“You insist that there is something a machine cannot do. If you will tell me precisely what it is that a machine cannot do, then I can always make a machine which will do just that!” – J. von Neumann
The journey of philosophy and cognitive psychology towards an understanding of thinking closely follows, for many important reasons, the quest for artificial intelligence. For much of cognitive psychology’s infancy, thinking, in humans, was viewed as closely analogous to the way in which a computer comes to produce output for a given symbolic input. That is to say, the behavior of the organism is produced by a calculation of sorts by the brain, a calculation inevitably performed after a mathematically describable function. Related to this is Church’s nonformal thesis that every computable function is recursively computable. Turing expanded upon Church’s thesis, by proposing that a machine can compute any mathematical function, as long as there is an effective procedure to obtain the rules of that function. It seemed to follow that the output of the human mind must be computable by some machine that could perform a similar enough function to the one operating in the mind, and that such a machine was realizable. The system expected to achieve this was the Universal Turing Machine, a system that takes symbols as input, computes a function, and yields an output. Inevitably, it was thought, the UTM would receive not only character strings and integers as inputs, but visual and auditory signals as well. Thus, research began to create computers that closely mimicked the responses of humans to given inputs; if we could achieve this, it was thought, we would know the function used by the brain, and we would therefore know exactly how thinking is performed. A behavioral test known as the Turing Test was created as the benchmark for such a function, the passing of which entailed getting humans to believe that the computer was, too, a human. Soon the line between output that was clearly produced by man, and that which was produced by machine would become blurred, at which point it surely would follow that some basic function of the human mind had been computed, and that A.I. had been realized. Whether responding to conversational input, or solving mathematical puzzles with proofs that exceeded Bertrand Russell’s own, these ostensibly cognitive activities performed by machines seemed to prove that Artificial Intelligence was gaining ground very quickly in achieving its goal. It is no wonder, then, as the field of Cognitive Psychology was coming into its own, that psychologists drew bidirectional analogies between A.I. and the human brain. The primary of these was the information-processing model of the human mind, possibly derived from the method of computation by serial computers at the time. However, the limits of the linear model of input to computation and then to output were soon exposed, producing astonishing dilemmas for those involved in the related fields. It became apparent that the Universal Turing Machine was unfit for many complex tasks that the brain seemed to perform with relative ease. For instance, visual systems in biological nervous systems induced behavioral output in milliseconds, while advanced computational algorithms in digital circuits, a million times faster than the biological systems, failed even to compare in performance (Churchland, Could a Machine Think? ). There seemed to be something about the nature of the brain, then, that separated A.I. from its goal. Two prominent philosophers proposed this idea, both around the same time, each stressing two different, yet important, perspectives. The first was Hubert Dreyfus, who argued that Artificial Intelligence systems lacked a grasp of the background context with which information should be processed. In this sense, the information-processing model, a foundational concept in Cognitive Psychology, was indeed lacking in its explanation of how information is integrated in the brain while thinking; surely processing was not achieved in the serial manner envisioned and implemented by A.I. researchers and engineers. Accessing all of the relevant data in such a way, it became clear, would take too long, even with the most efficient serial processing algorithms; a new approach was needed. The second well-publicized attack on A.I. came from John Searle, who argued that there must be certain fundamental aspects of biological brains, deeply embedded in its inherent formation, that were not yet understood by Psychologists, Neuroscientists, and A.I. theorists. One such aspect of biological brains, he argued, was the phenomenon of consciousness, and he created an elaborate thought experiment to illustrate what he thought was the absurdity of the A.I. research program’s attempt to create intelligence from the digital computer paradigm. In this experiment, Searle asks the reader to imagine a symbol-manipulating system (the nature of a Turing Machine) that could translate Chinese to English. He then asks the reader whether this system in fact “knows” English. In Searle’s view, it seems preposterous to answer in the affirmative, and therefore argues that symbol-manipulation alone does not provide a platform for conscious intelligence. The two arguments from these philosophers are similar in one important respect: they both call for a non-behavioral evaluation of intelligence; merely passing the Turing Test will not suffice for either of these perspectives of what intelligence entails. It is interesting to note that this occurred at a time when the Behaviorist approach was being replaced more and more by the Cognitive Psychological approach in the general field Psychology, and thus Cognitive Psychology was forced more and more into the deep-end, so to speak, of explaining human cognition. Psychology had to go beyond mere behavior, and had to do so independently of analogies from Computer Science. Neurobiology must have seemed like the best bet as to where to look.
