Machine as Mind

Machine as Mind
Herbert A. Simon
인지과학 협동과정
99132-801
심소영
Contents
1. Introduction
2. Nearly-Decomposable Systems
3. The Two Faces of AI
4. The View from Psychology
5. The Matter of Semantics
6. “Ill-Structured” Phenomena
7. The Processing of Language
8. Affect, Motivation, and Awareness
9. Conclusion: Computers Think
-- and Often Think like People
1. Introduction
• I will proceed from what psychological
research has learned about human mind, to
the characteristics we must bestow upon
computer programs when we wish those
programs to think. (Section 4 ~ 8)
• By “mind”, I means a system that produces
thought, viewed at a relatively high level of
aggregation. (Section 2)
1. Introduction
• The level of aggregation at which we model
phenomena
– The primitive of mind are symbols, complex
structure of symbols, and processes that operate
on symbols (requiring at least tens of
milliseconds). At this level, the same software
can be implemented with different kinds of
hardware.
1. Introduction
• Central thesis
– At this level of aggregation, conventional
computer can be, and have been, programmed
to represent symbol structures and carry out
processes on those structures in a manner that
parallels the way the human brain does it.
• Principal evidence
– Programs that do just that
1. Introduction
• Computer simulation of thinking is no more
thinking than a simulation of digestion is
digestion.
– The analogy is false. The materials of digestion
are chemical substances, which are not
replicated in computer simulation., but the
materials of thought are symbols, which can be
replicated in a great variety of materials
(including neurons and chips).
2. Nearly-Decomposable Systems
• Most complex systems are hierarchical and
nearly decomposable.
– E.g. Building - Rooms - Cubicles
• Nearly-Decomposable Systems
– can be analyzed at a particular level of
aggregation without detailed knowledge of the
structures at the levels below. Only aggregate
properties of the more microscopic systems
affect behavior at the higher level.
2. Nearly-Decomposable Systems
• Because mind behaves as a nearlydecomposable system, we can model
thinking at the symbolic level, without
concern for details of implementation at the
hardware level.
3. The Two Faces of AI
• AI can be approached in two ways.
– First, we can write programs without any
commitment to imitating the processes of
human intelligence.
• E.g. DEEPTHOUGHT
– Alternatively, we can write programs that
imitate closely the human processes.
• E.g. MATER (Baylor and Simon 1966)
3. The Two Faces of AI
• Chess-playing programs illustrate the two
approaches.
– DEEPTHOUGHT does not play in a humanoid
way, typically exploring 107 of branches of the
game tree before it makes its choice of move.
DEEPTHOUGHT rests on a combination of
brute force, unattainable by human players, and
extensive, mediocre chess knowledge.
3. The Two Faces of AI
– Human grandmasters seldom look at more than
100 branches. By searching the relevant
branches, they make up with chess knowledge
for their inability to carry out massive searches.
– MATER uses heuristics, so it looks at fewer
than 100 branches.
• Because my aim here is to consider machine
as mind, the remainder of my remarks are
concerned with programs that are intelligent
in more or less humanoid ways.
4. The View from Psychology
• How does intelligence look to contemporary
cognitive psychology?
4.1 Selective Heuristic Search
• Human problem solvers do not carry out
extensive searches.
• People use knowledge about the structure of
the problem space to form heuristics that
allow them to search extremely selectively.
4. The View from Psychology
4.2 Recognition: The Indexed Memory
• The grandmaster’s memory is like a large
indexed encyclopedia.
• The perceptually noticeable features of the
chessboard (the cues) trigger the appropriate
index entries and give access to the
corresponding information.
4. The View from Psychology
• Solving problems by responding to cues
that are visible only to experts is called
solving them by “intuition”. (solving by
recognition)
• In computers, recognition processes are
implemented by productions: the condition
sides serve as tests for the presence of cues,
the action sides hold the information that is
accessed when the cues are noticed.
4. The View from Psychology
• Items that serve to index semantic memory
are called “chunks”. An expert in any
domain must acquire some 50,000 chunks.
