Consciousness, Complexity, and Machines

The phenomenon of consciousness has recently been the focus of much debate and research, especially among experts in psychology and artificial intelligence. Researchers and philosophers alike are trying to solve the so-called mind-body problem by uniting the processes of the brain with the cognitive experiences of conscious awareness. This paper attempts to present several leading theories of consciousness, as well as the empirical evidence provided to support them. This first necessitates a rather laborious definition of consciousness and the phenomenal experiences typically associated with it. Finally, the implications of these conclusions to machine consciousness are examined.

Kihlstrom (1993) begins his paper by claiming that “consciousness is the central issue for psychology and for the broader interdisciplinary effort known as cognitive science” (p. 334). From this position, it should seem strange that consciousness has been defined in many varied and often incompatible ways. However, this is the case. Even once a coherent definition of consciousness has been established for purposes of discussion, it often reveals itself to be either incomplete or in disagreement with evidence gathered by introspection. Before the question of consciousness in machines can even be approached it is important to establish exactly what the question is. The latter will compose the bulk of this paper, with the former concluding it.

The mind-body problem can be summarized as follows: “How could the aggregation of millions of individually insentient neurons generate subjective awareness” (McGinn, 1991, 1)? In other words, how is it that the phenomenal experience that we call consciousness arises from the irritation of nerve tissue? This question of the relationship between the world we experience in contrast with the world as it actually exists has been discussed throughout the history of western philosophy, beginning as early as with Plato, who compared our perception to seeing the shadows of reality on a cave wall. The problem is an important one to the study of consciousness, a view that Farthing (1992) elaborates:

Some of the earliest attempts at “answering” this seeming disparity between the body and the mind came from the church. Gray (1999) writes that the church claimed the human being “consists of two distinct but intimately conjoined entities – a material body and an immaterial soul – a view referred to today as dualism” (p. 4). Descartes restricted this view to the idea that humans were unique in their ability for thought, and that this ability, the soul, was seated in the pineal gland (p. 5). Unfortunately, this view retains the underlying conviction that the mind is somehow of a different nature than the body, a stance that refutes the possibility of studying the most interesting features of consciousness by the scientific method. The modern, opposite view, is that of naturalism or materialism. “Materialism (materialist monism) is the belief that there is only one type of substance…mind and consciousness are functions of complexly organized matter” (Farthing, 1992, 65). This is the opinion upon which most scientists agree axiomatically, and the one that this paper will accept as true. Searle (1993) wrote that, “above all, consciousness is a biological phenomenon” (p. 310). It is the nature of consciousness and of the complexity from which it arises that is under such heated dispute; this is the topic of this paper. As Kihlstrom (1993) said, “my neuroscientific colleagues inform me that the mind-body problem has been solved, although there seems to be residual dispute over just what the solution actually is” (p. 334).

If consciousness must be a fundamentally physical process, we can begin to make attempts at defining it. This is a difficult proposition. Farthing (1992) makes the rather disparaging claim that “it is not possible to specify precise objective criteria for identifying consciousness” and that it cannot even be given a functional definition since its function is in doubt (p. 5). To say it cannot be defined makes studying it as a unified phenomenon nearly impossible. This seems to be avoiding the issue, and so further attempts must be made to at least narrow the field of what consciousness and phenomenal experience are and are not. This will lead to a working definition of consciousness that can be considered when discussing machine consciousness.

“The phenomenon of consciousness…needs to be distinguished from certain other phenomenon such as attention, knowledge, and self-consciousness” (Searle, 1993, 310). Knowledge, or what Pinker (1997) calls “access to information” (p. 135), refers to that data to which we are able to refer in order to make “verbal reports, rational thought[s], and deliberate decision[s]” (p.135). This definition also excludes “extremely complicated forms of self-referential consciousness that would normally be described as self-consciousness” (Searle, 1993, 311). Consciousness is sometimes used to describe what is at the center of attention. That is, while many things may be in the periphery of consciousness, only one thing (whether an object being observed or a thought) is at the center of attention. Baars (1997) draws a clear distinction between consciousness and attention, claiming that attention is the “access control mechanism that determines what will or will not become conscious” (p. 364). In this sense, Baars pools attention with information access, which we have already accepted as being different from consciousness. He goes on to exclude working memory from a definition of consciousness, and “argues that [working memory] may be a superstructure dependent on the fundamental features of consciousness” (p. 369). Now that we have established what consciousness is not, we will examine what it might be.