It seems that the primary mistake of the early A.I. research program, vindicating both Searle and Dreyfus, was the crude distinction between hardware and software. Given the UTM, it seemed only reasonable that one could compute the function of the mind on a software basis only; since it was the function that mattered, the engineering details were only relevant to a point. However, the human brain is not a UTM, and does not operate in a serial manner. Instead, the mechanism of the brain is inherently parallel; millions of neurons fire together to form a sum of input from many different areas, rather than a single stream from one point to another like in the UTM (Churchland, Could a Machine Think?). While the UTM works via binary signaling in an ordered sequence (if a and not b, then c), the brain contains networks of neurons, each network containing millions of inputs, with each input expressing a different weight on the system (depending for example, as one might imagine, on the input’s dendritic diameter or it’s degree of myelination), which altogether sum up to a probability for that node of the network, say a nerve, to send a signal. If the sum of the weights goes beyond a certain threshold, the system will fire and create an effect on the environment, and the new state will be fed back into the system through its sensory mechanisms. The network structure of this system became known as a neural network, and computational neuroscientists were able to model various elements of the nervous system based on this framework (Spivey, Continuity of Mind ). Furthermore, the system is also clearly dynamical, with bidirectional pathways to and from the target populations of processing units – for instance, populations of neurons in the human nervous system. This allows for a modulating ability between sensing and processing the environment. Thinking, therefore, cannot be a serial manipulation of symbols, just as Searle argued that it is not, but rather a far more complex and continuous, dynamical and parallel interaction of millions of efferent and afferent processing units. This does not mean that brains are not functional, that they aren’t “computers” in some sense; brains are computers, but in a radically different manner than was at first thought (Churchland, Could a Machine Think?).
With regards to thinking, an intriguing idea is the continuous manner, or “flow” of thinking. In natural language, we may say something such as, “think about your favorite food,” or we may tell someone to “think before you speak.” Although it may not always be useful to attempt to understand reality through the analysis of language, these phrases do give us insight into a general, “folk” understanding of what thinking is. If we use the understanding that inputs, fed into a system of neural networks such as a brain, cause a massively parallel and dynamical pattern of events, it is clear that the process will be continuous. This is due to the previously mentioned claim that as the nervous system processes inputs through the various sensory channels, the state of the system begins to change, and that this change is then fed back into the loop as new input. We might call this recursive process the foundation of the idea of cognition, as the brain continuously transitions to new states over time. As in the earlier examples from natural language, what happens when this process is applied with the conscious aim of moving the brain towards a state that “represents,” in some sense, an object in the world? It was an assumption of early folk psychology (a shared conceptual framework used to explain the behavior of ourselves and others), and even early Psychology, that beliefs (true or false statements about some subject) and desires were stored as symbolic representations in the brain. Paul Churchland argued that these ideas, having fermented for decades at the very basis of early Cognitive Psychology, were radically false and would soon be replaced by something more accurate. The answers were expected to come out of the growing field of Computational Neuroscience. The position that Churchland supports is known as eliminative materialism. It is materialist because it rejects dualism, thus requiring the perspective in question either to reduce to a theory coherent with the rest of neurobiology, or to be eliminated entirely by some new theory. It is an eliminative position precisely because it argues that such linguaformal representations, as described by folk psychology, do not exist in the brain at all (Churchland, Folk Psychology). Therefore, the eliminative materialist position requires a theory to explain basic mental processes, without using the idea that symbolic representations are accessed and processed by logical rules in the brain.
