1. Introduction
As a philosopher doing research on consciousness, my great fortune is to witness the very cutting edge of scientific research. I have learned about the importance of all kinds of things I otherwise would have found mystifying: thalamocortical loops and synchronous oscillatory neural firing, binocular rivalry experiments and change-blindness. Over the last 20 years, the hard problem of understanding how a brain produces consciousness has advanced from entirely intractable, and disreputable as a field of empirical investigation, to the point where leading neuroscientists such as Christof Koch claim that the problem is purely empirical and researchers should no longer be bothered by philosophical issues.Reference Koch1
Claims like this are hasty. Science is never done with philosophical problems, least of all the science of the mind. This article will review some of the ways data-gathering in neuroscience is useful and ways in which it is suspect. Section 2 begins concretely with an examination of the most commonly used and frequently criticized data collection device – fMRI. In Section 3, data becomes the mind itself in one of the most influential neuroscientific theories of consciousness, the Information Integration Theory. The serious philosophy begins in Section 4 with a general critique of information alone as a source of meaning, and Section 5 describes a way to extract meaning and consciousness from information by considering function. The mind, including consciousness, is designed to inform itself about its world, so any theory of the mind must explain how information is meaningful. I will argue that the meaning of consciousness is grounded in its function to represent the world as it appears at the present moment. fMRI and other neuroscience data are useful to the extent that they prove meaningful, both through proper methodology as well as philosophical foundation.
2. fMRI: What it Can and Cannot Show
Due to the advances in neuroscience and their implications for understanding the mind and its operations, the media regularly report results from technology with an increasing array of abbreviations: EEG, fMRI, TMS, DBS. One of the most familiar sources of data is the fMRI with its dark oval brains lit in shades of red, yellow, blue and purple.
fMRI stands for functional Magnetic Resonance Imaging, and it operates by sending a magnetic pulse that resonates the iron in blood, resulting in a faint but detectable signal. The well-known BOLD signal – blood-oxygen level dependent signal – is specifically attuned to the ratio between oxygenated and deoxygenated blood. The assumption is that neural activity lowers this ratio and thereby decreases the speed of decay in the signal. Physiologically, the reason for a lower ratio is the flow of blood, which increases following neural activity. Glucose and oxygen provided by blood flow restores energy to the depleted neurons. The dynamic of neural firing and metabolic restoration causes the lowered oxygenation ratio that the fMRI detects.Reference Aguirre 2
The details of fMRI operations are important to understanding some of the criticisms of the data. In a recent report, Martha Farah argues that several objections to fMRI are unwarranted.Reference Farah 3 One worry is that fMRI is only an indirect measure, because blood flow is used to indicate neural activity. Such an objection cannot be taken seriously unless we are also willing to rule out all kinds of ways that scientists make discoveries of one thing by investigating a correlated thing. To take just one case, most of subatomic physics research would be eliminated by this standard. The important issue is whether there is good reason to think the measure does indeed indicate the target phenomena, and there is. The relation between the BOLD signal and neural activity is well-established.
On the other hand, Farah notes other objections that should be addressed. The first concern is with reverse inference, which moves from brain activation to psychological process. When the amygdala is activated, for example, researchers often infer that emotions are involved.Reference Bonnet, Comte, Tatu, Millot, Moulin and Medeiros de Bustos 4 Legitimate reverse inference is possible, but only if prior research is done to validate forward inferences from the targeted psychological process to brain activation. In other words, you need to first determine how a particular brain area performs the function at issue, and then you can infer that when that brain area is activated, it is serving the pre-established function. So in the case of amygdala activation, reverse inference is supported by research showing that emotion processing produces amygdala activity.
