Philosophy of artificial intelligence

From Wikipedia, the free encyclopedia

The philosophy of artificial intelligence attempts to answer such questions as follows:

• Can a machine act intelligently? Can it solve any problem that a person would solve by thinking?
• Are human intelligence and machine intelligence the same? Is the human brain essentially a computer?
• Can a machine have a mind, mental states, and consciousness in the same way that a human being can? Can it feel how things are?

These three questions reflect the divergent interests of AI researchers, linguists, cognitive scientists and philosophers respectively. The scientific answers to these questions depend on the definition of "intelligence" and "consciousness" and exactly which "machines" are under discussion.

Important propositions in the philosophy of AI include:

• Turing's "polite convention": If a machine behaves as intelligently as a human being, then it is as intelligent as a human being.^[2]
• The Dartmouth proposal: "Every aspect of learning or any other feature of intelligence can be so precisely described that a machine can be made to simulate it."^[3]
• Newell and Simon's physical symbol system hypothesis: "A physical symbol system has the necessary and sufficient means of general intelligent action."^[4]
• Searle's strong AI hypothesis: "The appropriately programmed computer with the right inputs and outputs would thereby have a mind in exactly the same sense human beings have minds."^[5]
• Hobbes' mechanism: "For 'reason' ... is nothing but 'reckoning,' that is adding and subtracting, of the consequences of general names agreed upon for the 'marking' and 'signifying' of our thoughts..."^[6]

Can a machine display general intelligence?

Is it possible to create a machine that can solve all the problems humans solve using their intelligence? This question defines the scope of what machines will be able to do in the future and guides the direction of AI research. It only concerns the behavior of machines and ignores the issues of interest to psychologists, cognitive scientists and philosophers; to answer this question, it does not matter whether a machine is really thinking (as a person thinks) or is just acting like it is thinking.^[7]
The basic position of most AI researchers is summed up in this statement, which appeared in the proposal for the Dartmouth workshop of 1956:

• Every aspect of learning or any other feature of intelligence can be so precisely described that a machine can be made to simulate it.^[3]

Arguments against the basic premise must show that building a working AI system is impossible, because there is some practical limit to the abilities of computers or that there is some special quality of the human mind that is necessary for thinking and yet cannot be duplicated by a machine (or by the methods of current AI research). Arguments in favor of the basic premise must show that such a system is possible.

The first step to answering the question is to clearly define "intelligence".

Intelligence

The "standard interpretation" of the Turing test.^[8]

Turing test

Alan Turing^[9] reduced the problem of defining intelligence to a simple question about conversation. He suggests that: if a machine can answer any question put to it, using the same words that an ordinary person would, then we may call that machine intelligent. A modern version of his experimental design would use an online chat room, where one of the participants is a real person and one of the participants is a computer program. The program passes the test if no one can tell which of the two participants is human.^[2] Turing notes that no one (except philosophers) ever asks the question "can people think?" He writes "instead of arguing continually over this point, it is usual to have a polite convention that everyone thinks".^[10] Turing's test extends this polite convention to machines:

• If a machine acts as intelligently as human being, then it is as intelligent as a human being.

One criticism of the Turing test is that it is explicitly anthropomorphic^{[citation needed]}. If our ultimate goal is to create machines that are more intelligent than people, why should we insist that our machines must closely resemble people?^{[This quote needs a citation]} Russell and Norvig write that "aeronautical engineering texts do not define the goal of their field as 'making machines that fly so exactly like pigeons that they can fool other pigeons'".^[11]

Intelligent agent definition

Simple reflex agent

Recent A.I. research defines intelligence in terms of intelligent agents. An "agent" is something which perceives and acts in an environment. A "performance measure" defines what counts as success for the agent.^[12]

• If an agent acts so as to maximize the expected value of a performance measure based on past experience and knowledge then it is intelligent.^[13]

Definitions like this one try to capture the essence of intelligence. They have the advantage that, unlike the Turing test, they do not also test for human traits that we^[who?] may not want to consider intelligent, like the ability to be insulted or the temptation to lie^{[dubious – discuss]}. They have the disadvantage that they fail to make the commonsense^{[when defined as?]} differentiation between "things that think" and "things that do not". By this definition, even a thermostat has a rudimentary intelligence.^[14]

Arguments that a machine can display general intelligence

The brain can be simulated

An MRI scan of a normal adult human brain

Hubert Dreyfus describes this argument as claiming that "if the nervous system obeys the laws of physics and chemistry, which we have every reason to suppose it does, then .... we ... ought to be able to reproduce the behavior of the nervous system with some physical device".^[15] This argument, first introduced as early as 1943^[16] and vividly described by Hans Moravec in 1988,^[17] is now associated with futurist Ray Kurzweil, who estimates that computer power will be sufficient for a complete brain simulation by the year 2029.^[18] A non-real-time simulation of a thalamocortical model that has the size of the human brain (10¹¹ neurons) was performed in 2005^[19] and it took 50 days to simulate 1 second of brain dynamics on a cluster of 27 processors.

Few^[quantify] disagree that a brain simulation is possible in theory, even critics of AI such as Hubert Dreyfus and John Searle.^[20] However, Searle points out that, in principle, anything can be simulated by a computer; thus, bringing the definition to its breaking point leads to the conclusion that any process at all can technically be considered "computation". "What we wanted to know is what distinguishes the mind from thermostats and livers," he writes.^[21] Thus, merely mimicking the functioning of a brain would in itself be an admission of ignorance regarding intelligence and the nature of the mind^{[citation needed]}.

Human thinking is symbol processing

In 1963, Allen Newell and Herbert A. Simon proposed that "symbol manipulation" was the essence of both human and machine intelligence. They wrote:

• A physical symbol system has the necessary and sufficient means of general intelligent action.^[4]

This claim is very strong: it implies both that human thinking is a kind of symbol manipulation (because a symbol system is necessary for intelligence) and that machines can be intelligent (because a symbol system is sufficient for intelligence).^[22] Another version of this position was described by philosopher Hubert Dreyfus, who called it "the psychological assumption":

• The mind can be viewed as a device operating on bits of information according to formal rules.^[23]

A distinction is usually made^{[by whom?]} between the kind of high level symbols that directly correspond with objects in the world, such as and and the more complex "symbols" that are present in a machine like a neural network. Early research into AI, called "good old fashioned artificial intelligence" (GOFAI) by John Haugeland, focused on these kind of high level symbols.^[24]

Arguments against symbol processing

These arguments show that human thinking does not consist (solely) of high level symbol manipulation. They do not show that artificial intelligence is impossible, only that more than symbol processing is required.