Phenomenal experience, or phenomenal awareness, is the set of “percepts, memories, thoughts, beliefs, feelings, emotions, goals, and motives that we consciously experience…” (Kihlstrom, 1993, 334). The italics are mine. Neither the memories themselves, nor the percepts, embody consciousness. It is the sensation of subjective experience that constitutes consciousness. It is the difference between knowing that the ball reflects light which correspond to the color purple in the visible spectrum and actually experiencing the ball’s “purpleness.” The experienced qualities are often referred to as “qualia.” These qualia have a distinct first-person aspect to them; that is, we experience them with the implicit understanding that we are the agent or subject of them. Should it be said that the presence of these experiences, though not observable by a third party, indicates consciousness in a system? Perhaps not, as there are examples of otherwise “conscious” people having access to the information associated with seeing an object without actually having the experience of seeing it. Patients with “damage to the striate cortex of the occipital lobe…report a lack of visual experience in some portion of the visual field” (Kihlstrom, 1993, 337). This condition is known as blindsight. They do not experience the purple ball as we are so convinced that we do, “yet when forced to make guesses about properties of the object…[they were] more accurate than would be expected by chance alone” (p. 337). This implies that the conscious experience is separate from perception and even information access. Kihlstrom continues his argument:

Therefore, consciousness is not the information embodied by phenomenal awareness, as that can occur without phenomenal experience. But phenomenal awareness itself is left unexplained.

When discussing consciousness, it is necessary to divide discussion into two parts: the easy problems and the so-called hard problems. The hard problems consist of potentially unanswerable questions: “What makes us actually suffer the pain of a toothache or see the blue of the sky as blue” (Pinker, 1997, 131)? This aspect of consciousness is what Pinker calls “sentience…[something for which] the computational theory of the mind offers no insight; [nor] does any finding in neuroscience…” (p. 146-147). This is considered a hard problem because there may simply be no way to find an answer, or it may be a long time before we have even the ability to search for one. Introspection seems like the only way to even be convinced that such a phenomenon as sentience exists, and the problems with studying consciousness through the window of consciousness are numerous. The easy problems, on the other hand, are quantitative ones. They are ones which can be studied empirically and whose properties can be determined from behavior and observation. They are the questions with which science is concerned and which will be incorporated into a definition of consciousness. The hard problems will be ignored for the purposes of this paper for two reasons. First, it is no use creating a definition if we will have no way of determining whether anything fits that definition. Second, there is no empirical evidence dealing with this ethereal sentience. Instead, a definition will deal with the potential functions of, and quantitative properties of, consciousness.

Consciousness has been separated experimentally from perception, memory, and informational access (Baars, 1997). What remains is presented in Baars’s definition of “consciousness [as] the publicity organ of the brain, one that is used to access all its functions” (p. 370). But this is not the only view. Other theories do not segregate consciousness into its own organ. Bogen (1995) claims “that conscious awareness…is engendered by neuronal activity in and immediately around the intralaminar nuclei (ILN) of each thalamus” (p. 52). Amazingly, Bogen claims that in the presence of a pair of conscious mechanisms on the same brain, “there can be doubling of [consciousness]…this has been inferred from split-brain cats and monkeys” (p. 54). Sun (1999) lists several leading ideas: “consciousness as a system separate from that of the mind which works deliberately and serially, consciousness as a supervisory system, or consciousness as a dominant process in a pool of processes running in parallel…” (p. 529). Still other theories, such as connectionism from neural networks, describe consciousness as an “emergent property” of complexity (p. 529).