A second area of concern is due to the statistics of large numbers, which is an integral part of fMRI analysis. Each image is composed of about 100,000 voxels, units of measurement that represent a neural population within an area of about 1–3 mm (Ref. Reference Aguirre2, pp. S10–11). A voxel is like a pixel on a computer screen that creates an image in combination with an array of other pixels. Owing to the high number of voxels and the fact that the brain is always globally active, there is a good chance that a number of the active voxels are unrelated to the psychological process under investigation. In a vivid illustration of this problem, Farah tells the story of a dead fish that was put in a scanner with instructions to think about the emotions depicted in photographs shown. When the results were subjected to multiple analyses, it turned out that brain areas correlated to perspective-taking were activated. The researchers, Craig Bennett and colleagues,Reference Bennett, Baird, Miller and Wolford 5 showed that proper statistical methods prevent this sort of spurious result, so Farah concludes that the statistics of large numbers is not an inherent problem with fMRI technology.
Nonetheless, it is worth noting that both of the problems that Farah raises involve potentially inappropriate utilization of the data. While Farah is right that these problems can be corrected with proper methodology, the recent ‘replicability crisis’ in psychology casts doubt on the structures of scientific practice to effectively implement the required restrictions.Reference Pashler and Wagenmakers 6
In commenting on this crisis, Roger Giner-Sorolla highlights the demand for perfect results created by the hypercompetitive atmosphere of scientific publication and grant review.Reference Giner-Sorolla 7 The ever-increasing requirement for publication numbers means that, exponentially, more articles are vying for limited journal space. Likewise, the expectation that scientists fund their own labs with a shrinking pot of grant money means that the pressure to obtain significant data is extreme. Add to this the expense and time on an fMRI machine, and the consequences of a failed experiment are untenable. Policies need to be put in place as a check on this temptation: for example, by establishing a process for external statistical review of raw data and support for replication experiments to test the results.
Furthermore, there are structural limitations to fMRI that Farah does not consider. FMRI can show a brain area is active, but it cannot determine why. Measurements indicating greater activation quantify population size rather than activity level or functional significance.Reference Logothetis 8 There is no way to say what function that brain area performs relative to the task or if it performs any function at all. The scan only shows that there is activity correlated with the task.
More fundamentally, fMRI protocols often fail to recognize the importance of individual structural and functional differences. Statistical techniques average data across multiple trials and multiple subjects in order to smooth out the inevitable mental distractions or hardware glitches that disrupt the procedure. It is useful to pause a moment to appreciate the challenge of this data-gathering process. The subject needs to remain on task for an hour or more, head perfectly still, focused on some mundane activity such as looking alternately at houses and faces. No wonder that most studies rely on a pliant and well-disciplined population of college undergraduates for optimum results.
What those results are missing – in addition to information about more difficult-to-scan populations – is information about the particular brain functioning of each individual as well as the connectivity relations between brain areas. Various statistical methods have been developed in the last several years to try to capture these relations. Reference Mumford and Ramsey9,Reference Gayles and Molenaar10 As with the problems of researcher bias and the statistics of large numbers, the failure to model individual difference and functional connectivity is not an insurmountable problem. Methods can be developed to refine the interpretation of fMRI data in ways that yield these otherwise absent factors.
To be clear, my point is not that fMRI data is useless or false. My point is that we need to prevent interpretations of the data getting ahead of the development of proper methodology. Over the course of time, scientists will find the flaws in research procedures and correct for them. Given the global scientific community and the divergent motivations on the part of researchers, I have no doubt about the long-term self-corrective capacity of the scientific method. The worry in an age of 24-hour news and self-promotion is that insufficient time is given to the development of a methodology prior to the declaration of success.
This worry is amplified when billion-dollar information-gathering projects aim to totalize our understanding of the mind. Advocates for this sort of Big Science approach say that the problem with previous efforts is that they have been piecemeal, analyzing specific brain areas of interest or single neurons. What we need, they say, is to map the entire architecture of the brain at every level of description, from gene to synapse to connectome. While it is hard to object to spending more government resources on basic science, the likelihood that faulty conclusions will outpace the development of proper methodology is particularly serious given the enormous scale of these projects.
