Strong AI is artificial intelligence that matches or exceeds human intelligence—the intelligence of a machine that can successfully perform any intellectual task that a human being can. It is a primary goal of artificial intelligence research and an important topic for science fiction writers and futurists. Strong AI is also referred to as "artificial general intelligence" or as the ability to perform "general intelligent action". Science fiction, associates strong AI with such human traits as consciousness, sentience, sapience and self-awareness.
Some references emphasize a distinction between strong AI and "applied AI" (also called "narrow AI" or "weak AI"): the use of software to study or accomplish specific problem solving or reasoning tasks that do not encompass (or in some cases are completely outside of) the full range of human cognitive abilities.
Requirements of strong AI
Many different definitions of intelligence have been proposed (such as being able to pass the Turing test) but there is to date no definition that satisfies everyone. However, there is wide agreement among artificial intelligence researchers that all of the following traits require intelligence:
Work is underway to design machines that have these abilities and it is expected that strong AI would have most if not all of these capabilities.
There are other aspects of the human mind besides intelligence that also bear on the concept of strong AI:
None of these are necessary for strong AI—for example, it is not clear if consciousness is necessary for a machine to reason as well as human beings can. It is also not clear whether any of these traits are sufficient for intelligence: if a machine is built with a device that simulates the neural correlates of consciousness, would it automatically have the ability to represent knowledge or use natural language? It is also possible that some of these properties, such as sentience, naturally emerge from a fully intelligent machine, or that it becomes natural to ascribe these properties to machines once they begin to act in a way that is clearly intelligent. For example, intelligent action may be sufficient for sentience, rather than the other way around.
Research approaches
History of mainstream AI research
Modern AI research began in the middle 50s. The first generation of AI researchers were convinced that strong AI was possible and that it would exist in just a few decades. As AI pioneer Herbert Simon wrote in 1965: "machines will be capable, within twenty years, of doing any work a man can do." Their predictions were the inspiration for Stanley Kubrick and Arthur C. Clarke's character HAL 9000, who accurately embodied what AI researchers believed they could create by the year 2001. Of note is the fact that AI pioneer Marvin Minsky was a consultant on the project of making HAL 9000 as realistic as possible according to the consensus predictions of the time, having himself said on the subject in 1967, "Within a generation...the problem of creating 'artificial intelligence' will substantially be solved."
However, in the early 70s, it became obvious that researchers had grossly underestimated the difficulty of the project. The agencies that funded AI became skeptical of strong AI and put researchers under increasing pressure to produce useful technology, or "applied AI". As the eighties began, Japan's fifth generation computer project revived interest in strong AI, setting out a ten year timeline that included strong AI goals like "carry on a casual conversation". In response to this and the success of expert systems, both industry and government pumped money back into the field. However, the market for AI spectacularly collapsed in the late 80s and the goals of the fifth generation computer project were never fulfilled. For the second time in 20 years, AI researchers who had predicted the imminent arrival of strong AI had been shown to be fundamentally mistaken about what they could accomplish.
By the 1990s, AI researchers had gained a reputation for making promises they could not keep. AI researchers became reluctant to make any kind of prediction at all and avoid any mention of "human level" artificial intelligence, for fear of being labeled a "wild-eyed dreamer."
Mainstream AI research
For the most part, researchers today choose to focus on specific sub-problems where they can produce verifiable results and commercial applications, such as neural nets, computer vision or data mining.
Most mainstream AI researchers hope that strong AI can be developed by combining the programs that solve various subproblems using an integrated agent architecture, cognitive architecture or subsumption architecture. Hans Moravec wrote in 1988 "I am confident that this bottom-up route to artificial intelligence will one day meet the traditional top-down route more than half way, ready to provide the real world competence and the commonsense knowledge that has been so frustratingly elusive in reasoning programs. Fully intelligent machines will result when the metaphorical golden spike is driven uniting the two efforts."
