Not Final Version. Please quote and cite Philosophy of Science, 67, 444-465.


Shocking lessons from electric fish:

The theory and practice of multiple realization*

Brian L. Keeley†‡

Pitzer College


* Received June 1999; revised April 2000.

† Portions of this paper are taken from the author’s Ph.D. dissertation, written under the supervision of Sandra Mitchell and Patricia Churchland at the University of California, San Diego (UCSD), and finished in 1997. Early versions of this paper were presented to the Society for Philosophy and Psychology, the Southern Society for Philosophy and Psychology, and to the Philosophy/Neuroscience/Psychology Program at Washington University in St. Louis. John Bickle provided me with insightful suggestions; the last section of the paper would not have occurred without his input. Thanks also to an anonymous reviewer for this journal. Masashi Kawasaki, Calvin Wong, K.T. Moortgat, and Ted Bullock all provided feedback on the scientific details. However, any remaining errors are my responsibility. This work was supported by grants from the National Institute of Mental Health (NRSA #1 F31 MH10676-01), the McDonnell-Pew Center for Cognitive Neuroscience, and the UCSD Department of Philosophy.

‡ Send requests for reprints to the author, Pitzer College, 1050 North Mills Avenue, Claremont, CA 91711; email:



This paper explores the relationship between psychology and neurobiology in the context of cognitive science. Are the sciences that constitute cognitive science independent and theoretically autonomous, or is there a necessary interaction between them? I explore Fodor’s Multiple Realization Thesis (MRT) which starts with the fact of multiple realization and purports to derive the theoretical autonomy of special sciences (such as psychology) from structural sciences (such as neurobiology). After laying out the MRT, it is shown that, on closer inspection, the argument is either circular or self-undermining—the argument either assumes the very autonomy it seeks to demonstrate or the concluded autonomy is contradicted by the theoretical interdependence invoked by the premises of the argument. Next, I explore a concrete example of multiple realization in the explanation of animal behavior: the convergent evolution of jamming avoidance behaviors in three genera of weakly electric fish. Contrary to the image painted by the MRT, the work on these animals involves a high degree of interaction between the various levels of investigation. The fact that our understanding of electric fish behavior involves functional theories and multiple realization without the kind of disunified science that is supposed to follow from such a situation indicates that the mere fact of multiple realization cannot be the basis for an autonomous psychology.

1. Introduction

What is the relationship between the domain of psychology and the domain of neurobiology? More generally, what is the relationship between the various levels of science and between their target domains? For a quarter century, there has been an argument on the books that purports to show that higher level, so called "special sciences" are theoretically autonomous of lower level, "structural sciences;" that science is fundamentally disunified. The key fact in this argument is that proper explanations in these special sciences contain terms which are multiply realized in the terms of lower level sciences. In other words, special sciences make use of legitimate scientific generalizations that contain terms that can be physically constituted in a wide variety of ways.

This fact is relevant because it has been argued that multiple realizability renders the phenomena of such higher level sciences as psychology essentially opaque from the perspective of such lower level sciences as neurobiology. The canonical proponent of this argument, Jerry Fodor, presents the conclusion firmly: The physical constitution of special science generalizations is, "[...] entirely irrelevant to the truth of the generalizations, or to their interestingness, or to their degree of confirmation, or, indeed, to any of their epistemologically important properties" (1974/1975, 124, my emphasis). The generalizations of economics, psychology, etc., are best stated—indeed, can only be stated—in the terms of these sciences themselves. The study of how such high level generalizations are realized in physical, chemical, biological, or any other lower level terms has little, if anything, to tell us about these higher level phenomena. The notion is well-captured by Mitchell, et al (1997):

  • Relying on the essential incongruity of different taxonomies, [proponents of disunity] tend to argue that a question raised within a given conceptual framework can always be relevantly and satisfactorily answered without leaving that same framework. This would mean not just that higher sciences have no reason to fear preemption by more fundamental disciplines, but, furthermore, that they could afford not to care at all about what happens at the more fundamental level. (110)
  • Although this line of argument has been around for quite a while (see Putnam (1975) for a closely related argument), it has not lost much strength with age. Fodor himself still holds to its conclusions (see, for example, Fodor 1995). He is not alone. Horgan’s (1993) arguments owe a lot to it, as does Owens’ (1989). An unscientific survey of my bookshelf of works on cognitive science and philosophy of mind find an almost universal set of citations of Fodor (1974/1975) as one of the arguments for the autonomy of psychology. (The only works that do not cite Fodor are those which predate 1974.) It seems safe to say that the multiple realization thesis is alive and well.

    Much ink has been spilled attempting to counter the anti-reductionist components of the argument (cf., Churchland (1979), Hooker (1981), Bickle (1998)). In this paper, I intend to take a different tack and go after the argument’s implications for the theoretical autonomy of the special sciences. I shall argue that the argument from multiple realization to theoretical autonomy is not as strong as its proponents suggest. If anything, the study of multiply realized phenomena requires more, not less, interaction between the various levels of scientific investigation.

    In the next section, I lay out the multiple realization thesis and its alleged implications. Following this, I propose some reasons for doubting its validity. My strategy will then be to consider a concrete scientific case of multiple realization. So, in Section III, I take a close look at such a case: the convergent evolution of computational algorithms in several genera of weakly electric fish. This study will reveal how taxonomies at different levels of investigation coevolve, with discoveries at one level affecting the characterization of phenomena at other levels. In the final section of the paper, I deal with a variety of objections that arise for my case study and the conclusions I draw. In particular, I will address the objection that the study of fish is simply too unlike cognitive science to support the conclusions I draw. The point of the case study is to show that the mere presence of multiple realization in a theory does not by itself entail that the theory is independent of other levels of investigation. If psychology is indeed autonomous, it must be for reasons other than multiple realization of a functionally defined theoretical taxonomy—the traditional position within cognitive science.