3. The Promise and Pitfalls of Information Integration Theory
Even if all the methodological concerns about the collection and interpretation of data could be resolved, a critical philosophical problem is raised by the supposition that data are sufficient to understand the mind. The previous section considered the problem of data getting ahead of theory. This section considers the problem with data itself, the inability of data to be theory.
A theory of consciousness proposed by neuroscientist Guilio Tononi, the Information Integration Theory (IIT), has gained a great deal of attention due to its cogent integration of neuroscientific data, information processing mathematics, and philosophical argument.Reference Tononi 11 , Reference Tononi 12 In a recent collaboration,Reference Tononi and Koch 13 Tononi and Koch begin by describing five phenomenological axioms:
(1) Intrinsic existence: The subjective perspective is a real feature of consciousness, not an external relation.
(2) Composition: Consciousness is structured of different elements, such as colors, sounds, thoughts, etc.
(3) Information: Each conscious state can be differentiated from other possible experiences in terms of its elements.
(4) Integration: The elements of conscious experience are unified.
(5) Exclusion: Every conscious experience is what it is, neither more nor less.
Whether these five axioms are necessary and sufficient criteria for consciousness is a matter of dispute. Even so, phenomenological description is necessary to identify the target for explanation. Because ‘consciousness’ means many things to many people, a careful description is important to connect the neuropsychological explanation to the phenomena at issue.
According to IIT, the capacity of a system to integrate information accounts for these five features of consciousness, and it does so by showing how to answer questions about quantity and quality:
Quantity: What level of consciousness does the system exhibit? How many areas are involved? What sort of activity is required?
Quality: What kind of consciousness does the system have? What determines the difference between experiences of red and green, spicy and sweet?
These are very good questions. One obstacle to explanation is a tendency to think of consciousness as an undifferentiated monolith. Phrases such as ‘what it’s like’Reference Nagel 14 to have an experience, or the intrinsic ‘raw feels’Reference Kripke and Munitz 15 of consciousness tend to short-circuit further reflection on the complex nature of consciousness. The recognition of levels and varieties begins to open up phenomenological examination of the structure of experience in useful ways.
The quantity of consciousness considers aspects such as level of arousal, from high level activity in extreme fear or exhilaration to low level exhaustion or contemplation. Measures can extend to unconscious states as well, characterizing differences between sleep stages (NREM, REM, etc.) as well as clinical delineations between coma, minimal vegetative state, and so forth.Reference Tononi 16 Many brain areas might be involved, as when a jazz saxophonist integrates the sounds of the instruments, a complex understanding of improvisation, and audience response with motor movement in the ongoing development of a performance. Or very few brain areas may be activated for a repetitive task such as typing.Reference Tononi 16
The quality of consciousness concerns the features that are often simply equated with consciousness: sensory qualities such as color, sound, taste, smell, and texture; emotional feelings and moods, thought processes from object identification and classification to deliberation and decision-making. The qualities of consciousness characterize what it’s like to be conscious, so it seems from the perspective of consciousness to be all there is to experience. Yet as the five axioms suggest, other features are necessary to consciousness as well, such as subjectivity, unity, and uniqueness of content. Moreover, each quality can occur unconsciously, as research over the last decade demonstrates.Reference Dehaene 17
So a theory that accounts for the quantity and quality of consciousness is a valuable contribution to understanding this puzzling phenomenon. The problematic element in the theory is the claim that consciousness can be identified by the calculation of integrated information.