Artificial general intelligence
Artificial General Intelligence research aims to create AI that can replicate human-level intelligence completely, often called an Artificial General Intelligence (AGI) to distinguish from less ambitious AI projects. (The concept is derived from the psychometric notion of natural general intelligence (often denoted "g"), though no adherence to any particular theory of g is implied.) As yet, researchers have devoted little attention to AGI, with some claiming that intelligence is too complex to be completely replicated in the near term. Some small groups of computer scientists are doing AGI research, however. Organizations pursuing AGI include the Adaptive AI, Artificial General Intelligence Research Institute (AGIRI), Singularity Institute for Artificial Intelligence and Texai. One recent addition is Numenta, a project based on the theories of Jeff Hawkins, the creator of the Palm Pilot. While Numenta takes a computational approach to general intelligence, Hawkins is also the founder of the RedWood Neuroscience Institute, which explores conscious thought from a biological perspective. AND Corporation has been active in this field since 1990, and has developed machine intelligence processes based on phase coherence principles, having strong similarities to digital holography and QM with respect to quantum collapse of the wave function.
Simulated human brain model
Simulated human brain model could be one of the quickest means of achieving strong AI, as it doesn't require complete understanding of how intelligence works. Basically, a very powerful computer would simulate a human brain, often in the form of a network of neurons. For example, given a map of all (or most) of the neurons in a functional human brain, and a good understanding of how a single neuron works, it is theoretically possible for a computer program to simulate the working brain over time. Given some method of communication, this simulated brain might then be shown to be fully intelligent. The exact form of the simulation varies: instead of neurons, a simulation might use groups of neurons, or alternatively, individual molecules might be simulated. It's also unclear which portions of the human brain would need to be modeled: humans can still function while missing portions of their brains, and areas of the brain are associated with activities (such as breathing) that might not be necessary to think.
Speculation: human brains have developed to accommodate certain necessities, such as breathing and interpreting sensory input from a variety of sources. Without adequate simulations of these necessities (such as, for example, input that simulates the sensation of sufficient oxygen levels in the body), it is possible that an artificial brain could have difficulty functioning. In addition, human brains are reliant for stability on a number of mediating factors, including stages of development and external training. An artificial duplicate of the human brain, without input of mediation, could conceivably suffer from a number of cognitive and functional difficulties. In addition, the construction and sustenance of an artificial brain raises moral questions, namely regarding personhood, freedom, and death. Does a "brain in a box" constitute a person? What rights would such an entity have, under law or otherwise? Once activated, would human beings have the obligation to continue its operation? Would the shutdown of an artificial brain constitute death, sleep, unconsciousness, or some other state for which no human description exists? After all, an artificial brain is not subject to post-mortem cellular decay (and associated loss of function) as human brains are, so an artificial brain could, theoretically, resume functionality exactly as it would if it was before it was shut down?
This approach would require three things:
- Hardware. An extremely powerful computer would be required for such a model. Futurist Ray Kurzweil estimates 10 million MIPS, or ten petaflops. At least one special-purpose petaflops computer has already been built (the Riken MDGRAPE-3) and there are nine current computing projects (such as Blue Gene/P) to build more general purpose petaflops computers all of which should be completed by 2008, if not sooner. Most other attempted estimates of the brain's computational power equivalent have been rather higher, ranging from 100 million MIPS (100 petaflops) to 100 billion MIPS (100,000 petaflops). Using accurate Top500 projections, it might be estimated that such levels of computing power might be reached using the top-performing CPU-based supercomputers to be by ~2015 (for 100 petaflops), up to a more conservative estimate of ~2025 (for 100,000 petaflops). However, considering that GPU processing and Stream Processing power appears to double every year, these estimates will be reached much sooner using GPGPU processing as high-end GPU's set to arrive in early 2008 are already going to be able to process over 1 teraflop, which is 20x more powerful than a standard quad-core CPU. It should also be noted, however, that the overhead introduced by the modeling of the biological, chemical, and physical details of neural behaviour (especially on a molecular scale) might require a simulator to have access to computational power much greater than that of the brain itself and that current simulations and estimates do not account for the importance of Glial cells which outnumber neurons 10:1.