• It takes at least 10 years of intensive
training for a person to acquire the
information required for world-class
performance in any domain of expertise.
4. The View from Psychology
4.3 Seriality: The Limits of Attention
• Problems that cannot be solved by
recognition require the application of
sustained attention. Attention is closely
associated with human short-term memory.
• The need for all inputs and outputs of
attention-demanding tasks to pass through
short-term memory essentially serializes the
thinking process. We can only think of one
thing at a time.
4. The View from Psychology
• Hence, whatever parallel processes may be
going on at lower (neural) levels, at the
symbolic level the human mind is
fundamentally a serial machine.
4. The View from Psychology
4.4 The Architecture of Expert Systems
• Human Experts
– Search is highly selective, the selectivity is
based on heuristics stored in memory.
– The information accessed can be processed
further by a serial symbol-processing system.
4. The View from Psychology
• The AI experts systems
– have fewer chunks than the human experts and
make up for the deficiency by doing more
computing than people do. The difference is
quantitative, not qualitative: Both depend
heavily upon recognition, supplemented by a
little capacity for reasoning (i.e., search)
5. The Matter of Semantics
• It is claimed that the thinking of computers
is purely syntactical, that is, computers do
not have intentions, and their symbols do
not have semantic referents.
• The argument is refuted by concrete
examples of computer programs that have
goals and that demonstrably understand the
meanings of their symbols.
5. The Matter of Semantics
– Computer-driven van program has the intention
of driving along the road and creates internal
symbols that denote landscape features,
interprets them, and uses the symbols to guide
its steering and speed-control mechanisms
– Chess-playing program forms internal
representation that denotes the chess position
and intends to beat its opponent.
5. The Matter of Semantics
• There is no mystery about semantics and
human intentions.
– “Semantic” means that there is a
correspondence, a relation of denotation,
between symbols inside the head and objects
outside and the two programs have goals.
• It may be objected that computer does not
“understand” the meaning of its symbols or
the semantic operations on them, or the
goals it adopts.
5. The Matter of Semantics
– The word “understand” has something to do
with consciousness of meanings and intentions.
But my evidence that you are conscious is no
better than my evidence that the road-driving
computers are conscious..
• Semantic meaning
– a correspondence between the symbol and the
thing it denotes.
• Intention
– a correspondence between the goal symbol and
behavior appropriate to achieving the goal.
5. The Matter of Semantics
• Searl’s Chinese Room parable
– proves not that computer programs cannot
understand Chinese, but only that the particular
program Searl described does not understand
Chinese.
– Had he described a program that could receive
inputs from a sensory system and emit the
symbol “cha” in the presence of tea, we would
have to admit that it understood a little chinese.
6. “Ill-Structured” Phenomena
• “Ill-structured” means
– that the task has ill-defined or multidimensional goals,
– that its frame of reference or representation is
not clear or obvious,
– that there are no clear-cut procedures for
generating search paths or evaluating them.
• Use of NL, learning, scientific discovery
• When a problem is ill-structured,
– a first step is to impose some kind of structure
that allows it to be represented at least
approximately.
6. “Ill-Structured” Phenomena
• What does psychology tell us about
problem representations?
6.1 Forms of Representation
• Propositional Representations
– Situations may be represented in word or in
logical or mathematical notations
– The processing will resemble logical reasoning
or proof.
6. “Ill-Structured” Phenomena
• Pictorial Representations
– Situations may be represented in diagrams or
pictures.
– With processes to move them through time or
to search through a succession of their states.
• Most psychological research on
representations assumes one of the
representations mentioned.
6. “Ill-Structured” Phenomena
6.2 Equivalence of Representations
• What consequences does the form of
representation have for cognition?
• Informational & Computational Equivalence
– Two representations are informationally
equivalent if either one is logically derivable
from the other. If all the information available
in the one is available in the other.
– Two representations are computationally
equivalent if all the information easily available
in the one is easily available in the other.
6. “Ill-Structured” Phenomena
• Information is easily available if it can be obtained
from the explicit information with a small amount of
computation. (small relative to the capacities of the
processor)
• E.g. Arabic and Roman numerals are
informationally equivalent, but not
computationally equivalent.