Originally conceived by Claude Shannon as a way to quantify the messages transmitted by the telephone system, information is the reduction of uncertainty among possible states to the specification of a single state. So, for example, when only two alternatives are possible, 0 and 1, a switch set to 0 conveys 1 bit of information. When a system can be in 64 different positions, 6 bits of information can be conveyed by specifying its state. The unit of information, ‘bit’, is short for ‘binary digit’ because the amount of information is calculated in terms of the number of yes/no questions needed to determine the message content. 18 , 19
IIT draws on Shannon’s method of quantifying information in terms of the connections in a system, but the theory differs significantly in restricting its calculus to causally effective integration. Only the connections that interact to form what Tononi calls a complex determine the amount of information carried by a particular state. Consciousness, Tononi proposes, is constituted by the complex with the highest amount of integrated information, or Phi value.Reference Tononi, Boly, Massimini and Koch 20
By modeling neural interactions under various conditions, Tononi shows that high information values are produced when a system exhibits both functional specialization and functional integration. The human cortex, for example, is composed of areas devoted primarily to visual or auditory processing, yet these dedicated regions can be connected when seeing and hearing someone speak. Low Phi values are produced when regions function independently. Consequently, parallel processing can occur that remains unconscious because it is not integrated into the complex with the highest Phi value.Reference Tononi, Boly, Massimini and Koch 20
The promise of IIT is a mathematical calculation of the quantity and quality of consciousness:
Quantity: a high value means a high level of consciousness, as occurs in alert, multi-sensory integration. Conversely, during sleep and anesthesia, interconnections among cortical regions become weak or rhythmic, reducing the value of Phi.
Quality: the location in quality space specified by a state determines the nature of the experience as red or green, spicy or sweet.
Let me say a word about the idea of a quality space, because it is a remarkable conceptual achievement of the philosophy of mind and vision science. One of the most puzzling aspects of conscious experience is its qualitative character. A quality space opens experience to analysis by examining the internal relations of its features. For example, the pure redness of red is indescribable. We can say it is bold or vibrant or like the sound of a trumpet, but none of these descriptions is even remotely adequate. On the other hand, if red is compared with other colors, we can see that red is more like orange than it is like green.
Vision scientists examine the interrelation of stimuli, physiological mechanisms, and behavioral responses to develop theories about how colors are discriminated. This sort of analysis is what Tononi draws on when he describes the location in quality space measured by IIT. We start with the perceptual object (an apple, for example) and compare its color with other colors. This analysis yields the mechanism of opponent processing that physiologically produces the range of colors that the human visual system detects. The mathematical relation between the stimuli and the operation of the opponent process neurons forms the 3D color space. That particular state of the visual system therefore specifies which color the person sees.
There is insufficient space to fully explain the dimensions of a quality space, but the basic idea is that the sensory qualities of experience (colors, sounds, etc.) form internal relations that can be mapped. (For more detail, see the material listed in the Further Reading section.) These relations can then be mapped isomorphically onto neural relations to provide a topological explanation of the qualities of experience.Reference Kostic 21 In principle, every qualitative feature of consciousness can be similarly specified in terms of a neural quality space. According to IIT, just those features that are included in the main complex of integrated information are conscious. Other features that are activated but not integrated would remain unconscious.
Information is certainly pretty cool. What IIT does not seem to address, however, is the way integrated information is often processed unconsciously. While reading this sentence, for example, your brain is busy translating the colors and shapes on the page into letters and words. In addition to the visual image, all that occurs consciously is a sense of understanding, or perhaps an inner recitation of the text. Information about the individual letters and their semantic relations must be causally integrated in order to for you to read effectively, but if you were conscious of the letters themselves, you would have trouble reading. It’s a trees and forest problem.
I see three ways IIT might deal with unconscious processes like this. First, the semantic processes might not be integrated in the right sort of way to generate a Phi value.Reference Oizumi, Albantakis and Tononi 22 The technical details are too complicated to be worth spelling out, but I am skeptical that the vast array of unconscious processes would fall into this category. For one thing, many unconscious processes can be brought to consciousness by a shift of attention, which would not alter the causal structure of a system. Second, the semantic processes might be conscious after all, even though they do not seem conscious to you. Proponents could maintain that all information integrated into a complex must be conscious. But this response violates the phenomenological basis on which IIT rests. The claim is that the theory captures how conscious experience seems from the subjective perspective, yet the experience of reading does not seem to include the identification of letters. The third response suffers the same problem. You might be the host of multiple conscious complexes. While your reading experience is the major complex, from its perspective other minor processes ‘appear to be carried out unconsciously’,Reference Tononi, Boly, Massimini and Koch 20 yet are somehow conscious.