    2. The multiple realization thesis explained, defended and undermined

    In his 1974 paper, "Special sciences, or the disunity of science as a working hypothesis," Fodor argues from the multiple realizability of functionally defined generalizations to the theoretical independence of those sciences that make use of functionalist theories (Figure 1). For the sake of brevity, I will refer to this argument as the multiple realization thesis (MRT). The argument begins by noting that there are true, functionally defined, theories about certain high-level phenomena. These are the theories of the so-called "special sciences," e. g., economics, sociology, linguistics, and psychology. The taxonomies of these theories include terms that are functionally defined—defined in terms of their relationship to other terms in the theory, not in terms of their physical constitution. These terms are used in true, special science generalizations, e.g., Gresham’s law in economics, and a plethora of psychological generalizations.

    {Figure 1 about here}

    Second, the functionally-defined terms of these taxonomies are multiply realized; that is, there are a vast variety of physical instances of which any given special science theory would be true. Such sets of functionally equivalent entities need not have anything physical in common.1 For example, in economics, Gresham’s law states that bad money drives out the good. "Money" in this generalization can be gold or silver coins, printed pieces of paper-thin cloth, glass beads, coconuts, what have you. If we were to translate Gresham’s law into the language of the physical sciences, we would not find a single term with which to translate the economic term "money." The best we could do would be to replace "money" with a disjunctive set of physical descriptions corresponding to the physically different kinds of money (small cylindrical mass of gold or small cylindrical mass of silver or...). The straightforward translation fails because the special science category "money" does not map onto any corresponding natural kind in the language of the physical sciences. Such special science categories are, in Fodor’s colorful phrase, "wildly disjunctive."

    Third, lacking such a natural kind description in the language of physical theory, the special sciences and their functional characterizations cannot be reduced to the physical sciences. Lawlike special science generalizations cannot be translated into lawlike physical science generalizations, therefore special science laws cannot be reduced.

    Fourth, the irreducibility of psychology and other special sciences gives them a significant degree of independence from lower level, structural sciences. It is the functional language alone that can explain what these entities have in common and express the regularities that exist between them. Any attempt to translate the phenomena into the language of a lower level thereby looses the succinctly stated regularity. Due to the vast variety of realizations, it is pointless to study a phenomenon at a level of analysis at which it is not evident. Therefore, contrary to the hopes of some (Oppenheim and Putnam 1958), science is disunified. There are as many independent, separate sciences as there are levels of multiply realizable phenomena.

    Schematically, we can represent Fodor’s argument thus:

  • P1 - There are true, high-level, functional theories, i.e., there are special sciences.

    P2 - Some terms in the taxonomies of these theories are multiply realized; that is,

  • a) There are a vast variety of physical instances, of which any given functional theory would be true,


    b) The realizing physical instances are wildly disjunctive; they have nothing physically in common.

  • P3 - Legitimate, scientific laws cannot include wildly disjunctive terms.

  • Therefore,

  • C1 - The special sciences cannot be reduced to lower level sciences.
  • and,

  • C2 - The special sciences are independent of lower level sciences, i.e., science is disunified.
  • As developed by Fodor, this argument has significant appeal. It allows us to distinguish between the so called "hard" and "soft" sciences in a way that avoids casting aspersions on the higher level endeavors. Fodor (1975) argues that he has eked out a place for psychologists to dwell that does not "vitiate the psychologist’s claim to study mental phenomena," as previous accounts of the special sciences have done (1). However, too much independence—too much disunity—may be as bad as the potential vitiation that motivates Fodor.

    However, there are problems with the MRT. This trouble does not lay with the second or third premises of the argument, which I am willing here to accept. The fact of multiple realization is seemingly unassailable in workaday science. We do not have to resort to wild, science fiction thought experiments concerning Martians or artificially intelligent computers to find it. In an almost trivial sense, different humans multiply realize the same psychological states.2 At the same time, I do not wish to call into question Fodor’s conception of the role of disjunctive terms in scientific laws.

    The conclusions of the MRT are more problematic. As for the first conclusion, if the translational identity of higher level predicates with lower level ones was the only picture of reduction available to us, then arguments for unified science would be in hot water indeed. While I agree with Fodor that identity through translation is not a viable notion of reduction, the real questions are, (a) whether other notions of reduction are available and (b) whether unity of science through reduction is the only sense of unity available to us. To get from the premises to C1, one needs a theory of reduction in addition to the stated premises.3 Fodor is explicitly arguing against the specific reductionism of Oppenheim and Putnam (1958). The question is whether C1 still follows when invoking a different theory of reduction. However, while interesting, this is not the question of the present paper.4 At the moment, I have another target, namely the second conclusion: the disunity of science. Regardless of the validity of C1, the second conclusion does not follow.

    It is the first premise—that there are true, scientific generalizations in functional terms—that I wish to examine. There is more to this innocuous seeming assumption than we are led to believe. My concern can be brought into focus by asking a simple question: How do we identify the functional kinds that are required to express special scientific generalizations? That is, how are special science taxonomies individuated? The choice of answers to this question, I claim, presents the proponent of MRT with a dilemma.

    On the one hand, if one claims that an individual special science alone has the right and responsibility to define its own taxonomy, then the MRT begins to look suspiciously circular, since a degree of autonomy is assumed by the argument. The power to define a taxonomy—the language in which generalizations will be stated—is substantial. A significant task of any science is to carve up its realm of discourse. Therefore, one important way in which sciences at different levels can interact is through negotiation of mutually compatible taxonomies. Even when compatible taxonomies is not the goal, it is often useful to consider how related levels carve up their phenomena when deciding how to carve up one’s own.

    When one interprets the first premise as holding that this kind of interaction is prohibited, then it is simply circular to conclude that this type of interaction between levels should not occur. For example, it would come as no surprise that the categories of psychology are incompatible with those of neurobiology, to take Fodor’s example, if neurobiology was denied any role in delineating those categories in the first place. Far from providing us with an argument for the disunity of science based on multiple realization, this interpretation of the first premise simply builds disunity into the first premise of the argument. At the very least, we need to be given some reason to believe that the degree of independence entailed by the MRT is not significantly due to an assumed degree of independence of special sciences to define their own categories without consultation.