Defining consciousness as a functional, emergent property of complexity is an interesting proposition.

Competing theories identify significant organizational units ranging “from the quantum level to the level of the whole brain” (p. 181). Searle (1993) presents an alternative to the common neuron-centric theories of organization, saying that, “it might turn out that we have overestimated the importance of the neuron and the synapse. Perhaps the functional unit is a column or a whole array of neurons…” (p. 312). Neural network models present the neuron as being less individually important, and instead consider “features [to be] realized…[in the] complex interaction of huge numbers of cells” (p. 181).

If something has the necessary level of organization to be conscious, but does not have as many of those units, is it possible to claim that it is only partially conscious? In other words, when considering machine consciousness, is it possible for a machine to be semi-conscious? Some consider consciousness to be an all-or-nothing proposition, such as “a switch’s being on or off…this view of a unitary, indivisible consciousness is mistaken…” (Davies & Humphreys, 1993 ,224). For the purposes of this paper, we will imagine that consciousness need not occur as a result of a single configuration unique to the brain and can occur elsewhere and at other degrees of complexity. Assuming otherwise would negate our intent.

Neurons are often modeled like black boxes. They take inputs from surrounding neurons, and, when they reach a certain threshold, they fire, sending a signal to more surrounding neurons. Neurons can excite or inhibit, and can have different weighting values when they send information to other neurons. Neural networks are arrangements of similarly modeled neurons that are capable of performing computations and logical arguments. Of course, actual neurons are rather poorly understood, and are not remotely as simple as this model implies (Pinker, 1997, 99-111).

While there are many other subtleties to consciousness, we have covered those characteristics most relevant to our discussion about machines. So now consciousness needs a definition. From the evidence described above, it seems reasonable to characterize consciousness as the set of phenomenal, subjective experiences through which a system perceives and interacts with the world. Whether this is possible for a machine will be the subject of the remainder of this paper.

One difficulty with consciousness is that we may create a conscious machine and never know it. The first, and most famous test of whether a machine might be considered conscious was proposed by Alan Turing and is called, appropriately, the Turing test. Essentially, a human subject is placed in front of two computer terminals and is able to type text into each. One terminal is connected to another terminal where a person is typing the other side of the conversation. The other terminal has its content read and analyzed, and a machine generates a response to be sent to the human subject. If the subject is unable to determine through interrogation which terminal is connected to the computer and which to the human, then the machine would pass the consciousness test. Instead of criticizing this method directly, I will leave it to Searle. The two sides of the machine consciousness argument will now be presented.

“The Turing test,” Searle (1993) writes, “disposes us to make…the mistake of behaviorism and the mistake of computationalism” (p. 317). Searle opposes the view that consciousness can be analyzed through behavior. Searle says that a machine could behave as though it was conscious without actually being so. “There is no logical connection…between inner, subjective, qualitative mental states and external, publicly observable behavior” (p. 317). It is Searle’s next objection with which we are most concerned. That is, he claims “computational models of consciousness are not sufficient by themselves for consciousness” (p. 317). His reason goes as follows: a model for consciousness is fundamentally no different than a computational model for anything else. If we do not suppose a computational simulation of rain will make us wet, why believe that the “computational model of consciousness is somehow conscious” (p. 318)? He claims that there is a fundamental difference between the syntactic symbol manipulation of computation and the semantic content so apparent in phenomenal experience. He refers to his famous “Chinese room argument,” in which a “person goes through the computational steps of answering a question in Chinese but does not thereby acquire an understanding of Chinese” (p. 318). It was argued that the room as a system might be conscious, but even if the computational procedures were memorized and the experiment performed outside, Searle claims the same conclusion. But what if how we experience consciousness is the same as how people outside the room experience the thought experiment? In other words, what if we only understand Chinese in the same capacity that the people outside the room understand it? Couldn’t introspective analysis of consciousness be hiding many layers of preprocessing and giving us the illusory impression that we understand Chinese? In fact, this we know already to be the case for many mental phenomena (Kihlstrom, 1993, 334). Regardless, the heart of Searle’s position is that there is a difference between intrinsic properties of nature and observer-relative properties or symbols:

He admits that a computational interpretation can be assigned to anything, but that consciousness is not intrinsically computational. “Computation exists only relative to some agent or observer who imposes a computational interpretation on some phenomenon” (p. 318). So, while Searle adamantly asserts that consciousness is a biological phenomenon, it is not one born of computation. So, if a machine is meant to possess consciousness, he qualifies that it had better do more than simply crunch numbers.