Gödelian anti-mechanist arguments

In 1931, Kurt Gödel proved with an incompleteness theorem that it is always possible to construct a "Gödel statement" that a given consistent formal system of logic (such as a high-level symbol manipulation program) could not prove. Despite being a true statement, the constructed Gödel statement is unprovable in the given system. (The truth of the constructed Gödel statement is contingent on the consistency of the given system; applying the same process to a subtly inconsistent system will appear to succeed, but will actually yield a false "Gödel statement" instead.) More speculatively, Gödel conjectured that the human mind can correctly eventually determine the truth or falsity of any well-grounded mathematical statement (including any possible Gödel statement), and that therefore the human mind's power is not reducible to a mechanism.^[25] Philosopher John Lucas (since 1961) and Roger Penrose (since 1989) have championed this philosophical anti-mechanist argument.^[26] Gödelian anti-mechanist arguments tend to rely on the innocuous-seeming claim that a system of human mathematicians (or some idealization of human mathematicians) is both consistent (completely free of error) and believes fully in its own consistency (and can make all logical inferences that follow from its own consistency, including belief in its Gödel statement)^{[citation needed]}. This is provably impossible for a Turing machine^{[clarification needed]} (and, by an informal extension, any known type of mechanical computer) to do; therefore, the Gödelian concludes that human reasoning is too powerful to be captured in a machine^{[dubious – discuss]}.
However, the modern consensus in the scientific and mathematical community is that actual human reasoning is inconsistent; that any consistent "idealized version" H of human reasoning would logically be forced to adopt a healthy but counter-intuitive open-minded skepticism about the consistency of H (otherwise H is provably inconsistent); and that Gödel's theorems do not lead to any valid argument that humans have mathematical reasoning capabilities beyond what a machine could ever duplicate.^[27][28][29] This consensus that Gödelian anti-mechanist arguments are doomed to failure is laid out strongly in Artificial Intelligence: "any attempt to utilize (Gödel's incompleteness results) to attack the computationalist thesis is bound to be illegitimate, since these results are quite consistent with the computationalist thesis."^[30]

More pragmatically, Russell and Norvig note that Gödel's argument only applies to what can theoretically be proved, given an infinite amount of memory and time. In practice, real machines (including humans) have finite resources and will have difficulty proving many theorems. It is not necessary to prove everything in order to be intelligent^{[when defined as?]}.^[31]

Less formally, Douglas Hofstadter, in his Pulitzer prize winning book Gödel, Escher, Bach: An Eternal Golden Braid, states that these "Gödel-statements" always refer to the system itself, drawing an analogy to the way the Epimenides paradox uses statements that refer to themselves, such as "this statement is false" or "I am lying".^[32] But, of course, the Epimenides paradox applies to anything that makes statements, whether they are machines or humans, even Lucas himself. Consider:

• Lucas can't assert the truth of this statement.^[33]

This statement is true but cannot be asserted by Lucas. This shows that Lucas himself is subject to the same limits that he describes for machines, as are all people, and so Lucas's argument is pointless.^[34]
After concluding that human reasoning is non-computable, Penrose went on to controversially speculate that some kind of hypothetical non-computable processes involving the collapse of quantum mechanical states give humans a special advantage over existing computers. Existing quantum computers are only capable of reducing the complexity of Turing computable tasks and are still restricted to tasks within the scope of Turing machines.^{[citation needed][clarification needed]}. By Penrose and Lucas's arguments, existing quantum computers are not sufficient, so Penrose seeks for some other process involving new physics, for instance quantum gravity which might manifest new physics at the scale of the Planck mass via spontaneous quantum collapse of the wave function. These states, he suggested, occur both within neurons and also spanning more than one neuron.^[35] However, other scientists point out that there is no plausible organic mechanism in the brain for harnessing any sort of quantum computation, and furthermore that the timescale of quantum decoherence seems too fast to influence neuron firing.^[36]

Dreyfus: the primacy of unconscious skills

Hubert Dreyfus argued that human intelligence and expertise depended primarily on unconscious instincts rather than conscious symbolic manipulation, and argued that these unconscious skills would never be captured in formal rules.^[37]
Dreyfus's argument had been anticipated by Turing in his 1950 paper Computing machinery and intelligence, where he had classified this as the "argument from the informality of behavior."^[38] Turing argued in response that, just because we do not know the rules that govern a complex behavior, this does not mean that no such rules exist. He wrote: "we cannot so easily convince ourselves of the absence of complete laws of behaviour ... The only way we know of for finding such laws is scientific observation, and we certainly know of no circumstances under which we could say, 'We have searched enough. There are no such laws.'"^[39]

Russell and Norvig point out that, in the years since Dreyfus published his critique, progress has been made towards discovering the "rules" that govern unconscious reasoning.^[40] The situated movement in robotics research attempts to capture our unconscious skills at perception and attention.^[41]  Computational intelligence paradigms, such as neural nets, evolutionary algorithms and so on are mostly directed at simulated unconscious reasoning and learning. Statistical approaches to AI can make predictions which approach the accuracy of human intuitive guesses. Research into commonsense knowledge has focused on reproducing the "background" or context of knowledge. In fact, AI research in general has moved away from high level symbol manipulation or "GOFAI", towards new models that are intended to capture more of our unconscious reasoning. Historian and AI researcher Daniel Crevier wrote that "time has proven the accuracy and perceptiveness of some of Dreyfus's comments. Had he formulated them less aggressively, constructive actions they suggested might have been taken much earlier."^[42]

Can a machine have a mind, consciousness, and mental states?