    On the other hand, if the taxonomies of special sciences are developed by the special sciences in interaction with structural sciences, then the strength of the conclusion risks being significantly undermined. Unified science, in large part, consists in scientists at different levels of investigation working together to negotiate compatible taxonomies. Theorists at higher levels can suggest ways of grouping together phenomena that initially seem disparate from the point of view of the lower level, and vice versa. Data collected as part of investigations conducted at a lower level can motivate higher level theorists to redefine their ontologies in more precise ways. If we read the first premise as allowing this kind of give and take between theories at different levels of investigation, then the degree of independence entailed by the conclusion is at odds with the degree of interdependence suggested by the initial premise.

    We cannot conclude that special sciences are significantly independent and science disunified until we are clearer on the relationship between the levels entailed by the first premise. On my second reading of the MRT’s first premise, it hides an unspecified degree of intertheoretic interaction in the establishment of any given level’s scientific taxonomy. We cannot conclude that any given level of explanation is independent of other levels of explanation, if those other levels of explanation are consulted—off stage, as it were—in defining the theories employed at that given level. The image of science in which the various levels of scientific explanation work together to define mutually compatible taxonomies seems a far cry from the image of "disunified" science.

    Put bluntly, the dilemma I am posing is this: Either (1) the first premise simply assumes the disunity of science by assuming that the special sciences—and the special sciences alonehave the power to determine their taxonomies, in which case the MRT is circular, in that it assumes the very disunity it seeks to conclude, or (2) the first premise allows that the taxonomies of the special sciences are negotiated through the interaction of the special sciences with other sciences, in which case any disunity entailed by the rest of the argument is subverted by the unity invoked by the initial premise. Either way, the argument is not as powerful as it has traditionally been taken to be.

    In the next section, I want to suggest that the second horn of the dilemma is closer to being an accurate representation of how multiple realization plays out in the practice of science. It is important to consider real cases, because thought experiments and "toy" examples may lead us astray. In fact, it is often the devilishly difficult problem of understanding the behavior of complex systems that lead scientists to bridge levels of investigation (Crick 1988, 107). The consideration of real cases paints an image of a highly interactive science in which the language of explanation at one level is significantly influenced by evidence at other levels.

    3. The neuroethology of electroreception: a case study in multiple realization

    Neuroethology—the study of the neural basis of naturally occurring animal behavior—is an interdisciplinary endeavor that seeks to understand animal behavior at many different levels, combining neurobiology, ethology, computational modeling, and comparative biology (Keeley forthcoming-a). For purposes of illustration, I will focus on the neuroethology of weakly electric fish, although the conclusions I draw here are in no way restricted to these organisms. However, these animals represent a unique case in contemporary neurobiology. As Hardcastle (1996) notes, our understanding of these animals "... is widely regarded as one of the most complete neuroethological accounts of animal information processing to date" (49). The fact that this is our most complete understanding of a vertebrate system suggests that these animals are a good test bed for theories about what properties a mature scientific understanding should have.

    Consider three genera of fish: Eigenmannia, Apteronotus, and Gymnarchus. The members of all three genera are weakly electric fish; that is, they generate small electric fields using highly specialized "electric organs."5 Some aspects of the electric organ discharge (EOD) are under neural control and are thus part of the electric fish motor system. In addition, these fish possess sensory organs capable of detecting electricity. This combination of electromotor ability with an electrosensory capacity gives these fish a unique sensory system. By detecting small changes to their EOD caused by objects in their immediate environment, they collect sensory information about their world.6

    Neuroethologists, working in their capacity as ethologists and animal behaviorists, have discovered a number of regularities cast at the level of animal psychology. I will concentrate on one particularly robust behavioral regularity: the "jamming avoidance response". In wave-type electric fish, including the three genera in question, the electric organ is continuously discharged, but with a modulation that gives the resulting electric signal a sinusoidal character.7 Individual fish vary with respect to their baseline frequency; with variation occurring over a range that is determined by their species. Individual fish can control their EOD frequency and the most studied case of frequency change occurs when two fish of the same or similar frequency come into close physical proximity of one another.

    Sinusoidal waveforms have a curious property: when two sinusoidal waves of a similar frequency are added together, they interfere, resulting in a "beat" wave with a frequency equal to the difference between the two component waves (Figure 2a). You might imagine that a fish relying on its sinusoidal EOD to detect its environment would be adversely affected from having that signal effectively cancelled out by a neighbor with a similar frequency. If you indeed imagine this, you would be correct. Matsubara and Heiligenberg (1978) showed that the perceptual acuity of an electric fish presented with an interfering signal decreases as the difference (df) between the two frequencies—its own EOD and the interfering signal—decreases. Therefore, we say that such a fish is perceptually "jammed."

    {Figure 2 about here}

    To deal with this jamming, wave-type weakly electric fish have evolved a jamming avoidance response (JAR) in which each individual fish acts to change its EOD frequency to maintain a minimum df between its own frequency and that of its neighbors (Figure 2b; see Heiligenberg 1977). As an example, consider a scenario involving two fish: Fish 1 at 371Hz and Fish 2 at 370Hz. In the JAR, Fish 2 will shift its frequency downward, say to 365Hz, and Fish 1 will shift its frequency upward, say to 376Hz. The cognitive task in performing the JAR is to decide whether to shift one’s frequency up or down (Figure 3). This decision depends on which direction will increase the df. This, in turn, depends on whether one’s discharge is the higher or the lower frequency in the mix of the two signals detected at the surface of the skin. The JAR is a very robust behavior and is found in almost all wave-type genera, including not only Eigenmannia, but also Apteronotus and Gymnarchus.