Meaning and Context
“We are all living in the same environment, but each of us lives in a different world.” – Schopenhauer
The question soon arises, if symbolic representations are eliminated by a theory of continuously transitioning brain states, in what way is there meaning in the brain? In the old theory from folk psychology, cognitive theorists could rely on the idea of propositional attitudes to understand how beliefs created meaning by reference to objects in the world (Churchland, Folk Psychology). In the new view, however, with no discrete representations or attitudes in the brain, it may in fact be easier to develop a theory of meaning. Armed with the new perspective of the brain that entails parallel processing, Computational Neuroscientists have envisioned a system that can take multiple forms of input at once, allowing for a sense of context that may not have been achievable by the representational theory of the brain. Moreover, the continuous and recurrent nature of the new theory further allows for integration, as these populations continuously communicate with different networks in the brain, which then feed into the system as new input. By discarding the hardware/software distinction, the root of early A.I., the new view opens Cognitive Psychology to a better understanding of learning; an understanding in which the very structure of the brain changes over time. In Computational Neuroscience and A.I. theory, this is known as “training up the system,” in which connections are made between relevant populations of processing units, and the individual weights of those connections are strengthened or weakened, based on the outcomes of the training sets (Churchland, Could a Machine Think? ). In this sense, the resulting image is not of the brain as a storage space for, or processer of, information. Rather, we have in mind a flexible and rather robust, not to mention sensitive and dynamical, incredibly intricate system; one that never produces an output without incurring some form of learning. However, how does all of this add up to meaning? If I encounter some complex idea, for instance, if I read in my textbook the metaphor “love is a collaborative work of art,” how might I begin to understand this? In this example, each part of the metaphor is connected to various other networks in the system; some of them entailing metaphors themselves, but all of them linked somehow to more and more networks, each by different strengths. It is this association among networks that creates context and meaning, and it is easy to see how such a system will also create meaning that is very specific to the individual system. For instance, if the population of processing units, in this case neurons, corresponding to the idea of art has been trained to connect with networks pertaining to ideas like illusion, something to display, inherently trivial, then the metaphor will produce a profoundly different meaning to someone for whom art is connected with the networks that map to inherently valuable, meaningful expression of the self, gift to the world, unique perspective of reality. Meaning is thus based on the structure of the brain itself, rather than on a function computed by logical systems. Some A.I. researchers believe that the way to create machines that have meaning is simply to imitate such neural networks, using algorithms that rely on Bayesian probability to calculate the strength at which nodes should signal to each other in the network; presumably this can be simulated on a Universal Turing Machine.
“When people think about spring, surely they are not confused as to whether they are thinking about a season or something that goes boing – and if one word can correspond to two thoughts, thoughts can’t be words” – Steven Pinker
The previous illustration of the new view of meaning from Computational Neuroscience described the way in which a network as a whole can be structured so as to create a sense of context. How does do the populations of processing units themselves “stand for” the concepts associated in such a network, however, without being symbolic representations, as in the folk psychological view? Some Neurophilosophers, including Spivey and Churchland, have argued that the units in such a network become activated in specific patterns of distribution, and therefore create a vector particular to that activity: an activation pattern (Churchland, Neurosemantics). Although the vector is in no way itself a visual mental representation of the object in question, it provides, in a neurologically operational sense, a mathematical representation of this object. Spivey argues, however, that these representations are not completely accurate. Rather, there is again a Bayesian probability that a particular vector refers to an external object, a probability that is conditional upon other events in, and attributes of, the network (Spivey, The Continuity of Mind). Given that certain populations of neurons are firing, it becomes more or less probable that the vector refers to a particular object. Given this, one could imagine experiments where researchers observe the human brain in an fMRI machine, and predict which populations of activated neurons create the vector that “refers to”, for some person, some external object. In this sense, it would seem that objects are mapped in the brain in a meaningful way, and therefore it is difficult to discount representationism completely; it seems rather that the theory has been thoroughly revised, rather than eliminated (Churchland, Neurosemantics). This revision has allowed Computational Neuroscientists to overcome, perhaps a large part of, Dreyfus’ argument about such a system not integrating background information. Although the chess champion Deep Blue does not know whether or not he has won a game, equipped with such a system of parallel neural networks, perhaps he could simultaneously integrate enough concepts such as to create the context of “winning”.
Of course there are still far larger concepts to grapple with during this discussion. For instance, how do vectors, activation patterns in the brain, create qualia? One analogy that makes sense to me is that of sound waves to sound. Sound waves need some kind of interpretation by a nervous system in order to create the experience of sound, and perhaps, too, the vector codes correlating to green need an interpretive function to create the conscious experience of greenness. If such a function lies at a fundamental level of neurobiological systems, this would seem to be along Searle’s line of argument. Despite such an obvious wanting of explanation for these phenomena, I believe that the arguments illustrated in this essay move the conversation further towards a much more scientifically stable, so to speak, theory of cognition.