This is a philosophical question, related to the problem of other minds and the hard problem of consciousness. The question revolves around a position defined by John Searle as "strong AI":

• A physical symbol system can have a mind and mental states.^[5]

Searle distinguished this position from what he called "weak AI":

• A physical symbol system can act intelligently.^[5]

Searle introduced the terms to isolate strong AI from weak AI so he could focus on what he thought was the more interesting and debatable issue. He argued that even if we assume that we had a computer program that acted exactly like a human mind, there would still be a difficult philosophical question that needed to be answered.^[5]

Neither of Searle's two positions are of great concern to AI research, since they do not directly answer the question "can a machine display general intelligence?" (unless it can also be shown that consciousness is necessary for intelligence). Turing wrote "I do not wish to give the impression that I think there is no mystery about consciousness… [b]ut I do not think these mysteries necessarily need to be solved before we can answer the question [of whether machines can think]."^[43] Russell and Norvig agree: "Most AI researchers take the weak AI hypothesis for granted, and don't care about the strong AI hypothesis."^[44]

There are a few researchers who believe that consciousness is an essential element in intelligence, such as Igor Aleksander, Stan Franklin, Ron Sun, and Pentti Haikonen, although their definition of "consciousness" strays very close to "intelligence." (See artificial consciousness.)

Before we can answer this question, we must be clear what we mean by "minds", "mental states" and "consciousness".

Consciousness, minds, mental states, meaning

The words "mind" and "consciousness" are used by different communities in different ways. Some new age thinkers, for example, use the word "consciousness" to describe something similar to Bergson's "élan vital": an invisible, energetic fluid that permeates life and especially the mind.  Science fiction writers use the word to describe some essential property that makes us human: a machine or alien that is "conscious" will be presented as a fully human character, with intelligence, desires, will, insight, pride and so on. (Science fiction writers also use the words "sentience", "sapience," "self-awareness" or "ghost" - as in the Ghost in the Shell manga and anime series - to describe this essential human property). For others^[who?], the words "mind" or "consciousness" are used as a kind of secular synonym for the soul.

For philosophers, neuroscientists and cognitive scientists, the words are used in a way that is both more precise and more mundane: they refer to the familiar, everyday experience of having a "thought in your head", like a perception, a dream, an intention or a plan, and to the way we know something, or mean something or understand something^{[citation needed]}. "It's not hard to give a commonsense definition of consciousness" observes philosopher John Searle.^[45] What is mysterious and fascinating is not so much what it is but how it is: how does a lump of fatty tissue and electricity give rise to this (familiar) experience of perceiving, meaning or thinking?

Philosophers call this the hard problem of consciousness. It is the latest version of a classic problem in the philosophy of mind called the "mind-body problem."^[46] A related problem is the problem of meaning or understanding (which philosophers call "intentionality"): what is the connection between our thoughts and what we are thinking about (i.e. objects and situations out in the world)? A third issue is the problem of experience (or "phenomenology"): If two people see the same thing, do they have the same experience? Or are there things "inside their head" (called "qualia") that can be different from person to person?^[47]

Neurobiologists believe all these problems will be solved as we begin to identify the neural correlates of consciousness: the actual relationship between the machinery in our heads and its collective properties; such as the mind, experience and understanding. Some of the harshest critics of artificial intelligence agree that the brain is just a machine, and that consciousness and intelligence are the result of physical processes in the brain.^[48] The difficult philosophical question is this: can a computer program, running on a digital machine that shuffles the binary digits of zero and one, duplicate the ability of the neurons to create minds, with mental states (like understanding or perceiving), and ultimately, the experience of consciousness?

Arguments that a computer cannot have a mind and mental states

Searle's Chinese room

John Searle asks us to consider a thought experiment: suppose we have written a computer program that passes the Turing test and demonstrates "general intelligent action." Suppose, specifically that the program can converse in fluent Chinese. Write the program on 3x5 cards and give them to an ordinary person who does not speak Chinese. Lock the person into a room and have him follow the instructions on the cards. He will copy out Chinese characters and pass them in and out of the room through a slot. From the outside, it will appear that the Chinese room contains a fully intelligent person who speaks Chinese. The question is this: is there anyone (or anything) in the room that understands Chinese? That is, is there anything that has the mental state of understanding, or which has conscious awareness of what is being discussed in Chinese? The man is clearly not aware. The room cannot be aware. The cards certainly aren't aware. Searle concludes that the Chinese room, or any other physical symbol system, cannot have a mind.^[49]
Searle goes on to argue that actual mental states and consciousness require (yet to be described) "actual physical-chemical properties of actual human brains."^[50] He argues there are special "causal properties" of brains and neurons that gives rise to minds: in his words "brains cause minds."^[51]

Related arguments: Leibniz' mill, Davis's telephone exchange, Block's Chinese nation and Blockhead

Gottfried Leibniz made essentially the same argument as Searle in 1714, using the thought experiment of expanding the brain until it was the size of a mill.^[52] In 1974, Lawrence Davis imagined duplicating the brain using telephone lines and offices staffed by people, and in 1978 Ned Block envisioned the entire population of China involved in such a brain simulation. This thought experiment is called "the Chinese Nation" or "the Chinese Gym".^[53] Ned Block also proposed his Blockhead argument, which is a version of the Chinese room in which the program has been re-factored into a simple set of rules of the form "see this, do that", removing all mystery from the program.

Responses to the Chinese room

Responses to the Chinese room emphasize several different points.

• The systems reply and the virtual mind reply:^[54] This reply argues that the system, including the man, the program, the room, and the cards, is what understands Chinese. Searle claims that the man in the room is the only thing which could possibly "have a mind" or "understand", but others disagree, arguing that it is possible for there to be two minds in the same physical place, similar to the way a computer can simultaneously "be" two machines at once: one physical (like a Macintosh) and one "virtual" (like a word processor).
• Speed, power and complexity replies:^[55] Several critics point out that the man in the room would probably take millions of years to respond to a simple question, and would require "filing cabinets" of astronomical proportions. This brings the clarity of Searle's intuition into doubt.
• Robot reply:^[56] To truly understand, some believe the Chinese Room needs eyes and hands. Hans Moravec writes: 'If we could graft a robot to a reasoning program, we wouldn't need a person to provide the meaning anymore: it would come from the physical world."^[57]
• Brain simulator reply:^[58] What if the program simulates the sequence of nerve firings at the synapses of an actual brain of an actual Chinese speaker? The man in the room would be simulating an actual brain. This is a variation on the "systems reply" that appears more plausible because "the system" now clearly operates like a human brain, which strengthens the intuition that there is something besides the man in the room that could understand Chinese.
• Other minds reply and the epiphenomena reply:^[59] Several people have noted that Searle's argument is just a version of the problem of other minds, applied to machines. Since it is difficult to decide if people are "actually" thinking, we should not be surprised that it is difficult to answer the same question about machines.