    {Figure 3 about here}

    Walter Heiligenberg spent the better part of his career studying these animals. Eventually, he came to concentrate on the "computational rules" underlying the JAR in Eigenmannia (Figure 4, center column). He discovered that Eigenmannia solve the JAR using a complex computational algorithm. This algorithm represents an explanation of the JAR behavior. (I will present a sketch of this algorithm below. For those who wish to see more detail, see Heiligenberg, et al (1978); Heiligenberg and Bastian (1980); Heiligenberg (1991a,b).)

    {Figure 4 about here}

    Now, let us compare the three genera of weakly electric fish: Eigenmannia, Apteronotus, and Gymnarchus. The species of all three genera are electroreceptive. They are also all wave-type weakly electric fish. Fish of all three genera exhibit the JAR. Finally, the fish of all three genera carry out the JAR using the same computational algorithm. Nonetheless, there are important differences between these genera; differences with respect to how electroreception, the JAR, and the computational algorithm are physically realized in them. As we will see below, due to the process of natural selection, important components of Heiligenberg’s JAR algorithm are realized differently in the three genera of electric fish discussed here. These three genera of fish then multiply realize the JAR behavior and the computational algorithm that carries it out. Let us now turn to a discussion of these differences.

    First, Apteronotus differs from the other two genera in the nature of its electric organ (EO). In Eigenmannia and Gymnarchus, the EO is myogenic; that is, it is developmentally derived from muscle tissue. The EO of Apteronotus is neurogenic, i.e., derived from neural tissue. Although neurogenic and myogenic electric organs share many aspects of their gross anatomy, there are also significant differences in the physiology by which they generate electrical discharges. The most striking difference between myogenic and neurogenic EOs can be found in the effect of curare, a paralytic drug that acts specifically on the neuromuscular junction. Appropriate injections of curare will both paralyze the animal and silence the EOD in Eigenmannia and Gymnarchus. In Apteronotus, curare will effectively paralyze the animal, but the EOD is left unaffected. The differential affect of curare is a direct result of the compositional differences between the two types of electric organ (Bass 1986).

    There are more differences: Eigenmannia, Apteronotus, and Gymnarchus do not all share a close phylogenetic history. The electroreceptive sensory modality has independently evolved at least four different times in history, and these three genera come from separate, but convergently evolved lineages. Electroreception and the JAR in Eigenmannia and Apteronotus, on the one hand, and Gymnarchus, on the other, are the results of convergent evolution (Bullock, et al 1982, 1983; Northcutt 1986). Evidence suggests that the most recent common ancestor of these two lineages was not electroreceptive (Lauder and Liem 1983). As might be expected, evolution has led to different physical realizations of the JAR in these two orders.

    In Eigenmannia, the steps of the JAR computational algorithm are implemented in its nervous system in the following way (Figure 4, left side): Step (1) is carried out by two classes of electrosensory cells: P-type electroreceptors track changes in electrical amplitude, firing more often the stronger the electric signal presented to them. T-type electroreceptors, on the other hand, fire a single volley at the beginning of each cycle of a sinusoidal signal, and hence code phase information. Step (2)—integrating the locally collected amplitude and phase information in both time and space (over the surface of the fish)—occurs in the electrosensory lateral line lobe (ELL) located in the hindbrain. The ELL contains three different somatotopic maps that vary in their degree of spatial and temporal resolution and sends the results of its computations to the torus semicircularis (TS) in the midbrain. At this point, three synapses away from the sensory periphery, the real action begins. Collecting and comparing information from all over the body surface, the higher layers of this laminated structure carry out step (3). This step involves determining four things: a) when and where phase is advancing, b) when and where phase is delaying, c) when and where amplitude is increasing, and d) when and where amplitude is decreasing. In the lower lamina of the TS, a complex "neuronal democracy" is instantiated in which thousands of pair-wise comparisons of the results of step (3) are simultaneous made. These comparisons constitute step (4). Cells in the TS then project to the nucleus electrosensorius, where the results of the neuronal election are tallied and it is determined whether the EOD frequency should either be raised or lowered (step (5)). This result is put into action by the cells of the prepacemaker nucleus, which modulates the baseline frequency of the pacemaker nucleus (Pn). The Pn is endogenously oscillatory and controls the firing of the EO with a one-to-one mapped command signal, i.e., the frequency of the Pn is identical with the resulting frequency of the EO.

    The algorithm and structures involved in Apteronotus are the same (with the exception of the electric organ itself, as discussed above). This is not surprising given the close evolutionary relationship these two genera share (Alves-Gomes, et al 1995, Alves-Gomes 1996).

    A little more detail will round out the picture. Since Gymnarchus evolved this behavior separately, the neural basis of the JAR differs (Figure 4, right side). We know considerably less about Gymnarchus than about Eigenmannia and Apteronotus, but Masashi Kawasaki, a former student of Heiligenberg’s, has been able to identify the same algorithm underlying the JAR in Gymnarchus.8 One anatomical difference can be found in the receptor cells—the S-type and O-type electroreceptors—which are morphologically distinct from the P-type and T-type cells found in Eigenmannia and Apteronotus, but which function in much the same way (Bullock, et al 1975). S-type receptors are tuned to the amplitude of electric stimulation, whereas the O-type fire in a phase-locked manner. The primary difference in Gymnarchus is found in the ELL. While in Eigenmannia and Apteronotus this structure primarily serves the role of bringing together electrical information from larger and larger areas of skin surface, in Gymnarchus all of the important steps of the JAR algorithm are performed here. Relative to the ELL of Eigenmannia and Apteronotus, the ELL of Gymnarchus has a more complex computational architecture (and, not surprisingly, a more complex neural architecture, as well).

    The information in Table 1 summarizes the similarities and differences discussed above. The convergent evolution of the jamming avoidance response in the three genera of fish just described represents a concrete instance of multiple realization in the study of animal behavior. There are important similarities in their behavior and in the computational architectures implemented by different neural mechanisms. At the same time, there are significant differences in the ways that natural selection has designed the mechanisms of these fish. All three types of fish carry out the same sort of behaviors, following the same algorithm, but they do it with different neural structures.

