Transcript Could computers understand language?

Could computers
understand language?
Phenomenological (and other) arguments
against Artificial Intelligence
Staffan Larsson
Dept. of linguistics
Göteborg University
Overview
•
•
•
•
•
Artificial Intelligence
Phenomenological arguments
The BaB objection
Arguments related to evolution
Conclusion
Artificial Intelligence
The Man Machine?
• Descartes (1596-1650)
– Animals are to be considered as machines, automata
– Man has a soul and is therefore fundamentally
different from machines
• de la Mettrie (1709-1751)
– ”The man machine”
– The difference between human and machine is a
qualitative difference of complexity
– No fundamental difference!
• These questions are intimately related:
– Can machines think?
– Are humans machines?
The Turing test
• Can a machine be intelligent? Is ”artificial intelligence”
(AI) possible?
• Turing offers an operational definition of intelligence
• Turing (1912-1954): ”the Turing test”
– Test person A has a dialogue (via a text terminal) with B.
– A:s goal is to decide whether B is a human or a machine
– If B is a machine and manages to deceive A that B is a human, B
should be regarded as intelligent (able to think; ”a grade A
machine”)
– (This is a simplified version of the Turing test)
The Turing test and dialogue
• According to the Turing test – what is
fundamentally human?
– The ability to carry out a dialogue using natural
language
• Why is this fundamental?
– Assumption: In dialogue, all other human capabilities
show themselves (directly or indirectly)
• This means that ...
– ... in order to make a computer use natural language
in the same way and on the same level as a human, it
needs to be endowed with human-level intelligence
Artificial Intelligence
• Goal
– simulate human/intelligent behaviour/thinking
• Weak AI
– Machines can be made to act as if they were
intelligent
• Strong AI
– Agents that act intelligently have real, conscious
minds
• It is possible to believe in strong AI but not in
weak AI
A short history of AI
Cognitivism and GOFAI
• Descartes again:
– Understanding and thinking is forming and using symbolic
representations
• Until the mid-80’s, the paradigm of AI was cognitivism,
the idea that thinking is, essentially, symbol manipulation
• The physical symbol hypothesis (Newell & Simon):
– ”A physical symbol system has the necessary and sufficient
means for intelligent action.”
– All intelligent behaviour can be captured by a system that
reasons logically from a set of facts and rules that describe the
domain
• This is sometimes referred to as GOFAI (Good Old
Fashioned AI)
Dialogue systems as GOFAI?
• Since around 1986, GOFAI has been abandoned by many AI
researchers
– Instead, focus on connectionism, embodied interactive automata,
reinforcment learning, probabilistic methods, etc.
• However, a large part of current dialogue systems research adheres
to the GOFAI paradigm
– Information States, for example…
• Why?
– It seems to be the most workable method for the complex problems of
natural language dialogue
– It appears to be useful for improving on current human-computer
interfaces, although a major breakthrough of NL interfaces is needed to
prove this conclusively
• But is it also a step on the way towards ”human-level” natural
language understanding in computers?
– Does it scale up?
Phenomenological arguments
Some arguments against weak AI
• Ada Lovelace’s objection
– computers can only do what we tell them to
• Argument from disability
– claims (usually unsupported) of the form ”a machine can never
do X”
• The mathematical objection
– based on Gödel’s incompleteness theorem
• The argument from informality of behaviour
• (Searle’s Chinese Room
– argument concerns strong AI
– purports to show that producing intelligent behavoiur is not a
sufficient condition for being a mind)
Some problems in AI
• Computational complexity in real-time resource-bounded
applications
– Reasoning
– Planning for conjunctive goals
– Plan recognition
• Incompleteness of general FOL reasoning
– not to mention modal logic
• Frame problem
– updating the “world model”
• Endowing a computer with the common sense of a 4-year-old
– AI is still very far from this
• Humans don’t have problems with these things
• Is is possible that all these problems have a common cause?