- Software. Software to simulate the function of a brain would be required. This assumes that the human mind is the central nervous system and is governed by physical laws. Constructing the simulation would require a great deal of knowledge about the physical and functional operation of the human brain, and might require detailed information about a particular human brain's structure. Information would be required both of the function of different types of neurons, and of how they are connected. Note that the particular form of the software dictates the hardware necessary to run it. For example, an extremely detailed simulation including molecules or small groups of molecules would require enormously more processing power than a simulation that models neurons using a simple equation, and a more accurate model of a neuron would be expected to be much more expensive computationally than a simple model. The more neurons in the simulation, the more processing power it would require.
- Understanding. Finally, it requires sufficient understanding thereof to be able to model it mathematically. This could be done either by understanding the central nervous system, or by mapping and copying it. Neuroimaging technologies are improving rapidly, and Kurzweil predicts that a map of sufficient quality will become available on a similar timescale to the required computing power. However, the simulation would also have to capture the detailed cellular behaviour of neurons and glial cells, presently only understood in the broadest of outlines.
Once such a model is built, it will be easily altered and thus open to trial-and-error experimentation. This is likely to lead to huge advances in understanding, allowing the model's intelligence to be improved/motivations altered.
The Blue Brain project aims to use one of the fastest supercomputer architectures in the world, IBM's Blue Gene platform, to simulate a single neocortical column consisting of approximately 60,000 neurons and 5km of interconnecting synapses. The eventual goal of the project is to use supercomputers to replicate an entire brain.
The brain gets its power from performing many parallel operations, a standard computer from performing operations very quickly. It should be noted, however, that supercomputers also perform many operations in parallel. Good examples of this are the Cray and NEC vector computers which operate as a single machine but perform many calculations at once. Another example is any form of cluster computing, where multiple single computers operate as one.
The human brain has roughly 100 billion neurons operating simultaneously, connected by roughly 100 trillion synapses. By comparison, a modern computer microprocessor uses only 1.7 billion transistors. Although estimates of the brain's processing power put it at around 1014 neuron updates per second, it is expected that the first unoptimized simulations of a human brain in real time will require a computer capable of 1018 FLOPS. Non-real time simulations of the human brain (1011 neurons) were performed in 2005 and it took 50 days on a cluster of 27 processors to simulate 1 second of a model. By comparison a general purpose CPU (circa 2006) operates at a few GFLOPS (109 FLOPS). (each FLOP may require as many as 20,000 logic operations).
However, a neuron is estimated to spike 200 times per second (this giving an upper limit on the number of operations). Signals between them are transmitted at a maximum speed of 150 meters per second. A modern 2GHz processor operates at 2 billion cycles per second, or 10,000,000 times faster than a human neuron, and signals in electronic computers travel at roughly half the speed of light; faster than signals in humans by a factor of 1,000,000. The brain consumes about 20W of power whereas supercomputers may use as much as 1MW or an order of 100,000 more (note: Landauer limit is 3.5x1020 op/sec/watt at room temperature).
Neuro-silicon interfaces have also been proposed.
Critics of this approach believe that it is possible to achieve AI directly without imitating nature and often have used the analogy that early attempts to construct flying machines modeled them after birds, but that modern aircraft do not look like birds. The direct approach is used in AI - What is this, where it is shown that if we have a formal definition of AI, it can be found by enumerating all possible programs and then testing each of them to see whether it has produced Artificial Intelligence, or has not.
Artificial consciousness research
Artificial consciousness research aims to create and study artificially conscious systems. Igor Aleksander argues that the principles for creating a conscious machine already existed but that it would take forty years to train such a machine to understand language.