    {Table 1 about here}

    4. Implications of the case study for the autonomy of levels

    What are the implications of this case for the issues discussed in this paper? First and foremost, my reason for discussing electric fish is that is a clear and concrete example of multiple realization. Assuming that multiple realization is not restricted to the worlds of artifacts and thought experiments, if different physical instantiations of the jamming avoidance response behavior and the computational algorithm that underlies it is not an example of multiple realization in the biological world, I don’t know what is.

    Second, the scientific practice involved in the explanation of animal behavior is far from being a disunified collection of autonomous enterprises. We do not find ethologists plugging away at potential computational algorithms while ignoring the work of fish neurobiologists. Similarly, we do not find neurobiologists tracing pathways and recording from neurons in ignorance of the work of ethologists and others working at the level of whole organisms and their behavior. Instead, the science of neuroethology is highly interdisciplinary and interactive. Research in this field combines the interests of neurobiologists, ethologists, developmental biologists, evolutionary biologists, and psychophysicists, to name only the biggest participants.

    But the reason for this interaction is not simply a laudable spirit of camaraderie. Quite the contrary, there is a real need for such interaction because of the difficulty of the problems facing neuroethologists. In order to piece together an explanation of even as simple a behavior as the JAR in as simple an animal as the electric fish, data from many domains are required. Many domains of investigation informed the computational algorithm discovered by Heiligenberg. That the JAR was a significant behavior for the species and might be controlled by dedicated brain structures was contributed through the discovery of this behavior in the first place. The discovery of the response properties of two different classes of receptor cells suggested that the key inputs to the computational algorithm had something to do with the amplitude and temporal properties of the electric signal. Work at every level informed the others. Our understanding of the behavior, the neurobiology, and the computational algorithm coevolved together.

    The picture presented here looks like the second horn of my dilemma for the would-be proponent of the MRT: In the course of scientific practice, the theoretical taxonomy at any given level is continuously being negotiated between the sciences of the various levels involved. Without ethology, neurobiologists would not realize that the electrosensory lateral line lobe and torus semicircularis were structures for computing the steps required to carry out jamming avoidance behavior. Without neurobiology, ethologists would not have known that a taxonomy of inputs stressing amplitude and phase was crucial. What counts as a component of an explanation at one level of explanation is the result of input from many different levels of investigation. It is this large degree of interaction that undermines any "disunity" suggested by arguments such as the multiple realization thesis.

    The taxonomies at the different levels of explanation employed by neuroethologists have coevolved as a result of interaction between the levels. When scientific results at one level forces a change in that level’s taxonomy, these changes often percolate up or down to adjoining levels of explanation. Also, when a taxonomic discovery at one level occurs, this then immediately poses related taxonomic questions at other levels. For example, the discovery that Gymnarchus had convergently evolved a specific class of behaviors—a wave-type EOD and the JAR—it immediately opened the question of whether its computational algorithms and neural structures had similarly convergently evolved. These are just a few of the ways in which taxonomies at different levels can and do coevolve.

    5. Objections and conclusions

    In this final section of the paper, I wish to turn to some potential objections to the story I have so far presented. The most obvious worry is that, in this paper, I explicitly set out to explore the autonomy of psychology within the cognitive sciences, only to spin a tale of fish. One may justifiably wonder what fish have to do cognitive science.

    This objection can be broken down into two related parts. First, there is what I call the "neuroethology-ain’t-cognitive-science" worry. I have been discussing the science of neuroethology and suggesting that lessons drawn from it will apply to cognitive science. Given that cognitive science and neuroethology are different sciences with different target domains, this application of one to the other needs justification. Cognitive science has traditionally taken as its target domain intelligent behavior in humans. As Butler (1994) puts it, "The long-term goal of [research in cognitive science] is to explain human behavior by appeal to the cognitive capacities, or computational tasks, performed in the service of behavior. [...] The short term goals, defined by the projects actually undertaken by practicing cognitive scientists, is the explanation of individual cognitive capacities in terms of the functions that are computed in the service of these capacities" (130). Humans and human behavior are the touchstones of cognitive science.

    Neuroethology, on the other hand, is the study of the neural bases of animal behavior, and it typically concentrates on such animals as leeches, lobsters, bats, owls, and moles, not to mention the fish discussed herein. Furthermore, neuroethology is a biological science; it is concerned with the evolution of animals and behavior, much more so than traditional cognitive science is concerned with evolution. On the face of it, the two sciences appear to be very different, and I need to tell more of a story about how the two are relevantly related.

    My response is simply to refuse to concede that cognitive science and neuroethology are as different as the objection suggests. It is true that the two sciences have different target domains, but the structure of the two fields is remarkably similar. Replace "human behavior" with "non-human animal behavior" in the Butler quotation above, and one would have an excellent definition of neuroethology.

    The two fields share many affinities. First, they both ask many of the same questions about mechanisms and behavior, but where neuroethology asks those questions of all animals, cognitive science is primarily focused on humans. Second, both are highly interdisciplinary, and feature a very similar constitutive structure: Neurobiology and computational modeling play much the same role in each science. The contribution of developmental biology to neuroethology is analogous to the contribution of developmental cognitive psychology to cognitive science. Ethology’s focus on naturally occurring behavior is reflected in anthropology’s focus on natural human instances of cognition (e.g., Hutchins (1995)). Ethology’s focus on laboratory studies of the behavioral capacities of animals is reflected in psychology’s focus on the behavioral capacities of humans under controlled laboratory conditions. Finally, while philosophy may play a much more visible role in cognitive science, it also plays a role in neuroethology.9 From the perspective of philosophy of science, these two fields appear to be put together in much the same way.10

    The second part of the objection has less to do with the nature of the sciences involved than with the phenomena being studied. Cognitive science is the science of cognition. One may be tempted to insist that the JAR in weakly electric fish is not even a psychological phenomenon, much less a genuinely cognitive one. And, who would doubt that this ethological phenomenon can be fruitfully addressed physiologically? All that the arguments presented here show, the objection continues, is that the study of non-cognitive behavior needs to be highly integrated. This tells us nothing about the sciences of the cognitive. Let us call this second objection the "mere-electric-fish-behavior" worry.11

    The mere-electric-fish-behavior objection—that electric fish behavior is not "cognitive" enough—is a curious one. First, I might note in passing that Michael Tye (1997) has recently argued that, on his theory at least, fish are conscious. Second, I am not alone in seeing the relevance of work on electric fish to cognitive science. Hardcastle (1996) sees neuroethology, and particularly the work of Heiligenberg, as providing an object lesson on how to build a theory in cognitive science: "I suppose one may dispute whether electric fish are truly cognitive, but I still take his [Heiligenberg’s] research to present an exemplar of the type of model one would like to have of a thinking brain" (49, emphasis in original).