With this suggestion, IIT departs entirely from the way conscious experience seems from the subjective perspective. While pathologies such as split-brain cases and dissociative identity disorder may introduce doubt about the rule that there is one consciousness to a person, in the normal case the rule holds. The question that IIT fails to ask is: why? Why is information integrated into one unified consciousness per person? What function does conscious information integration perform that unconscious information does not perform?
4. The Limits of Information
The basic problem with fMRI data and integrated information is that they are meaningless, whereas consciousness is full of meaning. Sensory qualities such as taste and feel draw us toward some things and away from others. Colors enrich our visual perception to allow quick and subtle identification. Inner speech channels the stream of ideas along the path of thought. Think for a moment about the contents of your current conscious state: the text, your thoughts, the room you are in, bodily feelings of relaxation or stress. These contents make your consciousness meaningful; they constitute your view of the world. Because consciousness is your view of the world, it might be mistaken. Your stress might be unfounded, you may see the dress as black and blue when it is actually gold and white.Reference Corum 23 To understand the way consciousness presents the world to you requires understanding how and why it has acquired the contents you experience. In other words, an understanding of the meaning of consciousness requires an understanding of how conscious content represents the world.
While some claim that consciousness is more than its capacity to represent features of the world,Reference Block 24 I predict that a representational account will eventually overcome its skeptics. The better we understand how sensory and cognitive systems function to present different features of experience, the less motivation there will be to posit a phenomenal remainder.
IIT takes this position as well and attempts to capture conscious content through a specification of information relations among brain processes. However, unless we know how those brain processes are used to represent features of the world and why only some are conscious, the relations among brain processes are nothing more than physical, causal connections. Data alone are meaningless, and so data alone cannot explain the meaning conveyed by consciousness.
The reason that information relations are insufficient for meaning is not that meaning is some sort of magical sui generis feature of minds. The problem is that information is defined in relation to its source, where meaning must be defined in relation to its use.
To see why this is so, consider the attempt by Fred Dretske to extract meaning from information by delimiting the conditions under which a signal is guaranteed to signify its source.Reference Dretske 25 Dretske called these lawful correlations between signal and source natural signs. The sound of thunder, for example, carries information about the distance of the storm, because there is a lawful relation between the time and the distance the sound travels. By learning to exploit this regularity in nature, the sound comes to mean that a storm is coming.
The world is full of this sort of exploitable information, and natural selection has designed different animals to be attuned to different sorts of signals depending on the pressures in their environment. Flies increase defensive behaviors, such as jumping and running in response to an increase in threat stimuli.Reference Gibson, Gonzalez, Fernandez, Ramasamy, Tabachnik, Du, Felsen, Maire, Perona and Anderson 26 Archerfish track the velocity and direction of insect prey in order to aim a water spray timed to intersect its path.Reference Schuster, Wöhl, Griebsch and Klostermeier 27 Hummingbirds differentiate the replenish rate and amount of nectar according to the location and color of flowers.Reference Healy and Hurly 28 Animal researchers are continually discovering specialized mechanisms by which behavior can be tailored precisely to useful environmental information.
Dretske claims that because the signal is nomically correlated to its source, sensory information means the thing or state of affairs that caused it: threat stimuli mean danger to the fly; velocity and direction of visual stimuli mean prey to the archerfish; location and color mean nectar to the hummingbird. Information theory avoids skeptical problems about meaning – why does my perception of red mean something is red? – by tying perceptual content to its object by means of an objective lawful relation.
One problem with this interesting move is that it works too well. Call this the Problem of Error. Content is tied to the object so tightly that there is no possibility of a mistake. Natural signs always indicate their source, so they can never be wrong. Minds, as noted earlier, are often wrong. Flies respond to threat stimuli, such as shadows, even when no threat exists. Archerfish shoot at decoys with similar light and motion properties as prey. Hummingbirds continue to expect nectar from flowers even after the replenish rate has changed. If the correlation between signal and source were guaranteed, this sort of mistake could not happen. But of course it does happen all the time.