A related question is whether "consciousness" (as Searle understands it) exists. Searle argues that the experience of consciousness can't be detected by examining the behavior of a machine, a human being or any other animal. Daniel Dennett points out that natural selection cannot preserve a feature of an animal that has no effect on the behavior of the animal, and thus consciousness (as Searle understands it) can't be produced by natural selection. Therefore either natural selection did not produce consciousness, or "strong AI" is correct in that consciousness can be detected by suitably designed Turing test.

Is thinking a kind of computation?

The computational theory of mind or "computationalism" claims that the relationship between mind and brain is similar (if not identical) to the relationship between a running program and a computer. The idea has philosophical roots in Hobbes (who claimed reasoning was "nothing more than reckoning"), Leibniz (who attempted to create a logical calculus of all human ideas), Hume (who thought perception could be reduced to "atomic impressions") and even Kant (who analyzed all experience as controlled by formal rules).^[60] The latest version is associated with philosophers Hilary Putnam and Jerry Fodor.^[61]
This question bears on our earlier questions: if the human brain is a kind of computer then computers can be both intelligent and conscious, answering both the practical and philosophical questions of AI. In terms of the practical question of AI ("Can a machine display general intelligence?"), some versions of computationalism make the claim that (as Hobbes wrote):

• Reasoning is nothing but reckoning^[6]

In other words, our intelligence derives from a form of calculation, similar to arithmetic. This is the physical symbol system hypothesis discussed above, and it implies that artificial intelligence is possible. In terms of the philosophical question of AI ("Can a machine have mind, mental states and consciousness?"), most versions of computationalism claim that (as Stevan Harnad characterizes it):

• Mental states are just implementations of (the right) computer programs^[62]

This is John Searle's "strong AI" discussed above, and it is the real target of the Chinese room argument (according to Harnad).^[62]

Other related questions

Alan Turing noted that there are many arguments of the form "a machine will never do X", where X can be many things, such as:

Be kind, resourceful, beautiful, friendly, have initiative, have a sense of humor, tell right from wrong, make mistakes, fall in love, enjoy strawberries and cream, make someone fall in love with it, learn from experience, use words properly, be the subject of its own thought, have as much diversity of behaviour as a man, do something really new.^[63]

Turing argues that these objections are often based on naive assumptions about the versatility of machines or are "disguised forms of the argument from consciousness". Writing a program that exhibits one of these behaviors "will not make much of an impression."^[63] All of these arguments are tangential to the basic premise of AI, unless it can be shown that one of these traits is essential for general intelligence.

Can a machine have emotions?

If "emotions" are defined only in terms of their effect on behavior or on how they function inside an organism, then emotions can be viewed as a mechanism that an intelligent agent uses to maximize the utility of its actions. Given this definition of emotion, Hans Moravec believes that "robots in general will be quite emotional about being nice people".^[64] Fear is a source of urgency. Empathy is a necessary component of good human computer interaction. He says robots "will try to please you in an apparently selfless manner because it will get a thrill out of this positive reinforcement. You can interpret this as a kind of love."^[64] Daniel Crevier writes "Moravec's point is that emotions are just devices for channeling behavior in a direction beneficial to the survival of one's species."^[65]

However, emotions can also be defined in terms of their subjective quality, of what it feels like to have an emotion. The question of whether the machine actually feels an emotion, or whether it merely acts as if it is feeling an emotion is the philosophical question, "can a machine be conscious?" in another form.^[43]

Can a machine be self-aware?

"Self awareness", as noted above, is sometimes used by science fiction writers as a name for the essential human property that makes a character fully human. Turing strips away all other properties of human beings and reduces the question to "can a machine be the subject of its own thought?" Can it think about itself? Viewed in this way, it is obvious that a program can be written that can report on its own internal states, such as a debugger.^[63] Though arguably self-awareness often presumes a bit more capability; a machine that can ascribe meaning in some way to not only its own state but in general postulating questions without solid answers: the contextual nature of its existence now; how it compares to past states or plans for the future, the limits and value of its work product, how it perceives its performance to be valued-by or compared to others.

Can a machine be original or creative?

Turing reduces this to the question of whether a machine can "take us by surprise" and argues that this is obviously true, as any programmer can attest.^[66] He notes that, with enough storage capacity, a computer can behave in an astronomical number of different ways.^[67] It must be possible, even trivial, for a computer that can represent ideas to combine them in new ways. (Douglas Lenat's Automated Mathematician, as one example, combined ideas to discover new mathematical truths.)

In 2009, scientists at Aberystwyth University in Wales and the U.K's University of Cambridge designed a robot called Adam that they believe to be the first machine to independently come up with new scientific findings.^[68] Also in 2009, researchers at Cornell developed Eureqa, a computer program that extrapolates formulas to fit the data inputted, such as finding the laws of motion from a pendulum's motion.

Can a machine be benevolent or hostile?

This question (like many others in the philosophy of artificial intelligence) can be presented in two forms. "Hostility" can be defined in terms function or behavior, in which case "hostile" becomes synonymous with "dangerous". Or it can be defined in terms of intent: can a machine "deliberately" set out to do harm? The latter is the question "can a machine have conscious states?" (such as intentions) in another form.^[43]
The question of whether highly intelligent and completely autonomous machines would be dangerous has been examined in detail by futurists (such as the Singularity Institute). (The obvious element of drama has also made the subject popular in science fiction, which has considered many differently possible scenarios where intelligent machines pose a threat to mankind.)