    But these—perhaps sophistic—endorsements aside, I want to concede that it is possible to attempt a rescue the autonomy of psychology by postulating the unique status of cognition. However, I suspect that the price of this move may be too high. In any case, I will now show that, while this avenue is open to the would-be cognitive autonomist, it is necessarily a different argument than the MRT.

    The question raised by this line of objection is this: What is "special" about the special sciences? On Fodor’s original account, the characteristic property of special science theorizing is that these accounts necessarily make use of terms that are multiply realized. Using the MRT, the autonomy of the special sciences is then derived from this fact allegedly peculiar to these sciences. The essence of the MRT is that where there is multiple realization, there is irreducibility and autonomous high level sciences. This is the argument that has been almost universally held to be the right argument, and this is the argument I have targeted here.

    Yet, the mere-electric-fish-behavior objection changes this argument. In shifting the focus away from multiple realization per se to the unique nature of cognitive phenomena, this objection simply replaces the MRT with a new—and unexplicated—argument. On this new reading, multiple realization becomes an accidental property of psychology, and not the key feature which guarantees its autonomy, because it concedes that the JAR and the Heiligenberg computational algorithm are multiply realized in different kinds of fish without any resulting barrier to reduction or theoretical unification. Multiple realization is neither a necessary nor a sufficient condition for theoretical autonomy; instead, some special property of cognitive agents is (their intelligence, language, or consciousness perhaps).

    If I will be allowed the pun, according to the mere-electric-fish-behavior response, the issue multiple realization becomes a red herring. On this account, what is "special" about the special sciences is not the sciences themselves (as in the MRT) but what the special sciences are sciences of. In raising the mere-electric-fish-behavior worry, the ground of the autonomy argument shifts. Instead of the special sciences being special per se, the special sciences are the sciences of special things. Furthermore, these things are not special by virtue of being multiply realized. While this argument may well have its merits, it is a different argument than the Fodorian argument that has been the bedrock of psychological autonomy for the past quarter century.

    This new version of the argument is less a philosophy of science argument about the nature of explanation in general, but rather a philosophy of mind argument about the alleged nature of minds (or cognitive agents). This is very different than the story that Fodor spins in his classic paper. In that paper, psychology is not singled out as a unique science with special properties because of its special domain of study. Quite to the contrary, psychology is identified as one science among many—the special sciences—that all have an autonomous status due to the kinds of sciences they were: sciences that necessarily predicated taxonomies that contained multiply realized, functionally defined theoretical entities. Indeed, a significant aspect of the popularity of Fodor’s argument was the very fact that it gave psychology autonomy without adverting to the special status of minds and mental stuff, as Cartesian arguments did.

    So, we are left with the conclusion that the feature of special sciences that was supposed to guarantee their theoretical autonomy—the multiple realization of some of their theoretical terms—offers no such guarantee. Other sciences have multiple realization without theoretical autonomy, as the convergent evolution of jamming avoidance algorithms in weakly electric fish clearly demonstrates. If the autonomy of the special sciences is to be preserved—assuming it is wise to do so—it must be based on something else. However, I am willing to live with a philosophy of science in which the special sciences have an equally important role to play as the structural sciences, but which grants neither a more primary role. That is the lesson taught by the neuroethology of electric fish. I will leave it to the reader to judge how shocking this lesson is.



  • Akins, Kathleen (1993), "What is it like to be Boring and Myopic?" in B. Dahlbom (ed.), Dennett and His Critics: Demystifying Mind. Oxford: Blackwell, 124-160.

    Alves-Gomes, Josť A. (1996), The Phylogeny and Evolutionary History of the South American Electric Fishes (Order Gymnotiformes). Unpublished Ph.D. thesis, University of California, San Diego.

    ———, G. Ort‚ M. Haygood, W. Heiligenberg, and A. Meyer (1995), "Phylogenetic Analysis of the South American Electric Fishes (Order Gymnotiformes) and the Evolution of Their Electrogenic System: A Synthesis Based on Morphology, Electrophysiology, and Mitochondrial Sequence Data", Molecular Biology and Evolution 12: 298-318.

    Bass, Andrew (1986), "Electric Organs Revisited", in Bullock and Heiligenberg (1986).

    Bechtel, William and J. Mundale (1999), "Multiple Realizability Revisited: Linking Cognitive and Neural States", Philosophy of Science 66: 175-207.

    Bickle, J. (1998), Psychoneural Reduction: The New Wave. Cambridge, MA: MIT Press (A Bradford Book).

    Bullock, Theodore Holmes, K. Behrend, and W. Heiligenberg (1975), "Comparison of the Jamming Avoidance Responses in Gymnotoid and Gymnarchid Electric Fish: A Case of Convergent Evolution of Behavior and its Sensory Basis", Journal of Comparative Physiology 103: 97-121.

    ———, D. A. Brodznick, and R. G. Northcutt (1983), "The Phylogenetic Distribution of Electroreception: Evidence for Convergent Evolution of a Primitive Vertebrate Sense Modality", Brain Research Reviews 6: 25-46.

    ———, R. H. Hamstra, and H. Scheich (1972), "The jamming avoidance response of high frequency electric fish: I. General features", Journal of Comparative Physiology, 77, 1-22.

