15 Foundations and prospects Flashcards
he big questions
- What does it mean to think?
- Are machines able to think?
- What is intelligence?
- Can machines be intelligent?
- What does it mean to be conscious?
- Can machines be conscious?
- What is mind?
- Can machines have mind?
Weak vs. strong AI
- *Weak AI**: Machines can be made to act as if they are intelligent
- *Strong AI**: Machines can be made that are intelligent, have minds, and are conscious
The Turing Test
In an attempt to answer the question “Can machines think?”, Alan Turing (1950) proposed the Turing test for intelligence: The computer shall have a conversation with an interrogator for 5 minutes and have a 30% chance of fooling the interrogator into believing it is human. Turing believed that by year 2000, a computer with a storage of 109 units will pass the Turing test. So far, no computer has passed the test. Such a machine will qualify as weak AI (“as if intelligent”).
Objections to intelligent machines
Turing considered many objections to AI
• Argument from disability
• The mathematical objection
• The argument from informality
Argument from disability: A machine can never do X
• X = to be kind, friendly, make mistakes, have sense of humor, fall in love, do something really new, . . .
• Counter: Many such “impossibility claims” are unsupported, and some can be refuted
Mathematical objections to AI
An AI program is a formal system implemented on a computer, and subject to theoretical limits, e.g. the incompleteness theorem (Gödel): In any formal system powerful enough to do arithmetic, there are true statements that cannot be proved. Humans can overcome formal limits, e.g. by “meta-transfer” to other formalisms and are therefore inherently superior.
Counter-arguments:
- Computers are finite machines, and are strictly not subject to Gödel’s theorem
- Intelligent humans also suffer from inability to prove all true statements
- The brain is a deterministic physical device (some argue against this) and subject to the same formal limits as as computer
Informality objection to AI
Proposition (Dreyfus):
• Human behavior is too complex to be captured by a simple set of rules
• Since computers can only follow rules (can only do what the are told to), they cannot generate intelligent behavior on human level
This critique is directed towards simple first-order logic rule-based systems without learning, “GOFAI - Good Old Fashioned AI”. Modern AI includes other reasoning and learning methods:
• Generalization from examples
• Supervised, unsupervised and reinforcement learning • Learning with very large feature sets
• Directed sensing
Thus, AI makes progress to overcome the critique.
Strong AI - machine consciousness
Even if machines can be made to act as if they are intelligent (weak AI), “real” machine intelligence must have consciousness (strong AI). The machine must be aware of its own mental state and actions, be aware of its own beliefs, desires and intentions. Turing rejected this requirement, because we do not even know that other humans have consciousness, we can only observe their external behavior. Many will nevertheless require strong AI before they accept a machine as intelligent.
Can machines have mental states?
Functionalism answer: If the computer provides same answer to a problem as a human would (same function), it must have the same internal mental state.
Biological naturalism answer: Mental states are high-level and emergent features that are caused by neural activity in the brain that cannot be replicated by other means.
The mind-body problem
Ancient question:
How is mind (soul, consciousness) related to body (brain)?
Dualist view: Mind and body are fundamentally different categories of existence
Materialist view: “Brains cause minds” (Searle). I.e. the brain is the “hardware” for the mind “software”
Accepting the materialist view, can a machine have consciousness?
The Chinese room
Argument by Searle (1980):
Human (“CPU”) with no knowledge of Chinese operates in a closed room with a rulebook (“program”) and a stack of paper (“memory”).
Human receives slips of paper with (for him non-intelligible) Chinese text, follows rules mechanically and returns sensible replies in Chinese.
From the outside, it seems that the Chinese room behaves intelligently, yet the human has no idea of what he is responding to the inputs (just follows the rules).
This demonstrates that a system that passes Turing test need not be intelligent or conscious
The Systems reply
The Systems reply (McCarthy)
The Chinese room argument relies on following claims:
Certain kinds of objects are incapable of conscious understanding (in this case, Chinese)
The human, paper, and rule book are objects of this kind
If each of the objects is incapable of conscious understanding, then any system constructed from the objects is incapable of conscious understanding
Therefore there is no conscious understanding in the Chinese room
In the “Systems reply” to Searle (McCarthy and others), the third claim is not accepted. If it was true, how could
(conscious) humans be made of (unconscious) molecules?
onsciousness as emergent property
In more recent work, Searle claims that consciousness is an emergent property of properly arranged neurons, and only (biological) neurons. (Most) AI researchers agree that consciousness is an emergent property, but that the physical components underlying it can be neurons or electronic components or some other mechanism. Searle’s argument is not more founded on “facts” than the opposite (AI) argument.
Can the strong AI question be settled?
Consciousness is not a well defined or well understood phenomena. We do not know what kind of experiment can be used to determine consciousness in a computer. Question could be settled if we discovered how consciousness can be reduced to other phenomena. As no such reduction is known, the strong AI question will remain open.
Tentative answers to some “big questions”
- *Weak AI** (machines can be made that act as if they are intelligent).
- Many AI programs do in fact exhibit “intelligence”.
- Arguments against weak AI are needlessly pessimistic.
- *Strong AI** (machines can be made that are intelligent and conscious).
- Difficult to prove either impossibility or possibility of this claim.
- The answer is not important for further progress for (weak) AI.
Recapitulation: AI as agent design
The AI “project” can be seen as the design of intelligent agents. Different agent designs are possible, from reflex agents to deliberative knowledge-based ones. Different paradigms are being used: logical, probabilistic, “neural”. Do we have the necessary tools to build a complete, general-purpose agent?
Model- and utility-based agent
State-of-the-art
Interaction with the environment
• Improved greatly in recent years: cameras, MEMS, . . .
• Dominant new environment: the Internet
Keeping track of environment’s state:
• Perception and updating of internal representation
• Filtering methods for tracking uncertain environments
• Mostly low-level and propositional
• Need to improve ability to recognize higher-level objects, relations, scenes, etc.
Evaluate and select actions:
• Simple methods for planning and deciding exist
• Real-world complexity require strong abstraction ability (hierarchies) • Great deal of development is needed
Utility as expression of preference:
• MEU is sound in principle, but depends on realistic utility functions • Need to extract utility information from humans to guide agents
Learning capabilities
• Basic learning technology has progressed rapidly in recent years, sometimes with abilities that exceed human learning ability
However, little progress on how to learn higher level concepts from lower level (input) concepts
• Without such generalization ability, agents must be taught manually by humans
Hybrid agent architecture
Ability to incorporate different types of reasoning and decision making (from reflex to deliberation). Learning from experience (compiling).