One issue is that machines may acquire the autonomy and intelligence required to be dangerous very quickly. Vernor Vinge has suggested that over just a few years, computers will suddenly become thousands or millions of times more intelligent than humans. He calls this "the Singularity."^[69] He suggests that it may be somewhat or possibly very dangerous for humans.^[70] This is discussed by a philosophy called Singularitarianism.

In 2009, academics and technical experts attended a conference to discuss the potential impact of robots and computers and the impact of the hypothetical possibility that they could become self-sufficient and able to make their own decisions. They discussed the possibility and the extent to which computers and robots might be able to acquire any level of autonomy, and to what degree they could use such abilities to possibly pose any threat or hazard. They noted that some machines have acquired various forms of semi-autonomy, including being able to find power sources on their own and being able to independently choose targets to attack with weapons. They also noted that some computer viruses can evade elimination and have achieved "cockroach intelligence." They noted that self-awareness as depicted in science-fiction is probably unlikely, but that there were other potential hazards and pitfalls.^[69]

Some experts and academics have questioned the use of robots for military combat, especially when such robots are given some degree of autonomous functions.^[71] The US Navy has funded a report which indicates that as military robots become more complex, there should be greater attention to implications of their ability to make autonomous decisions.^[72][73]

The President of the Association for the Advancement of Artificial Intelligence has commissioned a study to look at this issue.^[74] They point to programs like the Language Acquisition Device which can emulate human interaction.

Some have suggested a need to build "Friendly AI", meaning that the advances which are already occurring with AI should also include an effort to make AI intrinsically friendly and humane.^[75]

Can a machine have a soul?

Finally, those who believe in the existence of a soul may argue that "Thinking is a function of man's immortal soul." Alan Turing called this "the theological objection". He writes

In attempting to construct such machines we should not be irreverently usurping His power of creating souls, any more than we are in the procreation of children: rather we are, in either case, instruments of His will providing mansions for the souls that He creates.

Neurophilosophy

From Wikipedia, the free encyclopedia

Neurophilosophy or philosophy of neuroscience is the interdisciplinary study of neuroscience and philosophy that explores the relevance of neuroscientific studies to the arguments traditionally categorized as philosophy of mind. The philosophy of neuroscience attempts to clarify neuroscientific methods and results using the conceptual rigor and methods of philosophy of science.

Specific issues

Below is a list of specific issues important to philosophy of neuroscience:

• "The indirectness of studies of mind and brain"^[1]
• "Computational or representational analysis of brain processing"^[2]
• "Relations between psychological and neuroscientific inquiries"^[3]
• Modularity of mind^[2]
• What constitutes adequate explanation in neuroscience?^[4]
• "Location of cognitive function"^[5]

The indirectness of studies of mind and brain

Many of the methods and techniques central to neuroscientific discovery rely on assumptions that can limit the interpretation of the data. Philosophers of Neuroscience have discussed such assumptions in the use of functional Magnetic Resonance Imaging,^[6][7] Dissociation in Cognitive Neuropsychology,^[8][9] single unit recording,^[10] and computational neuroscience.^[11] Following are descriptions of many of the current controversies and debates about the methods employed in neuroscience.

fMRI

Many fMRI studies rely heavily on the assumption of "localization of function"^[12](same as functional specialization). Localization of function means that many cognitive functions can be localized to specific brain regions. A good example of functional localization comes from studies of the motor cortex.^[13] There seem to be different groups of cells in the motor cortex responsible for controlling different groups of muscles. Many philosophers of neuroscience criticize fMRI for relying too heavily on this assumption. Michael Anderson points out that subtraction method fMRI misses a lot of brain information that is important to the cognitive processes.^[14] Subtraction fMRI only shows the differences between the task activation and the control activation, but many of the brain areas activated in the control are obviously important for the task as well.

Some philosophers entirely reject any notion of localization of function and thus believe fMRI studies to be profoundly misguided.^[15] These philosophers maintain that brain processing acts holistically, that large sections of the brain are involved in processing most cognitive tasks (see holism in neurology and the modularity section below). One way to understand their objection to the idea of localization of function is the radio repair man thought experiment.^[16] In this thought experiment, a radio repair man opens up a radio and rips out a tube. The radio begins whistling loudly and the radio repair man declares that he must have ripped out the anti-whistling tube. There is not really any anti-whistling tube in the radio and it is obvious that the radio repair man has confounded function with effect. This criticism was originally targeted at the logic used by neuropsychological brain lesion experiments, but the criticism is still applicable to neuroimaging. These considerations are similar to Van Orden's and Paap's criticism of circularity in neuroimaging logic.^[17] According to them, neuroimagers assume that their theory of cognitive component parcellation is correct and that these components divide cleanly into feed-forward modules. These assumptions are necessary to justify their inference of brain localization. The logic is circular if the researcher then use the appearance of brain region activation as proof of the correctness of their cognitive theories.

A different problematic methodological assumption within fMRI research is the use of reverse inference^[18] A reverse inference is when the activation of a brain region is used to infer the presence of a given cognitive process. Poldrack points out that the strength of this inference depends critically on the likelihood that a given task employs a given cognitive process and the likelihood of that pattern of brain activation given that cognitive process. In other words, the strength of reverse inference is based upon the selectivity of the task used as well as the selectivity of the brain region activation. A 2011 article published in the NY times has been heavily criticized for misusing reverse inference.^[19] In the study, participants were shown pictures of their iPhones and the researchers measured activation of the insula. The researches took insula activation as evidence of feelings of love and concluded that people loved their iPhones. Critics were quick to point out that the insula is not a very selective piece of cortex, and therefore not amenable to reverse inference.

The Neuropsychologist Max Coltheart took the problems with reverse inference a step further and challenged neuroimagers to give one instance in which neuroimaging had informed psychological theory^[20] Coltheart takes the burden of proof to be an instance where the brain imaging data is consistent with one theory but inconsistent with another theory. Roskies maintains that Coltheart's ultra cognitive position makes his challenge unwinnable.^[21] Since Coltheart maintains that the implementation of a cognitive state has no bearing on the function of that cognitive state, then it is impossible to find neuroimaging data that will be able to comment on psychological theories in the way Coltheart demands. Neuroimaging data will always be relegated to the lower level of implementation and be unable to selectively determine one or another cognitive theory. In a 2006 article, Richard Henson suggests that forward inference can be used to infer dissociation of function at the psychological level.^[22] He suggests that these kinds of inferences can be made when there is crossing activations between two task types in two brain regions and there is no change in activation in a mutual control region.