– They all seem to be related to representations and symbol manipulation
The argument from informality of behaviour
(Dreyfus, Winograd, Weizenbaum)
• Human behaviour based on our everyday
commonsense background understanding /
holistic context
– allows us to experience what is currently relevant,
and deal with tings and people
– crucial to understanding language
– involves utterance situation, activity, institution,
cultural setting
• In its widest sense, the background involves all
of human culture and experience
• Dreyfus argues that the background has the
form of dispositions, or informal know-how
– Normally, ”one simply knows what to do”
– a form of skill rather than propositional knowing-that
– inarticulate, to some extent pre-conceptual
• To achieve GOFAI,
– this know-how, along with interests, feelings,
motivations, and bodily capacities that go to make a
human being,
– would have to be conveyed to the computer as
knowledge in the form of a huge and complex belief
system
• ”The background cannot be formalised”
– There are no reasons to think that humans represent
and manipulate the background explicitly
• Human behaviour is far too complex to be
described by a set of formal rules of the kind that
can be followed by computers
– (”unwritten rules” are not just unwritten, but unwritable; any written version will leave something out
and overspecify something else)
– (This applies mainly to Von Neumann computers
programmed in the standard way; we’ll get to artificial
neural networks later)
Problems with formalising
commonsense background
• How is everyday knowledge organized so
that one can make inferences from it?
– Ontological engineering: finding the primitive
elements in which the ontology bottoms out
• How can skills or know-how be
represented as knowing-that?
• How can relevant knowledge be brought to
bear in particular situations?
CYC (Lenat) and natural language
• Formalise common sense
– The kind of knowledge we need to understand NL
– using general categories that make no reference to
specific uses of the knowledge (context free)
• Lenat’s ambitions:
– it’s premature to try to give a computer skills and
feelings required for actually coping with things and
people
– L. is satisfied if CYC can understand books and
articles and answer questions about them
CYC vs. NL
• Example (Lenat)
– ”Mary saw a dog in the window. She wanted it.”
• Dreyfus:
– this sentence seems to appeal to
• our ability to imagine how we would feel in the situation
• know-how for getting around in the world (e.g. getting closer to
something on the other side of a barrier)
– rather than requiring us to consult facts about dogs and windows
and normal human reactions
• So feelings and coping skills that were excluded to
simplify the problem return
– We shouldn’t be surprised; this is the presupposition behind the
Turing Test – that understanding human language cannot be
isolated from other human capabilities
CYC vs. NL
• How can relevant knowledge be brought to bear in
particular situations?
– categorize the situation
– search through all facts, following rules to find the facts possibly
relevant in this situation
– deduce which facts are actually relevant
• How deal with complexity?
– Lenat: add meta-knowledge
• Dreyfus:
– meta-knowledge just makes things worse; more meaningless
facts
– CYC is based on an untested traditional assumption that people
store context-free facts and use meta-rules to cut down the
search space
Everyday skills vs. rules
• Dreyfus suggests testing this assumption
– by looking at the phenomenology of everyday know-how
– Heidegger, Merleau-Ponty, Pierre Bourdieu
• What counts as facts depends on our skills; e.g. giftgiving (Bourdieu)
– If it is not to constitute an insult, the counter-gift must be deferred
and different, because the immediate return of an exact identical
object clearly amounts to a refusal....
– It is all a question of style, which means in this case timing and
choice of occasion...
– ...the same act – giving, giving in return, offering one’s services,
etc. – can have completely different meanings at different times.
Everyday skills vs. rules
• Having acquired the necessary social skill,
– one does not need to recognize the situation as appropriate for
gift-giving, and decide rationally what gift to give
– ”one simply responds in the appropriate circumstances by giving
an appropriate gift”
• Humans can
– skilfully cope with changing events and motivations
– project understanding onto new situations
– understand social innovations
• one can do something that has not so far counted as
appropriate...