    ———, and W. Heiligenberg (eds.), (1986), Electroreception. New York: John Wiley & Sons.

    ———, R. G. Northcutt, and D. A. Brodznick (1982), "Evolution of Electroreception", Trends in Neuroscience 5: 50-53.

    Butler, Keith (1994), "Neural Constraints in Cognitive Science", Minds and Machines 4: 129-162.

    Churchland, Paul M. (1979), Scientific Realism and the Plasticity of Mind. Cambridge Studies in Philosophy. Cambridge: Cambridge University Press.

    Crick, Francis (1988), What Mad Pursuit: A Personal View of Scientific Discovery. New York: Basic Books.

    Dennett, Daniel C. (1978), "Why Not the Whole Iguana?" Behavioral and Brain Sciences 1: 103-104.

    Fodor, Jerry A. (1974/1975), "Special Sciences, or the Disunity of Science as a Working Hypothesis". Originally appeared in Synthese 28: 77-115. Reprinted in Fodor (1975, 9-25) and Fodor (1981, 127-145). Fodor (1975) cited here.

    ———. (1975), The Language of Thought. Cambridge, MA: Harvard University Press.

    ———. (1981), Representations: Philosophical Essays on the Foundations of Cognitive Science. Cambridge, MA: MIT Press (A Bradford Book).

    ———. (1995), "West Coast Fuzzy: Why We Don’t Know How Brains Work", Times Literary Supplement 4821: 5-6. (Review of Paul M. Churchland’s Engine of Reason, the Seat of the Soul.)

    Hardcastle, Valerie Gray (1996), How to Build a Theory in Cognitive Science. Albany, NY: State University of New York Press.

    Heiligenberg, Walter (1977), Principles of Electrolocation and Jamming Avoidance in Electric Fish: A Neuroethological Approach. Berlin: Springer-Verlag.

    ———. (1991a), Neural Nets in Electric Fish. Cambridge, MA: MIT Press (A Bradford Book).

    ———. (1991b), "The Jamming Avoidance Response of the Electric Fish, Eigenmannia: Computational Rules and Neuronal Implementation", Seminars in the Neurosciences 3: 3-18.

    ———, C. Baker, and J. Matsubara (1978), "The Jamming Avoidance Response in Eigenmannia Revisited: The Structure of a Neuronal Democracy", Journal of Comparative Physiology 127: 267-286.

    ———, and J. Bastian (1980), "The Control of Eigenmannia’s Pacemaker by Distributed Evaluation of Electroreceptive Afferences", Journal of Comparative Physiology 136: 113-133.

    Hooker, Clifford A. (1981), "Towards a General Theory of Reductionism; Part III", Dialogue 20: 496-529.

    Horgan, Terence (1993), "Nonreductive Materialism and the Explanatory Autonomy of Psychology", in S. J. Wagner and R. Warner, (eds.) Naturalism: A Critical Appraisal. Notre Dame, Indiana: University of Notre Dame Press, 295-320.

    Hutchins, Edwin (1995), Cognition in the Wild. Cambridge, MA: MIT Press.

    Kawasaki, Masashi (1993), "Independently Evolved Jamming Avoidance Responses Employ Identical Computational Algorithms: A Behavioral Study of the African Electric Fish, Gymnarchus niloticus", Journal of Comparative Physiology A 173: 9-22.

    ———. (1994), "The African Wave-type Electric Fish, Gymnarchus niloticus, Lacks a Corollary Discharge Mechanisms for Electrosensory Gating", Journal of Comparative Physiology A 174: 133-144.

    ———. (1996), "Comparative Analysis of the Jamming Avoidance Response in African and South American Wave-type Fishes", Biological Bulletin 191: 103-108.

    ———, and Y.-X. Guo (1996), "Neuronal Circuitry for Comparison of Timing in the Electrosensory Lateral Line Lobe of the African Wave-type Electric Fish, Gymnarchus niloticus", The Journal of Neuroscience 16: 380-391.

    ———. (1998), "Parallel Projection of Amplitude and Phase Information from the Hindbrain to the Midbrain of the African Electric Fish, Gymnarchus niloticus", Journal of Neuroscience 18: 7599-7611.

    Keeley, Brian L. (1999), "Fixing Content and Function in Neurobiological Systems: The Neuroethology of Electroreception", Biology & Philosophy 14: 395-430.

    ———. (forthcoming-a), "Neuroethology and the Philosophy of Cognitive Science", Philosophy of Science 67 (proceedings).

    ———. (forthcoming-b), "Making Senses of the Senses: Individuating Modalities in Human and Other Animals."

    Lauder, G. V. and K. F. Liem (1983), "Patterns of Diversity and Evolution in Ray-finned Fishes", in R. G. Northcutt and R. E. Davis (eds.), Fish Neurobiology, vol. 1, 1-24. Ann Arbor: University of Michigan Press.

    Lloyd, Dan (1989). Simple Minds. Cambridge, MA: MIT Press (A Bradford Book).

    Matsubara, J. and W. Heiligenberg (1978), "How well do Electric Fish Electrolocate under Jamming?" Journal of Comparative Physiology A 12: 285-290.

    Mitchell, Sandra D., L. Daston, G. Gigerenzer, N. Sesardic, and P. B. Sloep (1997), "The Whys and Hows of Interdisciplinarity", in P. Weingart, S.D. Mitchell, P. Richerson, and S. Maasen, (eds.), Human by Nature: Between Biology and the Social Sciences. Mahwah, New Jersey: Lawrence Erlbaum Associates.

    Northcutt, R. G. (1986), "Electroreception in Nonteleost Bony Fishes", in Bullock and Heiligenberg (1986, 257-285).

    Oppenheim, Paul and H. Putnam, (1958), "Unity of Science as a Working Hypothesis". Minnesota Studies in the Philosophy of Science, vol. II, 3-36.

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

    Putnam, Hilary (1975), "Philosophy and Our Mental Life", in Mind, Language, and Reality. Philosophical Papers, vol. 2. Cambridge: Cambridge University Press.

