Transcript ppt - Department of Computer Science and Engineering

INTERNAL REPRESENTATIONS:
An approach to AI methodology.
Course Seminar
By
Sumesh.M. K.
(05408008)
Course: CS621,
Offered by Prof. Pushpak Bhattacharyya, IITB.
INTERNAL REPRESENTATIONS:
AN APPROACH TO AI
1
Abstract: The theory of internal representations which is at the heart of
cognitive science has assumed central stage in the contemporary
discussions on mind and artificial intelligence. I discuss two problems
concerning internal representations as it arise in the artificial intelligence
framework. The first problem is-'meaning barrier'- of forming the correct
structure of representation which is regarded as 'the central task facing
the artificial-intelligence community'. The different models of conceptbased perception are suggested as attempts at solving this problem. A
machine with a developed concept-based perception can be rightly taken
as a 'thinking machine'. The second problem is-'experience barrier' as I
would call it-of having perceptual experience. To address this problem an
alternative account of the notion of internal representations-'derivations'and a 'derivational framework' of mind are offered. I translate this
approach into the framework of artificial intelligence to see the possibility
of an 'experiencing machine', a machine that not only 'thinks' but thinks
like human beings. I conclude with some remarks on certain independent
evidences in favour of the derivational approach outlined.
INTERNAL REPRESENTATIONS:
AN APPROACH TO AI
2
The Notion ‘Artificial Intelligence’
Intelligence
Mind
Problem in the study of mind ; Lack of Data?
Theory of Mind-Black box analogy-Transcendental
Arguments
Intuitive suggestions from all intellectual disciplines
about Mind
Case: Arts, Paintings, Music, Mathematics,
Philosophy etc,
The movie Matrix
INTERNAL REPRESENTATIONS:
AN APPROACH TO AI
3
INTERNAL REPRESENTATIONS:
AN APPROACH TO AI
4
Rene Magritte about The Human Condition
“I placed in front of a window, seen from a room, a painting
representing exactly that part of the landscape which was
hidden from view by the painting. Therefore, the tree
represented in the painting hid from view the tree situated
behind it, outside the room. It existed for the spectator, as it
were, simultaneously in his mind, as both inside the room in the
painting, and outside in the real landscape. Which is how we
see the world: we see it as being outside ourselves even though
it is only a mental representation of it that we experience inside
ourselves”.
INTERNAL REPRESENTATIONS:
AN APPROACH TO AI
5
# Kant's Model of Mind
Perception(passive) & Conception (active)
Chomsky UG
Genie's case
Hubel and Wiesel
Neuroscience
# What about AI? What model of mind is
presupposed?
INTERNAL REPRESENTATIONS:
AN APPROACH TO AI
6
Where do we place Turing's Question
# The problem of Mental RepresentationsHigh-level perception(concepts begins to play its role)
-The structure of representations?
-How it is formed?
-How it is influenced by context? (PNAS)
-How can perceptions radically reshape themselves when
necessary?
-concepts, meaning, understanding?
# Modularity Vs Non-modularity
possibility of representation module (single “correct”
representation in allINTERNAL
situations,
no flexibility)
REPRESENTATIONS:
AN APPROACH TO AI
7
INTERNAL REPRESENTATIONS:
AN APPROACH TO AI
8
# AI and The problem of Internal Representations
What is the correct structure for representations?
-predicate calculus?frames?scripts?semantic
networks?…
Our focus Short-term active representations as these
are the direct product of perception.
The problem of relevance: To determine which
part of the data are relevant to a given representation,
a complex filtering process is required.
The problem of Organization: How are these
data put into the correct form for the representation?
INTERNAL REPRESENTATIONS:
AN APPROACH TO AI
9
# The traditional approach to AI
- select a preferred type of high-level representational
structure.
-select the data assumed to be relevant to the
task.(human programmer/hand-code a representation)
The problem of representation is ignored
E.g. face detection
but c.f, Machine learning,speech processing (stops
short of modeling at the conceptual level)
INTERNAL REPRESENTATIONS:
AN APPROACH TO AI
10
# The "meaning barrier“
-The formation of appropriate representations lies at the heart of
the high-level cognitive abilities. The problem of high-level
perception forms the central task facing the AI community, the task
of understanding how to draw meaning out of the world.