• ...and have it recognized in retrospect as having been just
the right thing to do
Everyday skills vs. rules
• When things go wrong - when we fail –
there is a breakdown
– In such situations, we need to reflect and
reason, and may have to learn and apply
formal rules
• but it is a mistake to
– read these rules back into the normal situation
and
– appeal to such rules for a causal explanation
of skilful behaviour
Analogy and metaphor
• ... pervade language (example from Lenat):
– ”Texaco lost a major ruling in its legal battle with
Pennzoil. The supreme court dismantled Texaco’s
protection against having to post a crippling $12
billion appeals bond, pushing Texaco to the brink of a
Chapter 11 filing” (Wall Street Journal)
• The example drives home the point that,
– far from overinflating the need for real-world
knowledge in language understanding,
– the usual arguments about disambiguation barely
scratch the surface
Analogy and metaphor
• ... pervade language (example from Lenat):
– ”Texaco lost a major ruling in its legal battle with
Pennzoil. The supreme court dismantled Texaco’s
protection against having to post a crippling $12
billion appeals bond, pushing Texaco to the brink of a
Chapter 11 filing” (Wall Street Journal)
• The example drives home the point that,
– far from overinflating the need for real-world
knowledge in language understanding,
– the usual arguments about disambiguation barely
scratch the surface
Analogy and metaphor
• Dealing with metaphors is a nonrepresentational mental capacity (Searle)
– ”Sally is a block of ice” could not be analyzed
by listing the features that Sally and ice have
in common
• Metaphors function by association
– We have to learn from vast experience how to
respond to thousands of typical cases
Background and NL
• NL interpretation problems
– Analogy, metaphor
– It is notoriously hard to exactly pin down implicatures and
presuppositions of natural language utterances
• It appears that full It appears that full disambiguation and
understanding of natural language requires access to this
background knowledge
– Still, this does not normally cause us any problems
• Most (if not all) dialogue systems assume that context can be
formalised
– This is perhaps plausible for ”micro-worlds”, i.e. highly limited and
systematic domains, such as train timetables or programming a VCR
– So we can still do practically useful AI work for limited domains
Counter-argument 1
• Phenomenology studies how things
appear to us, not how they are
– The fact that we are often not aware of
reasoning simply indicates that a lot of
reasoning is non-conscious
Dreyfus’ account of skill acquisition
• 5 stages
– 1. Beginner student: Rule-based processing;
• learning and applying rules for manipulating context-free elements
• There is thus a grain of truth in GOFAI
– 2. Understanding the domain; seeing meaningful aspects, rather than
context-free features
– 3. Setting goals and looking at the current situation in terms of what is
relevant
– 4. Seeing a situation as having a certain significance toward a certain
outcome
– 5. Expert: The ability of instantaneously selecting correct responses
(dispositions)
• (Note: this is how adults typically learn; Infants, on the other hand...
– learn by imitation
– ”pick up on a style” that pervades his/her society)
Response to counter-argument 1
• There is no reason to suppose that the
beginner’s features and rules (or any
features and rules) play any role in expert
performance
– That we once followed a rule in tying our
shoelaces does not mean we are still
following the same rule unconsciously
– ”Since we needed training wheels when
learning how to ride a bike, we must now be
using invisible training wheels.”
Counter-argument 2
• This argument only applies to GOFAI!
• A lot of modern AI is not GOFAI
– interactionist AI (Brooks, Chapman, Agre)
– connectionism / neural networks
– reinforcement learning
Interactionist AI
• No need for a representation of the world
– instead, look to the world as we experience it
• Behaviour can be purposive without the agent having in mind a goal
or purpose
– In many situations, it is obvious what needs to be done
– Once you’ve done that, the next thing is likely to be obvious too
– Complex series of actions result, without the need for complex decisions
or planning
• However, Interactionist AI does not address problem of informal
background familiarity
– programmers have to predigest the domain and decide what is relevant
– systems lack ability to discriminate relevant distinctions in the skill
domain...
– ... and learn new distinctions from experience
Connectionism
• Apparently does not require being given a theory
of a domain in order to behave intelligently
– Finding a theory = finding invariant features in terms
of which situations can be mapped onto responses
• Starting with random weights, will neural nets
trained on same date pick out the same
invariants?
– No; it appears the ”tabula rasa” assumption (random
initial weights) is wrong
Learning & generalisation
• Learning depends on the ability to
generalise
• Good generalisation cannot be achieved
without a good deal of background
knowledge
• Example: trees/hidden tanks
• A network must share our commonsense
understanding ot the world if it is to share
our sense of appropriate generalisation
Reinforcement learning
• Idea: learn from interacting with the world
– Feed back reinforcement signal measuring the
immediate cost or benefit of an action
– Enables unsupervised learning
– (The ”target representation” in humans is neural
networks)
• Dreyfus: To build human intelligence, need to
improve this method
– assigning fairly accurate actions to novel situations
– reinforcement-learning device must ”exhibit global
sensitivity by encountering situations under a
perspective and actively seeking relevant input”
(An aside: human reinforcement)
• Currently, programmer must supply machine
with rule formulating what to feed back as
reinforcement
• What is the reinforcement signal for humans?
– Survival?
– Pleasure vs. pain?
• Requires having needs, desires, emotions
• Which in turn may depend on the abilities and
vulnerabilities of a biological body
Progress?
• Unfortunately, all current learning techniques rely on the
previous availability of explicitly represented knowledge
– but as we have seen, Dreyfus argues that commonsense
background cannot be captured in explicit representations
• Russel & Norvig, in Artificial Intelligence -A Modern
Approach (1999)
– ”In our view, this is a good reason for a serious redesign of
current models of neural processing so that they can take
advantage of previously learned knowledge. There has been
some progress in this direction.”