Another problem with information theory, call it the Specification Problem, is that it fails to specify which aspect of the source is represented. Take the information about the color of flowers that the hummingbird uses to represent nectar. The same visual signal that carries information about the color red, carries information about the specific light reflectance properties of hue, saturation, and brightness. It also carries information about the ambient environment necessary to generate and transmit that wavelength frequency. Information about the hummingbird’s retinal configuration and visual processing system is also nested in the same physical process by which the color is represented. In order to be successful in re-identifying the flower, the hummingbird must be able to extract just the color information and in such a way that the range of reds under various lighting conditions and visual angles still represent the same flower.
In other words, the wealth of information available in the world is a theoretical disadvantage as well as an advantage. The theory needs somehow to say how and why the signal manages to mean only ‘red’ when it carries all kinds of information in addition to ‘red’.
Dretske has tried numerous ways to solve both of these problems. His original response to the Specification Problem draws on the role ofchannel conditions in communication theory, and the distinction between analog and digital signal coding. In calculating the amount of information transmitted by a signal, the conditions necessary for transmission must be specified. In communication, this means calculating the signal-to-noise ratio to determine the statistical probability that a message will be successfully transmitted. If the signal-to-noise ratio is 0 then as much noise as signal is transmitted, rendering the probability of successful transmission at chance. The goal for communication, then, is to specify the channel conditions necessary for an optimum signal-to-noise ratio.
In a theory of meaning, the goal is not to increase the probability of successful transmission; it is to specify the content of the transmission. A signal is either meaningful or it is not. It makes no sense to say a signal means ‘a 0.5 probability of blue’. It either means ‘blue’ or not. (It could, of course, mean ‘a 0.5 probability of blue’, but then it would either mean that or not.) For Dretske, channel conditions rule out alternative possible sources so that a signal carries information about its source with the probability of 1. A signal acquires meaning, or informational content, only given this absolute guarantee that it carries information about its source.Reference Dretske 29
Further specification of content occurs through the process of digitalization. Perception carries information in analog form; every variation in the source is encoded in the signal. Like an analog clock that spatially represents every gradation of temporal movement, perception represents every gradation of redness. Cognition or conceptualization carries information in digital form by abstracting from the details of the analog information to represent a range. So a digital clock only shows the time in discrete minute-long units, and the concept ‘red’ represents all the gradations of red as the same color.
Through channel conditions and digitalization, Dretske solves the Specification Problem. The hummingbird representation means red, because that representation occurs with a probability of 1 when the source is red, and it abstracts from specific properties of hue, brightness, and saturation to represent a range of reds.
So far so good; however, the Problem of Error proves more difficult to solve. Because the signal is correlated by a probability of 1 with its source, there is no possibility for the signal to be wrong about its source. The very advantages of natural signs as a ground of knowledge – their objectivity, their natural necessity, their ubiquity – makes them insufficient as a ground of meaning.
Two of the ways Dretske has tried to resolve the Problem of Error are instructive. First, he has suggested that error can occur after a critical training period. A child is trained to recognize birds by putting her in an appropriate informational relationship to birds (this is a bird, this is a bird, this is not a bird, etc.). Once she has acquired this concept, it is possible for the concept to be triggered by something that is not a bird (airplane). This is a case of misrepresentation.Reference Dretske 29
Among other difficulties, this answer is hopelessly vague about when the ‘critical training period’ ends. Are we to take the child’s misidentification of the airplane as a false belief or as an indication that she has not yet got the right concept? More troubling, why not say she has successfully acquired a concept that means ‘bird or airplane’? The possibility of a disjunctive concept would undermine much of the value in the requirement of correlation between signal and source with a probability of 1 as it would open the way for all kinds of hodgepodge collections to count as concepts.