Franklin’s Intelligent Distribution Agent
Stan Franklin (1995, 2003) defines an autonomous agent as possessing functional consciousness when it is capable of several of the functions of consciousness as identified by Bernard Baars’ Global Workspace Theory (GWT). His brain child IDA (Intelligent Distribution Agent) is a software implementation of GWT, which makes it functionally conscious by definition. IDA’s task is to negotiate new assignments for sailors in the US Navy after they end a tour of duty, by matching each individual’s skills and preferences with the Navy’s needs. IDA interacts with Navy databases and communicates with the sailors via natural language email dialog while obeying a large set of Navy policies. The IDA computational model was developed during 1996-2001 at Stan Franklin’s "Conscious" Software Research Group at the University of Memphis. It "consists of approximately a quarter-million lines of Java code, and almost completely consumes the resources of a 2001 high-end workstation." It relies heavily on codelets, which are "special purpose, relatively independent, mini-agent[s] typically implemented as a small piece of code running as a separate thread." In IDA’s top-down architecture, high-level cognitive functions are explicitly modeled; see Franklin (1995) and Franklin (2003) for details. While IDA is functionally conscious by definition, Franklin does “not attribute phenomenal consciousness to [his] own 'conscious' software agent, IDA, in spite of her many human-like behaviours. This in spite of watching several US Navy detailers repeatedly nodding their heads saying 'Yes, that’s how I do it' while watching IDA’s internal and external actions as she performs her task."
Haikonen’s cognitive architecture
Pentti Haikonen (2003) considers classical rule-based computing inadequate for achieving AC: "the brain is definitely not a computer. Thinking is not an execution of programmed strings of commands. The brain is not a numerical calculator either. We do not think by numbers." Rather than trying to achieve mind and consciousness by identifying and implementing their underlying computational rules, Haikonen proposes "a special cognitive architecture to reproduce the processes of perception, inner imagery, inner speech, pain, pleasure, emotions and the cognitive functions behind these. This bottom-up architecture would produce higher-level functions by the power of the elementary processing units, the artificial neurons, without algorithms or programs". Haikonen believes that, when implemented with sufficient complexity, this architecture will develop consciousness, which he considers to be "a style and way of operation, characterized by distributed signal representation, perception process, cross-modality reporting and availability for retrospection." Haikonen is not alone in this process view of consciousness, or the view that AC will spontaneously emerge in autonomous agents that have a suitable neuro-inspired architecture of complexity; these are shared by many, e.g. Freeman (1999) and Cotterill (2003). A low-complexity implementation of the architecture proposed by Haikonen (2004) was reportedly not capable of AC, but did exhibit emotions as expected.
Emergence
Some have suggested that intelligence can arise as an emergent quality from the convergence of random, man-made technologies. Human sentience — or any other biological and naturally occurring intelligence — arises out of the natural process of species evolution and an individual's experiences.
Origin of the term: John Searle's strong AI
The term "strong AI" was adopted from the name of an argument in the philosophy of artificial intelligence first identified by John Searle as part of his Chinese room argument in 1980. He wanted to distinguish between two different hypotheses about artificial intelligence:
- An artificial intelligence system can think and have a mind.
- An artificial intelligence system can (only) act like it thinks and has a mind.
The first one is called "the strong AI hypothesis" and the second is "the weak AI hypothesis" because the first one makes the stronger statement: it assumes something special has happened to the machine that goes beyond all its abilities that we can test. Searle referred to the "strong AI hypothesis" as "strong AI". This usage, which is fundamentally different than the subject of this article, is common in academic AI research and textbooks.
The term "strong AI" is now used to describe any artificial intelligence system that acts like it has a mind, regardless of whether a philosopher would be able to determine if it actually has a mind or not. Dijkstra has been quoted as saying, "The question of whether a computer can think is no more interesting than the question of whether a submarine can swim." As Russell and Norvig write: "Most AI researchers take the weak AI hypothesis for granted, and don't care about the strong AI hypothesis." AI researchers are interested in a related statement (that some sources confusingly call "the strong AI hypothesis"):
- An artificial intelligence system can think (or act like it thinks) as well or better than people do.
This assertion, which hinges on the breadth and power of machine intelligence, is the subject of this article.
No comments:
Post a Comment