soon. These assumptions end up causing scientists to abstract over the neurophysiology improperly in that they deny diversity of function or multi-tasking all the way down the line prior to examining the data for exactly these possibilities.

As a result, neuroscientists cannot use the data they get to support their claims of function, for they are assuming local and specific functions prior to gathering appropriate data for that claim. But, gathering the appropriate data is simply beyond our kin at the moment, for we have no way to approach studying the brain except through a modularist lens. Basically, we are stuck theoretically and empirically in a counter-productive circle. What do all the brain data we have amassed tell us about how the brain works? Precious little thus far.