-On the one side of the barrier, models in low level perception, yet
not complex enough to be ‘meaningful’.
-On the other side , high level cognitive modeling has started with
conceptual representations(predicate logic/nodes in a semantic
network etc), meaning is already built-in.
And there is gap between the two.
INTERNAL REPRESENTATIONS:
AN APPROACH TO AI
11
# Traditional AI : characterized by an objectivist view of
perception,representation etc.
PSSH (Newell & Simon '76) upon which most of the
traditional AI enterprise has been built, posits that thinking
occurs thru’ the manipulation of symbolic representations,
which are composed of atomic symbolic primitives. Such
representations are rigid, fixed, binary entities.
# However, recently, Connectionist models whose
distributed representations are highly context-dependent
(Rumelhart & McClelland ’86). Here no representational
primitives in internal processing, as each representation is a
vector in a multidimensional space,whose position can adjust
flexibly to changes in context.
-Recurrent connections(Elman), Classifier-system models
(Holland) etc,.
INTERNAL REPRESENTATIONS:
12
AN APPROACH TO AI
INTERNAL REPRESENTATIONS:
AN APPROACH TO AI
13
# Case study: the problem of representation is
overlooked
Bacon : a program as a model of scientific discovery
SME: a computational model of analogy-making-the
Structure Mapping Engine e.g.,.analogy of the solar
system.
# To deal with the greater flexibility of human perception
and representation, integration of task-oriented
processes with high-level perception is necessary.
INTERNAL REPRESENTATIONS:
AN APPROACH TO AI
14
micro domains?
E.g.: Chapman's "Sonja" program(1991)
Hofstadter group's Copycat/Tabletop
architecture
Shrager (1990)
Further models
Connectionist networks, classifier systems,...
INTERNAL REPRESENTATIONS:
AN APPROACH TO AI
15
Summary problem 1
Concepts and perception together
make cognition possible. But
researchers in AI often try to bypass
perception while modeling cognition. A
system can not have cognition unless it
have the processes that build-up
appropriate perceptual representations.
Integrating perceptual processes into a
cognitive model leads to flexible
internal representations. Recently,
many models that encompass this idea
have been suggested. These models
take AI closer towards the workings of
mind.
INTERNAL REPRESENTATIONS:
AN APPROACH TO AI
16
Problem 2
Experiential aspect of a mental state:
Two elements: Objective and Subjective
The subjective (slide 23)
Problem for RTM:
Intuitions: Inverted Spectrum
: Brain in a Vat
INTERNAL REPRESENTATIONS:
AN APPROACH TO AI
17
INTERNAL REPRESENTATIONS:
AN APPROACH TO AI
18
INTERNAL REPRESENTATIONS:
AN APPROACH TO AI
19
INTERNAL REPRESENTATIONS:
AN APPROACH TO AI
20
INTERNAL REPRESENTATIONS:
AN APPROACH TO AI
21
The concept of representation; Derivation
no emphasis on content
Systematicity
arbitrary
Consumption
fluidity
Internal representations are internal states whose functional
role is to bear certain specifiable contents. Further we want to
respect the associated idea that much of the experiential
elements may be grounded not in the rigid activity of inner
vehicles, but in complex interactions involving neural, bodily
and environmental factors.
E.g.:Christopher Longuert-Higgins Robot
H2O model
INTERNAL REPRESENTATIONS:
AN APPROACH TO AI
22
Self-reference
Feed back loop
Godel's theorem
The second problem is how to account for the
experiential/phenomenal properties of mental
representations.