One final assumption worth mentioning is the assumption of pure insertion in fMRI.^[23] The assumption of pure insertion is the assumption that a single cognitive process can be inserted into another set of cognitive process without effecting the functioning of the rest. For example, if you wanted to find the reading comprehension area of the brain, you might scan participants while they were presented with a word and while they were presented with a non-word (e.g. "Floob"). If you infer that the resulting difference in brain pattern represents the regions of the brain involved in reading comprehension, you have assumed that these changes are not reflective of changes in task difficulty or differential recruitment between tasks. The term pure insertion was coined by Donders as a criticism of reaction time methods.

Recently, researchers have begun using a new functional imaging technique called resting state functional connectivity MRI.^[24] Subjects' brains are scanned while the subject sits idly in the scanner. By looking at the natural fluctuations in the bold pattern while the subject is at rest, the researchers can see which brain regions co-vary in activation together. They can use the patterns of covariance to construct maps of functionally linked brain areas. It is worth noting that the name "functional connectivity" is somewhat misleading since the data only indicates co-variation. Still, this is a powerful method for studying large networks throughout the brain. There are a couple of important methodological issues that need to be addressed. Firstly, there are many different possible brain mappings that could be used to define the brain regions for the network. The results could vary significantly depending on the brain region chosen. Secondly, what mathematical techniques are best about to characterize these brain regions?

The brain regions of interest are somewhat constrained by the size of the voxels. Rs-fcMRI uses voxels that are few millimeters cubed so the brain regions will have to be defined on a larger scale. Two of the statistical methods that are commonly applied to network analysis can work on the single voxel spatial scale, but graph theory methods are extremely sensitive to the way nodes are defined. Brains regions can be divided according to their cellular architectural, according to their connectivity, or according to physiological measures. Alternatively, you could take a theory neutral approach and randomly divide the cortex into partitions of the size of your choosing. As mentioned earlier, there are several approaches to network analysis once the your brain regions have been defined. Seed based analysis begins with an a priori defined seed region and finds all of the regions that are functionally connected to that region. Wig et al. caution that the resulting network structure will not give any information concerning the inter-connectivity of the identified regions or the relations of those regions to regions other than the seed region. Another approach is to use independent component analysis to create spatio-temporal component maps and the components are sorted by components that carry information of interest and those that are caused by noise. Wigs et al. once again warns that inference of functional brain region communities is difficult under ICA. ICA also has the issue of imposing orthogonality on the data.^[25] Graph theory uses a matrix to characterize covariance between regions which is then transformed into a network map. The problem with graph theory analysis is that network mapping is heavily influenced by a priori brain region and connectivity (nodes and edges), thus the researcher is at risk for cherry picking regions and connections according to their own theories. However, graph theory analysis is extremely valuable since it is the only method that gives pair-wise relationships between nodes. ICA has the added advantage of being a fairly principled method. It seems that using both methods will be important in uncovering the network connectivity of the brain. Mumford et al. hoped to avoid these issues and use a principled approach that could determine pair-wise relationships using a statistical technique adopted from analysis of gene co-expression networks.

Dissociation in cognitive neuropsychology

Cognitive Neuropsychology studies brain damaged patients and uses the patterns of selective impairment in order to make inferences on the underlying cognitive structure. Dissociation between cognitive functions is taken to be evidence that these functions are independent. Theorists have identified several key assumptions that are needed to justify these inferences:^[26] 1) Functional Modularity- the mind is organized into functionally separate cognitive modules. 2). Anatomical Modularity- the brain is organized into functionally separate modules. This assumption is very similar to the assumption of functional localization. These assumptions differ from the assumption of functional modularity, because it is possible to have separable cognitive modules that are implemented by diffuse patterns of brain activation. 3)Universality- The basic organization of functional and anatomical modularity is the same for all normal humans. This assumption is needed if we are to make any claim about functional organization based on dissociation that extrapolates from the instance of a case study to the population. 4) Transparency / Subtractivity- the mind does not undergo substantial reorganization following brain damage. It is possible to remove one functional module without significantly altering the overall structure of the system. This assumption is necessary in order to justify using brain damaged patients in order to make inferences about the cognitive architecture of healthy people.

There are three principal types of evidence in cognitive neuropsychology: association, single dissociation and double dissociation.^[27] Association inferences observe that certain deficits are likely to co-occur. For example, there are many cases who have deficits in both abstract and concrete word comprehension following brain damage. Association studies are considered the weakest form of evidence, because the results could be accounted for by damage to neighboring brain regions and not damage to a single cognitive system.^[28] Single Dissociation inferences observe that one cognitive faculty can be spared while another can be damaged following brain damage. This pattern indicates that a) the two tasks employ different cognitive systems b) the two tasks occupy the same system and the damaged task is downstream from the spared task or c) that the spared task requires fewer cognitive resources than the damaged task. The "gold standard" for cognitive neuropsychology is the double dissociation. Double dissociation occurs when brain damage impairs task A in Patient1 but spares task B and brain damage spares task A in Patient 2 but damages task B. It is assumed that one instance of double dissociation is sufficient proof to infer separate cognitive modules in the performance of the tasks.