– But no such research is cited
• So R & N admit that this is a real problem. In fact it is still
the exact same problem that Dreyfus pointed out
originally
– There is still nothing to indicate that Dreyfus is wrong when
arguing against the possibility of getting computers to learn
commonsense background knowledge
The BaB objection
The argument from infant
development (Weizenbaum)
• (Based on writings by child psychologist Erik Erikson)
• The essence of human being depends crucially on the fact that
humans are born of a mother, are raised by a mother and father, and
have a human body
– ”Every organism is socialized by dealing with problems that confront it”
(Weizenbaum)
– For humans, the problems include breaking the symbiosis with the
mother after the infant period
– This is fundamental to the human constitution; it lays the ground for all
future dealings with other people
• Men and machines have radically different constitutions and origins
– Humans are born by a mother and father
– Machines are built by humans
• OK, so we need to give AI systems a human or human-like body,
and let human parents raise them
The argument from language as
social comittment (Winograd)
• The essence of human communication is commitment,
an essentially social and moral attitude
• Speech acts work by imposing commitments on speaker
and hearer
• If one cannot be held (morally) responsible for one’s
actions, one cannot enter into commitments
• Computers are not human
– so they cannot be held morally responsible
– therefore, they cannot enter into commitments
• Therefore, machines can never be made to truly and
fully understand language
• OK, so we need to treat these AI computers exactly as
humans, and hold them morally responsible
The argument from ”human being”/Dasein
(Heidegger, Dreyfus)
•
Heidegger’s project in Being and Time
– Develop an ontology for describing human being
– What it’s like to be human
•
This can, according to Heidegger, only be understood ”from the inside”
– H:s text is not intended to be understandable by anyone who is not a human
•
Such an explanation is not possible, according to H.; human being cannot
be understood ”from scratch”
– Yet it is exactly such an explanation that is the goal of AI
•
According to Heidegger/Dreyfus, AI is impossible because (among other
things)
– Infants are, strictly speaking, not yet fully human; they must first be socialised
into a society and a social world
– Only humans so socialized can fully understand other humans
– Since cultures are different, humans socialized into one culture may have
problems understanding humans from another culture
– Machines are not socialised, they are programmed by humans
•
OK, so we need to socialise AI systems into society!
This leads to...
The ”Build-a-Baby” (BaB) objection
• So, we can do real AI, provided we can build
robot infants that are raised by parents and
socialised into society by human beings who
treat them as equals
– This probably requires people to actually think that
these AI systems are human
– These systems will have the same ethical status as
humans
• If this is line of ”research” to be pursued, it raises
some serious ethical problems
– C.f. movies ”A.I.” and ”Bladerunner”
• Is the goal of AI to build more humans?
Arguments related to evolution
The ”humans are animals”
argument
• What reason do we have to think that nonconscious reasoning operates by formal
reasoning?
• Humans have evolved from animals, so
presumably some non-formal thinking is
still part of the human mind
– Hard to tell a priori how much
The argument from the role of
emotions
• Classical AI deals first with rationality
• Possibly, we might want to add emotions as an
additional layer of complexity
• However, it seems plausible to assume that emotions
are more basic than rationality (Damasio: The Feeling of
what happens)
– Animals have emotions but not abstract rational reasoning
– The human infant is emotional but not rational
• So machines should be emotional before they are made
rational
– unfortunately, no-one has a clue how to make machines
emotional
The argument from brain matter
and evolution
• Weak AI assumes that physical-level simulation
is unnecessary for intelligence
• However, evolution has a reputation for finding
and exploiting available shortcuts
– works by ”patching” on previous mechanisms
• If there are any unique properties of biological
brain-matter that offers some possible
improvement to cognition, it is likely they have
been exploited
• If so, it is not clear if these properties can be
emulated by silicon-based computers
The argument from giving a damn
• Humans care; machines don’t give a damn (Haugeland)
• Caring (about surviving, for example) comes from
instincts (drives) which animals, but not machines, have
• Caring about things is intimately related to the evolution
of living organisms
– Having a biological body
• So, can evolution be simulated?
– Winograd argues that the only simulation that would do the job
would need to be as complex as real evolution
– So in 3,5 billion years, we can have AI!
Conclusion
Summary
• Dreyfus et al has not proved conclusively that
weak AI is impossible (which Dreyfus admits)
• However, they point out that
– there are no reasons to believe it is possible,
– there are reasons for believing it is not possible
• Specific problems are pointed out
– indicating what we may need to do if we want to
achieve weak AI
• Provides some interesting ideas which allow us
to think not only about AI...
– ... but about human intelligence and behaviour