• E.g. Representation of the same problem
as a set of declarative propositions in
PROLOG, as a node-link diagram in LISP.
6. “Ill-Structured” Phenomena
6.3 Representations Used by People
• There is much evidence that people use
mental pictures to represent problems, but
there is little evidence that people use
propositions in predicate calculus.
– Even in problems with mathematical
formalisms, the processes resemble heuristic
search more than logical reasoning.
6. “Ill-Structured” Phenomena
– In algebra and physics, subjects typically
convert a problem from natural language into
diagrams and then into equations.
– Experiment with presentation(* and +) and a
sentence, “The star is above/below the plus”
• Whatever the form of representation, the
processing of information resembles
heuristic search rather than theorem proving
6. “Ill-Structured” Phenomena
6.4 Insight Problems (“Aha!” experiences)
• Problems that tend to be solved suddenly,
after a long period of fruitless struggle.
• “Insight” that lead to change in
representation and solution of the mutilated
checkerboard problem can be explained by
mechanisms of attention focusing.
6. “Ill-Structured” Phenomena
The representations people use (both
propositional and pictorial) can be
simulated by computers.
7. The Processing of Language
• Whatever the role it plays in thought,
natural language is the principal medium of
communication between people.
• Far more has been learned about the relation
between natural language and thinking from
computer programs that use language inputs
or outputs to perform concrete tasks.
7. The Processing of Language
7.1 Some Programs that Understand Language
• Novak’s ISMC program (1977)
– extracts the information from natural-language
descriptions of physics problems, and
transforms it into an internal “semantic”
representation suitable for a problem-solving
system.
7. The Processing of Language
• Hayes and Simon’s UNDERSTAND
program (1974)
– reads natural-language instructions for puzzles
and creates internal representations(“pictures”)
of the problem situations and interpretations of
the puzzle rules for operating on them.
• These programs give us specific models of
how people extract meaning from discourse
with semantic knowledge in memory.
7. The Processing of Language
7.2 Acquiring Language
• Siklossy’s program ZBIE (1972)
– was given (internal representations of) a simple
picture (a dog chasing a cat) and a sentence
describing the scene.
– With the aid of a carefully designed sequence
of such examples, it gradually learned to
associate nouns with the objects in the pictures
and other words with their properties and the
relations.
7. The Processing of Language
7.3 Will Our Knowledge of Language Scale?
• These illustrations involve relatively simple
language with a limited vocabulary.
• To demonstrate an understanding of human
thinking, we do not need to model thinking
in the most complex situations we can
imagine. Our theory explain the phenomena
in range of situations that would call for
genuine thinking in human.
7. The Processing of Language
7.4 Discovery and Creativity
• Making scientific discoveries is both illstructured and creative. These activities
have been simulated by computer.
• BACON program (Simon et al. 1987)
– When given the data available to the scientists
in historically important situations, it has
discovered Kepler’s Third Law, etc..
7. The Processing of Language
• KEKADA program (Simon et al. 1988)
– plans experimental strategies, responding to the
information gained from each experiment to
plan the next one.
– is able to track Faraday’s strategy.
• Programs like BACON and KEKADA show
that scientists use essentially the same kinds
of processes as those identified in more
prosaic kinds of problem solving.
8. Affect, Motivation, and
Awareness
• Motivation selects particular tasks for
attention and diverts attention from others.
• If affect and cognition interact largely
through the mechanisms of attention, then it
is reasonable to pursue our research on
these two components of mental behavior
independently.
• Many of the symbolic processes are in
conscious awareness, and awareness has
implications for the easy of testing.
9. Conclusion: Computers Think
and Often Think like People
• Computers can be programmed, and have
been programmed, to simulate at a symbolic
level the processes that are used in human
thinking.
• The human mind does not reach its goals
mysteriously or miraculously. Even its
sudden insights are explainable in terms of
recognition processes, well-informed search,
and changes in representation motivated by
shifts in attention.