Sun (1999), on the contrary, assumes “the sufficiency and necessity of mechanistic explanations. By mechanistic explanations, I mean any concrete physical processes, that is, computational processes…” (p. 530). He thereby works off the hypothesis that every phenomenological distinction is, at base, a computational distinction. The purpose of his paper is to propose what a computational model of consciousness should look like. “I proposed the model CLARION…which embodies the ‘representational difference’ view of consciousness” (p. 535). That is, his representation contains two layers that correspond roughly to what might be called the conscious and the unconscious. By comparing the behavior of his model with human behavior, he was able to garner information about the function of human consciousness. It does not seem clear whether Sun intended for his system to actually be conscious is unclear, though I suspect not. His significance was in supporting with experiment the conviction that computation was sufficient, and necessary within his definition of it, to explain consciousness.

Cook (1999) found what I believe to be both a plausible and practical medium in his paper on simulating consciousness. He designed a “two-hemisphere system with a division of labor for nuclear and fringe information-processing,” hypothesizing that it should outperform a “one-hemisphere system or a two-identical-hemisphere system” (Cook, 1999, 68). He recognized that his system is not conscious, but that it “contains important qualitative elements of human consciousness” and might be considered conscious in a “quantitatively ‘deficient’” (semi-conscious) sense (p. 91). Truly noteworthy are Cook’s opening words, which embody my personal conclusions about machine consciousness:

“The greatest difficulty in simulating consciousness is defining it” (Cook, 1999, 62). However, once we dealt with “the dilemma of having to deal with both objective brain states and subjective psychological states” (p. 90), a working definition of consciousness was formulated. This definition was viewed in terms of machine consciousness, focusing primarily on computation. Arguments both for and against the plausibility of creating consciousness in a computationally-based machine were examined. Finally, the opposing views were reconciled with a more temperate and practical stance that, while a machine may never have the phenomenal consciousness that we experience, they may certainly exhibit all the quantitative and behavioral qualities associated with human consciousness. Thus, functionally conscious machines can be utilitarian in helping us to better understand one of man’s greatest mysteries: sentience.

Baars, B. J. (1997). Some essential differences between consciousness and attention, perception, and working memory. Consciousness and Cognition, 6, 363-371.

Bogen, J. E. (1995). On the neurophysiology of consciousness: I. An overview. Consciousness and Cognition, 6, 52-62.

Cook, N. D. (1999). Simulating consciousness in a bilateral neural network: “Nuclear” and “fringe” awareness. Consciousness and Cognition, 8, 62-93.

Davies, M., & Humphreys, G. W. (1993). Consciousness. Cambridge, Massachusetts: Blackwell Publishers.

Kihlstrom, J. F. (1993). The continuum of consciousness. Consciousness and Cognition, 2, 334-354.

McGinn, C. (1991). The problem of consciousness. Cambridge, Massachusetts: Basil Blackwell.

Ornstein, R. E. (1977). The psychology of consciousness. New York: Harcourt Brace Janovich, Inc.

Pinker, S. (1997). How the mind works. New York: W. W. Norton & Company.

Revonsuo, A. (1999). Binding and the phenomenal unity of consciousness. Consciousness and Cognition, 8, 173-185.

Searle, J. R. (1993). The problem of consciousness. Consciousness and Cognition, 2, 310-319.

Sun, R. (1999). Accounting for the computational basis of consciousness: A connectionist approach. Consciousness and Cognition, 8, 529-565.