Dretske’s second solution is to take the teleosemantic turn: a representation means something when its function is to be correlated with that thing. So the child’s concept means ‘bird’ because it has the function of being correlated with birds. When things go well, only birds trigger the bird concept; the concept means ‘bird’, because it is causally correlated with birds. (By this point the language of probability and information has dropped out in favor of terms such as ‘biologically necessary’ and ‘natural sign’.) Misrepresentation occurs when one of the intermediary stimuli that usually indicates birds occurs in the absence of a bird.Reference Dretske and Bogdan 30 Although it has the function of representing birds, the ‘bird’ concept fails to fulfill this function.
5. From Information to Meaning
The move to function is a good one, but Dreske is still too enamored of information to realize that successful representation is determined by what is done with it rather than by what caused it. If functional definition is oriented toward its cause, then any means of being caused prevents error. Rather than say the bird concept has failed, we could say that it successfully represents the range of intermediary stimuli that cause it. If the function of a representation is simply to carry information about its cause, then there is always some way to specify the cause such that the representation is carrying it.
Instead, the functional definition of a representation must be specified by the way it is used. A ‘bird’ concept represents birds, because it is used to do things with birds: to track them visually, to think about them, to capture and eat them. A ‘bird’ concept is successful when these activities are successful. In Ruth Millikan’s formulation,Reference Millikan 31 which I favor, a representation means something when it varies in accordance with that thing, because that co-variation relation has helped serve useful functions. My concept means ‘bird’, because it has helped me to do things with birds.
Others have also offered theories of meaning that involve use to address issues of misrepresentation. Donald DavidsonReference Davidson 32 solves the problem of specification by means of the linguistic community that interprets beliefs and actions as rational. You can determine that my concept means ‘bird’ by the way I use the word to make true statements. A worry about depending on a linguistic community to specify meaning is that this move excludes non-human animals, pre-linguistic infants, and all contents of consciousness that are not regulated by language. Since there is good reason to think the mind is capable of non-linguistic representation, a biological rather than social ground of meaning is preferable.
Jerry Fodor, who has also expressed concern about information as a source of meaning,Reference Fodor 33 argues that errors asymmetrically depend on a lawful relation between a veridical representation and its cause. The airplane would not have been mistaken for a bird, unless birds reliably cause bird representations. Therefore, using ‘bird’ to represent an airplane is an error. As discussed above, the trouble with reliable causation is that meaning is defined in terms of its source rather than its use. Consequently, the theory cannot account for the ways representations can be reliably caused yet mistaken, as when Venus reliably causes both my representation of the last star in the morning and the first star in the evening, yet I fail to know they are the same. Conversely, representations can be meaningful yet not reliably caused, as when I mistake twins for the same person. Fodor says these cases can’t happen often or action would be irrational (Ref. Reference Fodor34, pp. 27–54), yet they seem to happen all the time.
On Millikan’s account, the actual history of successful and unsuccessful use of a representation articulates both the basis for meaning as well as change in meaning. When something other than birds triggers my bird concept, this counts as a misrepresentation, because only birds have accounted for the success of my concept in the past. This new thing – the airplane – will likely fail to sustain my needs with regard to birds. I won’t be able to capture and eat it, for example. If, on the other hand, a new thing does sustain my needs – perhaps I move near the sea and learn to catch flying fish regularly – then my concept is equivocal; it includes both birds and flying fish. This sort of equivocation is problematic for an informational theory, because information defines success in terms of causal source. Equivocation is not a problem for Millikan’s consumer theory, because error is still possible. Some third thing, that useless airplane, may trigger my equivocal concept and yet fail to satisfy my needs. So airplanes cannot count as ‘birds’. Or, more importantly, future needs may require a disambiguation between birds and flying fish. Perhaps I relocate inland and no longer catch flying fish. This change in environment would motivate a distinction between these two ‘kinds’ of bird.