E.g.,My conscious Tsunami experience -”bluish splashy
way for me” has two parts, one represents the approaching
giant waves (primary, qualitative content), second
derivationally represents the experience/the awareness of it
(secondary, subjective content). The idea is, these
elementary components integrate to form a unitary
conscious state
INTERNAL REPRESENTATIONS:
AN APPROACH TO AI
23
INTERNAL REPRESENTATIONS:
AN APPROACH TO AI
24
The argument for the derivationality of Phenomenality is :
1) A mental state M of subject S has P when, and only when, S is aware
of M;
2) Awareness of X requires mental representation of X; therefore,
3) M has P when, and only when, S has a mental state M*, such that M*
derivationally represents M.
The mechanism that make this kind of MR possible, I conjecture, is a
kind of 'loop' an interaction between M and M* . Here M reaches up to
M* and influences it , while being influenced by M*. Even one mental
representation works as a network. In this sense MR does not stand for
other (external) objects. It operates by means of the direct causal
properties of M and M* and employs feed back loop and forward loop
Any theoretical framework that takes mental representations as arbitrary,
systematical, consumable, fluid, self -referential information-bearing
structures can be called derivational.
INTERNAL REPRESENTATIONS:
25
AN APPROACH TO AI
INTERNAL REPRESENTATIONS:
AN APPROACH TO AI
26
I now wish to go back to the example, to see the new tools are of any
use to address it. My conscious Tsunami experience -”bluish
splashy way for me”, as we have seen, has two components.The first
element represents the approaching giant waves (primary,
qualitative, yet objective content) and taken as M in the argument for
derivationality. The second element, second derivationally
represents the experience/the awareness of it (secondary, subjective
content)which is taken as M* in the argument which in turn
influence M in the manner of a loop.
The idea is, these derivations from M to M* and vice versa form a
unitary conscious state
I wish to add some empirical evidence in support of this
approach.The recent model of a single neuron as a network suggests
that representations are neither IDEAS as Hume conceived it nor
static communicative ideas as many representationalists argued.
INTERNAL REPRESENTATIONS:
AN APPROACH TO AI
27
Conclusions:
Recognizing the centrality of perceptual
process makes AI more in tune with the
findings of Cognitive Science. It yields
human perception-like flexible representations
and actions. Within a domain and within a
level these AI representations work as internal
derivations. With added features like selfreference, feed back loop etc, it may work
exactly as mental derivations.
INTERNAL REPRESENTATIONS:
AN APPROACH TO AI
28
Core references: (Besides CS621 notes & references, Homepages and Websites of theorists and
teams working in Cognitive Science).
Chomsky, N.
Press.
(2000) New Horizons in The Study of Language and Mind. Cambridge Univ
(2005) "Three Factors in Language Design" Linguistic Inquiry V 36, Winter '05. pp1-22.
Darwin, Charles (1859)
Fodor, J. A.
The Origin Of Species.
(1983) The modularity of Mind. Cambridge: Bradford Books, MIT Press.
(1998) Concepts: Where Cognitive Science Went Wrong. Oxford. OUP.
(2000) The Mind Doesn't Work That Way: the scope and limits of computational
psychology. Cambridge: Bradford Books, MIT Press.
Haugeland, John (ed) (1997)Mind Design II Philosophy Psychology Artificial Intelligence,
Revised and enlarged edition
Bradford Book, The MIT Press Cambridge, Massachusetts.
INTERNAL REPRESENTATIONS:
AN APPROACH TO AI
29
Hofstadter, D (1979) Godel, Escher, Bach: An Eternal Golden Braid, The Harvester
Press, London.
(1995) Fluid Concepts and Creative Analogies-Computer Models
of the Fundamental Mechanisms of Thought. Penguin.
Kant, Immanuel (1787) The Critique of Pure Reason.
Nagel, Thomas (1974) "What is it like to be a Bat?" in Philosophical review.
Chua, Hannah Faye., Boland, Julie E., and Nisbett, Richard E. (2005) "Cultural
variation in eye movements during scene perception". PNAS August 30, 2005 vol. 102
no. 35 12633.
Turing, Alan M.
59,236,pp:433-60
(1950)
"Computing Machinery and Intelligence", Mind V
INTERNAL REPRESENTATIONS:
AN APPROACH TO AI
30