    Tye, Michael (1997), "The Problem of Simple Minds: Is There Anything it is Like to be a Honey Bee?" Philosophical Studies 88: 289-317.

  • Footnotes

    1. The notion of "functional equivalence" is due to Putnam (1975), although he prefers the term "functional isomorphism."

    2. Although, see Bechtel and Mundale (1999). They derive conclusions very similar to mine through a sustained attack on the second premise of Fodor’s argument and the intelligibility of his notion of multiple realization. While I am sympathetic to their arguments, this is not the course I pursue here.

    3. John Bickle noted this in a spoken commentary on an earlier draft of this paper presented to the 1998 Society for Philosophy and Psychology annual meeting.

    4. Bickle (1998, Chapter 4) addresses this question directly and at length.

    5. They are called weakly electric fish to distinguish them from their more powerful cousins, the strongly electric fish, such as the South American electric eel or the Mediterranean Torpedo. The electrical discharges of weakly electric fish are too small to be detected with the naked hand. In pet stores, Eigenmannia are typically known as "glass knife" fish. Two common species of Apteronotus are "brown" and "black ghosts." Gymnarchus are harder to find in pet stores, but when they are, they are found under the African name, "aba aba."

    6. See Keeley (1999) for a discussion of the 20th century discovery of the non-human modality of electroreception. Keeley (forthcoming-b) explores the issue of individuating non-human sensory modalities.

    7. Not all weakly electric fish are wave-type. Many are pulse-type. In pulse-type electric fish, the electric organ is discharged in a quick pulse—allowing a "snap-shot" of the world, as it were—instead of being discharged continuously. I will not discuss these fish here.

    8. The lack of corollary discharge (i.e., an efference copy) in Gymnarchus niloticus was demonstrated in Kawasaki (1994). The hypothesis that the JAR algorithm is identical in Eigenmannia, Apteronotus, and Gymnarchus is developed primarily in Kawasaki (1993); see also, Kawasaki (1996). Further details of the neural implementation of the JAR in Gymnarchus niloticus are presented in Kawasaki and Guo (1996, 1998).

    9. See, for example, Dennett (1978), Lloyd (1989), Akins (1993), and Tye (1997).

    10. I make the case for the structural similarities between neuroethology and cognitive science more fully in Keeley (forthcoming-a).

    11. The term is due to Bickle, who writes of the "mere-sea-slug-representation" objection to attempts to explain the phenomena of learning and memory via studies of Aplysia californica (Bickle 1998).



    Figure 1. A diagrammatic representation of multiple realization: A special science law is stated in the terms of that special (Sna). The special science terms can be realized in a variety of ways when described in the terms of the lower level science (Pna). Therefore, we say that the terms of the special science are best translated into disjunctive predicates in the terms of the lower level science. The single special science law translates into a plethora of seemingly unrelated lower level regularities. (Adapted from Fodor (1974/1975, 20).)


    Figure 2. The Jamming Avoidance Response Explained: A) The problem: When two fish of similar electric organ discharge (EOD) frequency are in close physical proximity, both fish are perceptually jammed. In the diagram, Fish 1 has an EOD frequency slightly higher than that of Fish 2. The combination of these two sinusoidal signals in the environment results in interference pattern with a "beat" frequency equal to frequency difference (df) of the two signals. This interference results in a loss of perceptual acuity for both fish, i.e., both fish are "jammed." B) The solution: Based solely on information collected from their perceptually jammed electrosensory system (see Figure 3), the fish with the higher frequency (Fish 1) will increase its EOD frequency, while the fish with the lower frequency (Fish 2) will lower its EOD frequency. This increases the df and decreases the amount of interference between the two signals and thus eliminates the jamming. Abstracting from experimental results (Bullock, et al 1972, 19), the EOD frequencies of two fish are tracked over time as their separate aquaria are alternately electrically connected and isolated with conductive wires. When connected, the two fish shift their EODs away from one another (this is the Jamming Avoidance Response). (In this original experiment, it was not possible to tell which fish was which.) When isolated, they drift back towards their original frequencies.


    Figure 3. The essence of the Heiligenberg Algorithm: The perceptual task confronting a jammed fish is to determine whether it needs to shift its frequency up or down. This, in turn, requires the fish to extract from its perceptually jammed electrosensory system whether its own contribution to the combined signal is higher or lower in frequency. According to Heiligenberg (1991a), this feat is accomplished by taking advantage of subtle differences between the detected signal at different points on a given fish’s body. This difference is a result of the fact that the electrosensory cells scattered across the body of the fish are directionally sensitive; they are better at detecting a signal perpendicular to the local skin surface than one parallel. Therefore, whereas the EOD signal of Fish 1 is relatively uniform in its contribution to the detected signal at both point A and B, the contribution of Fish 2 is stronger at A than at B. As a result, comparing the signal detected at different points on the body will reveal differences in amplitude and timing ("phase"). Although this telltale difference seems straightforward, to reliably extract it and utilize it under the plethora of jamming scenarios (different relative frequencies, amplitudes, locations, and orientations) requires a relatively complicated algorithm. This neural computation involves a number of steps (see Figure 4). Instantiating what Heiligenberg refers to as a "neuronal democracy," in several steps, the phase and amplitude of the signal detected all over the body are compared in stages and in parallel. This, then allows the fish to compute whether it is contributing the lower or higher EOD frequency to the detected signal, which in turn tells it whether it needs to shifts its frequency up or down to alleviate the jamming. (Adapted from Heiligenberg (1991a, 28-29)


    Figure 4. Comparison of anatomical instantiations of Heiligenberg’s jamming avoidance response algorithm in three genera of weakly electric fish: The computational algorithm (center column) describes the transformation of sensory representations into appropriate behavior (see Heiligenberg 1991a,b). This computational architecture is closely mirrored in the neuroanatomy of individuals of two New World genera, Eigenmannia and Apteronotus (left column). The same computational architecture is realized quite differently in the neuroanatomy of individuals of the African genus, Gymnarchus (right column).