Many theorists criticize cognitive neuropsychology for its dependence on double dissociations. In one widely cited study, Joula and Plunkett used a model connectionist system to demonstrate that double dissociation behavioral patterns can occur through random lesions of a single module.^[29] They created a multilayer connectionist system trained to pronounce words. They repeatedly simulated random destruction of nodes and connections in the system and plotted the resulting performance on a scatter plot. The results showed deficits in irregular noun pronunciation with spared regular verb pronunciation in some cases and deficits in regular verb pronunciation with spared irregular noun pronunciation. These results suggest that a single instance of double dissociation is insufficient to justify inference to multiple systems.^[30]

Charter offers a theoretical case in which double dissociation logic can be faulty.^[31] If two tasks, task A and task B, use almost all of the same systems but differ by one mutually exclusive module apiece, then the selective lesioning of those two modules would seem to indicate that A and B use different systems. Charter uses the example of someone who is allergic to peanuts but not shrimp and someone who is allergic to shrimp and not peanuts. He argues that double dissociation logic leads one to infer that peanuts and shrimp are digested by different systems. John Dunn offers another objection to double dissociation.^[32] He claims that it is easy to demonstrate the existence of a true deficit but difficult to show that another function is truly spared. As more data is accumulated, the value of your results will converge on an effect size of zero, but there will always be a positive value greater than zero that has more statistical power than zero. Therefore, it is impossible to be fully confident that a given double dissociation actually exists.

On a different note, Alphonso Caramazza has given a principled reason for rejecting the use of group studies in cognitive neuropsychology.^[33] Studies of brain damaged patients can either take the form of a single case study, in which an individual's behavior is characterized and used as evidence, or group studies, in which a group of patients displaying the same deficit have their behavior characterized and averaged. In order to justify grouping a set of patient data together, the researcher must know that the group is homogenous, that their behavior is equivalent in every theoretically meaningful way. In brain damaged patients, this can only be accomplished a posteriori by analyzing the behavior patterns of all the individuals in the group. Thus according to Caramazza, any group study is either the equivalent of a set of single case studies or is theoretically unjustified. Newcombe and Marshall pointed out that there are some cases (they use Geschwind's syndrome as an example) and that group studies might still serve as a useful heuristic in cognitive neuropsychological studies.^[34]

Single unit recordings

It is commonly understood in neuroscience that information is encoded in the brain by the firing patterns of neurons.^[35] Many of the philosophical questions surrounding the neural code are related to questions about representation and computation that are discussed below. There are other methodological questions including whether neurons represent information through an average firing rate or whether there is information represented by the temporal dynamics. There are similar questions about whether neurons represent information individually or as a population.

Computational neuroscience

Many of the philosophical controversies surrounding computational neuroscience involve the role of simulation and modeling as explanation. Carl Craver has been especially vocal about such interpretations.^[36] Jones and Love wrote an especially critical article targeted at Bayesian behavioral modeling that did not constrain the modeling parameters by psychological or neurological considerations^[37] Eric Winsberg has written about the role of computer modeling and simulation in science generally, but his characterization is applicable to computational neuroscience.^[38]

Computation and representation in the brain

The computational theory of mind has been widespread in neuroscience since the cognitive revolution in the 1960s. This section will begin with a historical overview of computational neuroscience and then discuss various competing theories and controversies within the field.

Historical overview

Computational neuroscience began in the 1930s and 1940s with two groups of researchers. The first group consisted of Alan Turing, Alonzo Church and John von Neumann, who were working to develop computing machines and the mathematical underpinnings of computer science.^[39] This work culminated in the theoretical development of so-called Turing machines and the Church–Turing thesis, which formalized the mathematics underlying computability theory. The second group consisted of Warren McCulloch and Walter Pitts who were working to develop the first artificial neural networks. McCulloch and Pitts were the first to hypothesize that neurons could be used to implement a logical calculus that could explain cognition. They used their toy neurons to develop logic gates that could make computations.^[40] However these developments failed to take hold in the psychological sciences and neuroscience until the mid-1950s and 1960s. Behaviorism had dominated the psychology until the 1950s when new developments in a variety of fields overturned behaviorist theory in favor of a cognitive theory. From the beginning of the cognitive revolution, computational theory played a major role in theoretical developments. Minsky and McCarthy's work in artificial intelligence, Newell and Simon's computer simulations, and Noam Chomsky's importation of information theory into linguistics were all heavily reliant on computational assumptions.^[41] By the early 1960s, Hilary Putnam was arguing in favor of machine functionalism in which the brain instantiated Turing machines. By this point computational theories were firmly fixed in psychology and neuroscience. By the mid-1980s, a group of researchers began using multilayer feed-forward analog neural networks that could be trained to perform a variety of tasks. The work by researchers like Sejnowski, Rosenberg, Rumelhart, and McClelland were labeled as connectionism, and the discipline has continued since then.^[42] The connectionist mindset was embraced by Paul and Patricia Churchland who then developed their "state space semantics" using concepts from connectionist theory. Connectionism was also condemned by researchers such as Fodor, Pylyshyn, and Pinker. The tension between the connectionists and the classicists is still being debated today.

Representation

One of the reasons that computational theories are appealing is that computers have the ability to manipulate representations to give meaningful output. Digital computers use strings of 1s and 0s in order to represent the content such as this Wikipedia page. Most cognitive scientists posit that our brains use some form of representational code that is carried in the firing patterns of neurons. Computational accounts seem to offer an easy way of explaining how our brains carry and manipulate the perceptions, thoughts, feelings, and actions that make up our everyday experience.^[43] While most theorists maintain that representation is an important part of cognition, the exact nature of that representation is highly debated. The two main arguments come from advocates of symbolic representations and advocates of associationist representations.

Symbolic representational accounts have been famously championed by Fodor and Pinker. Symbolic representation means that the objects are represented by symbols and are processed through rule governed manipulations that are sensation to the constitutive structure. The fact that symbolic representation is sensitive to the structure of the representations is a major part of its appeal. Fodor proposed the Language of Thought Hypothesis in which mental representations manipulated in the same way that language is syntactically manipulated in order to produce thought. According to Fodor, the language of thought hypothesis explains the systematicity and productivity seen in both language and thought.^[44] Blechner^[45] proposed that dreams reveal a figurative-affective representation that may represent the language of thought or, more properly, the substrate of thought. Blechner’s model accounts for how evolutionarily earlier forms of humans and non-human mammals could think, dream, and solve problems without language. Waking thought became a later development in which underlying figurative-affective thinking is transformed into communicable linguistic forms. It would also account for the way problem-solving can occur in dreams.^[46]

Associativist representations are most often described with connectionist systems. In connectionist systems, representations are distributed across all the nodes and connection weights of the system and thus are said to be sub symbolic.^[47] It is worth noting that a connectionist system is capable of implementing a symbolic system. There are several important aspects of neural nets that suggest that distributed parallel processing provides a better basis for cognitive functions than symbolic processing. Firstly, the inspiration for these systems came from the brain itself indicating biological relevance. Secondly, these systems are capable of storing content addressable memory, which is far more efficient than memory searches in symbolic systems. Thirdly, neural nets are resilient to damage while even minor damage can disable a symbolic system. Lastly, soft constraints and generalization when processing novel stimuli allow nets to behave more flexibly than symbolic systems.