Although information is insufficient for meaning, it does play a role: objective, physical correlations among events are the reason for representational success. If the world were not full of information, animals would not be able exploit the stable environmental relations necessary for survival. But information alone cannot specify meaning, because it fails to include the evolutionary, developmental and dynamic dimensions of representation. If content is specified by use, then species history, individual history and current situation all figure in determining what a representation means.
And here at last is the moral of the story: no amount of information can tell us anything about the mind unless we know how the system uses that information. In fMRI research, assumptions about use are smuggled into the research methodology by designing tasks that depend on established skills: identify houses and faces, push the button when you hear the tone, say which items on the list you recognize. On the assumption we know what it would mean to accomplish these tasks, we can look to see what the brain is doing and draw conclusions about the meaning of various neural activations. Because these assumptions are generally warranted, conclusions based on them have proven beneficial. We understand some of what has gone wrong when particular brain areas are damaged and when neurotransmitters stimulate or inhibit activation. This understanding is based on a common evolutionary history and a roughly similar developmental history among human beings. To the extent that we can determine the evolutionary and developmental pressures on non-human animals, we can begin to understand how their minds function as well.
With consciousness things are more difficult because there is no simple correlation between a world state and a conscious state. We know roughly what it means to identify a house; we are much less clear about what it means to be conscious of a house. Information Integration Theory posits that the quantity and quality of information integrated determines the level of consciousness, the degree of Phi. But there is still the question of why integrated information is conscious.
Here again, I believe function can provide an answer. In my view, consciousness serves the biological function of representing the current moment – what is happening now. Notice again the contents of your current conscious state. The various aspects of your self and world that appear to you now constitute your conscious experience. Each element of your conscious experience – for example, the color of the text you are reading – has the function representing something in the world that you use. The way things appear to you, as colored, textured, shaped, etc., is a matter of how you use those features in the world.
Likewise, your conscious experience as a whole is constituted by a conjunction of elements that itself has a function, the function of representing the present moment. Not all of your current sensations and thoughts are incorporated into this representation, only those that provide the best approximation of how the world is now, relative to your goals and actions. The function of representing the current moment is to keep track of how things are going in the world. You need to have conscious representations in order to respond to novel conditions, either a difficulty in accomplishing a task or an opportunity that might be worth taking.Reference Droege 35 , Reference Droege 36 By articulating representational function, what the mind does, a theory can identify the physical structures that accomplish that function. Theory moves beyond correlation to explanation.
In the end, Information Integration Theory might provide an understanding of how brain processes produce consciousness, even though it cannot explain why those processes are conscious. A philosophical theory that can account for the function of consciousness is a necessary supplement to any data-based, informational analysis of the brain.
My title ‘Not by Data Alone’ is an oblique reference to the Bible passage from Matthew: 37 People do not live by bread alone, but by every word that comes from the mouth of God.
One point of this passage is that physical sustenance is insufficient for a meaningful life. In Christianity, meaning comes from God. In evolution, it comes from representational function. These explanations need not be opposed, but that is an issue to consider another day. What I hope to have shown here is that physical information is insufficient for meaning, and meaning constitutes consciousness. A mind cannot live by data alone.
Acknowledgements
This article was written for the symposium, Reconsidering Humanity, organized by Urban Strandberg and Christian Munthe. I am grateful for the generous support of the University of Gothenburg and the Swedish research foundation Riksbankens Jubileumsfond, which made my participation in the symposium and the following workshop possible. A draft of this article benefited considerably from comments by all the participants at the workshop, especially Palle Dahlstedt, Barry Smith and Christian Munthe. Great thanks are due as well to Daniel and Alice Martin for their invaluable contribution to our discussions.
About the Author
Paula Droege is a Teaching Professor in the Philosophy Department at Pennsylvania State University. Her research on philosophical theories of consciousness proposes an essential role for temporal representation in conscious states. She is the author of Caging the Beast: A Theory of Sensory Consciousness, and articles on the role of consciousness in memory, free will and delusions, as well as animal consciousness.