The Churchlands described representation in a connectionist system in terms of state space. The content of the system is represented by an n-dimensional vector where the n= the number of nodes in the system and the direction of the vector is determined by the activation pattern of the nodes. Fodor rejected this method of representation on the grounds that two different connectionist systems could not have the same content.^[48] Further mathematical analysis of connectionist system relieved that connectionist systems that could contain similar content could be mapped graphically to reveal clusters of nodes that were important to representing the content.^[49] Unfortunately for the Churchlands, state space vector comparison was not amenable to this type of analysis. Recently, Nicholas Shea has offered his own account for content within connectionist systems that employs the concepts developed through cluster analysis.

Views on computation

Computational neuroscience is committed to the position that the brain is some sort of computer, but what does it mean to be a computer? The definition of a computation must be narrow enough so that we limit the number of objects that can be called computers. For example, it might seem problematic to have a definition wide enough to allow stomachs and weather systems to be involved in computations. However, it is also necessary to have a definition broad enough to allow all of the wide varieties of computational systems to compute. For example, if the definition of computation is limited to syntactic manipulation of symbolic representations, then most connectionist systems would not be able to compute.^[50] Rick Grush distinguishes between computation as a tool for simulation and computation as a theoretical stance in cognitive neuroscience.^[51] For the former, anything that can be computationally modeled counts as computing. In the latter case, the brain is a computing function that is distinct from systems like fluid dynamic systems and the planetary orbits in this regard. The challenge for any computational definition is to keep the two senses distinct.

Alternatively, some theorists choose to accept a narrow or wide definition for theoretical reasons. Pancomputationalism is the position that everything can be said to compute. This view has been criticized by Piccinini on the grounds that such a definition makes computation trivial to the point where it is robbed of its explanatory value.^[52]

The simplest definition of computations is that a system can be said to be computing when a computational description can be mapped onto the physical description. This is an extremely broad definition of computation and it ends up endorsing a form of pancomputationalism. Putnam and Searle, who are often credited with this view, maintain that computation is observer-related. In other words, if you want to view a system as computing then you can say that it is computing. Piccinini points out that, in this view, not only is everything computing, but also everything is computing in an indefinite number of ways.^[53] Since it is possible to apply an indefinite number of computational descriptions to a given system, the system ends up computing an indefinite number of tasks.

The most common view of computation is the semantic account of computation. Semantic approaches use a similar notion of computation as the mapping approaches with the added constraint that the system must manipulate representations with semantic content. Note from the earlier discussion of representation that both the Churchlands' connectionist systems and Fodor's symbolic systems use this notion of computation. In fact, Fodor is famously credited as saying "No computation without representation".^[54] Computational states can be individuated by an externalized appeal to content in a broad sense (i.e. the object in the external world) or by internalist appeal to the narrow sense content (content defined by the properties of the system).^[55] In order to fix the content of the representation, it is often necessary to appeal to the information contained within the system. Grush provides a criticism of the semantic account.^[51] He points out that appeal to the informational content of a system to demonstrate representation by the system. He uses his coffee cup as an example of a system that contains information, such as the heat conductance of the coffee cup and the time since the coffee was poured, but is too mundane to compute in any robust sense. Semantic computationalists try to escape this criticism by appealing to the evolutionary history of system. This is called the biosemantic account. Grush uses the example of his feet, saying that by this account his feet would not be computing the amount of food he had eaten because their structure had not been evolutionarily selected for that purpose. Grush replies to the appeal to biosemantics with a thought experiment. Imagine that lightning strikes a swamp somewhere and creates an exact copy of you. According to the biosemantic account, this swamp-you would be incapable of computation because there is no evolutionary history with which to justify assigning representational content. The idea that for two physically identical structures one can be said to be computing while the other is not should be disturbing to any physicalist.

There are also syntactic or structural accounts for computation. These accounts do not need to rely on representation. However, it is possible to use both structure and representation as constrains on computational mapping. Shagrir identifies several philosophers of neuroscience who espouse structural accounts. According to him, Fodor and Pylyshyn require some sort of syntactic constraint on their theory of computation. This is consistent with their rejection of connectionist systems on the grounds of systematicity. He also identifies Piccinini as a structuralist quoting his 2008 paper: "the generation of output strings of digits from input strings of digits in accordance with a general rule that depends on the properties of the strings and (possibly) on the internal state of the system".^[56] Though Piccinini undoubtedly espouses structuralist views in that paper, he claims that mechanistic accounts of computation avoid reference to either syntax or representation.^[55] It is possible that Piccinini thinks that there are differences between syntactic and structural accounts of computation that Shagrir does not respect.

In his view of mechanistic computation, Piccinini asserts that functional mechanisms process vehicles in a manner sensitive to the differences between different portions of the vehicle, and thus can be said to generically compute. He claims that these vehicles are medium-independent, meaning that the mapping function will be the same regardless of the physical implementation. Computing systems can be differentiated based upon the vehicle structure and the mechanistic perspective can account for errors in computation.

Dynamical systems theory presents itself as an alternative to computational explanations of cognition. These theories are staunchly anti-computational and anti-representational. Dynamical systems are defined as systems that change over time in accordance with a mathematical equation. Dynamical systems theory claims that human cognition is a dynamical model in the same sense computationalists claim that the human mind is a computer.^[57] A common objection leveled at dynamical systems theory is that dynamical systems are computable and therefore a subset of computationalism. Van Gelder is quick to point out that there is a big difference between being a computer and being computable. Making the definition of computing wide enough to incorporate dynamical models would effectively embrace pancomputationalism.