Metaphor

1
(No Transcript)
2
Metaphor Machine Learning

• Science is all a metaphor

3
To Cover Outline

• The Research Question
• Grounding An example of a computational approach
  to body-based metaphor
• Srini Narayanan, 1999.
• Directions for improvement Automated discovery
  of metaphors in absence of pre-existing target
  domain model.
• Aims of my research.

4
The Research Question

• How can we automatically extend complex physical
  understandings of the world via metaphor-like
  processes to increasingly abstract understandings
  of the world?

5
Narayanan (1999)A computational Model

• A target domain is explained in terms of a source
  domain.
• Not new.
• Interpretation of the target domain is carried
  out by simulating within the source domain.
• Also not new.

6
Narayanan (1999)A computational Model

• The source domain is a sensory-motor control
  model of actions (e.g. walking), represented as a
  stochastic petri-net.
• The target domain is a temporally extended belief
  net modeling events in international economic
  policies.
• Metaphorical mappings map entities, processes and
  parameters between domains.

7
Narayanan (1999)Example

• Metaphor Liberalization plan stumbling.
• Simulation of stumbling in source domain provides
  the following inferences
• Context Ongoing plan, difficulties exist.
• Possibility of suspension.
• Possibility of failure.
• Further inferences from target domain
• Goal is free trade, deregulation.

8
Narayanan (1999)Strengths

• An attempt to see how abstract understandings can
  arise metaphorically from more concrete ones.
• Shows that body-centred processes can be
  redeployed to this purpose.
• An illustration of the ideas of Lakoff Johnson
  (1980).

9
Narayanan (1999)Limitations

• Source domain is merely a simulation (of e.g.
  walking).
• It is discrete this could be a problem It has
  never been shown to work in the world.
• It is a model of content (e.g. entities it has
  predicates) is it embodied enough?
• Target domain is pre-constructed. Much of the
  complexity is hidden.
• The mapping between the two is not automated.
  Its hand coded and simple.
• The model is contrived to prove a point its
  not particularly useful.

10
Narayanan (1999)To-do list

• Address limitations.
• Act in the world or in a world-simulation.
• Obtain metaphor/abstraction automatically.
• Baby-steps Dont use a pre-constructed target
  domain.
• Want to do more than just explain want to do
  something useful.
• The idea is to start with a concrete
  understanding and then see how we can extend it.
• So we need to work in source domain(s).
• What abstract understandings can we conjure from
  it (them)?
• What happens if we relax, combine, reconfigure or
  otherwise extend parts of the source domain(s) to
  do more interesting reasoning? I call this
  metaphor. Do you?

11
Overarching Aims

• Simulate people or animals?
• Problematic.
• Do things with machines better, meaner, more
  intelligently?
• That would be nice.
• Create something that is useful?
• Why, yes.

12
Recap Outline

• The Research Question
• Grounding An example of a computational approach
  to body-based metaphor
• Srini Narayanan, 1999.
• Directions for improvement Automated discovery
  of metaphors in absence of pre-existing target
  domain model.
• Aims of my research.

13
Some References

• Approaches to machine discovery using heuristics
  (Eurisko, Absolver)
• Lenat, D.B. Brown, J.S. (1984). Why AM and
  EURISKO appear to work. Artificial Intelligence,
  23, 269-294.
• Prieditis, A. (1993). Machine Discovery of
  Effective Admissible Heuristics. Machine
  Learning, 12, 117-141.
• A computational approach to body-centered
  metaphor
• Narayanan, S. (1999). Moving Right Along A
  Computational Model of Metaphoric Reasoning about
  Events. AAAI 1999, 121-127.
• A computational approach to automated inference
  of active spatial structure
• Philipona, D., ORegan, K., Nada, J.P., and
  Coenen, O.J. (2004). Perception of the structure
  of the physical world using unknown sensors and
  effectors. Advances in Neural Information
  Processing Systems, 15.

14
(No Transcript)
15
(No Transcript)
16
(No Transcript)
17
(No Transcript)
18
(No Transcript)
19
(No Transcript)
20
(No Transcript)
21

22
Recap Outline

• The Research Question
• An example of a computational approach to
  body-based metaphor
• Srini Narayanan, 1999.
• Directions for improvement Automated discovery
  of metaphors in absence of pre-existing target
  domain model.
• An example of work on machine discovery of
  concepts in abstract domains
• Douglas Lenat, 1984 (Eurisko).
• Directions for improvement Application to
  real-world domains.
• Aims of my research.

23
Lenat (1984)Automated Discovery

• Automated Mathematician Eurisko Programs that
  automate the discovery of new concepts using
  heuristic operators.
• Concepts expressed as frames containing LISP
  structures.
• Eurisko allowed heuristic operators to operate on
  heuristic operators to develop new heuristic
  operators

24
Lenat (1984)Example from AM

• List equality can be generalized to equality of
  list lengths.
• Thus, any list can be canonized to be an
  unordered list of an arbitrary symbol T.
• A join operation on such an unordered list is
  equivalent to addition Each such canonized
  unordered list is a called a number.
• A multiplication operation is discovered through
  a particular application of operators on lists
  and operators on operators.

25
Lenat (1984)Strengths

• Discoveries are largely automated The system is
  capable of finding new things.
• Exploits the overlap between data structures and
  processes The system can evolve its own
  discovery processes.

26
Lenat (1984)Limitations

• What can be found by the system is determined by
  choice of representation and by choice of initial
  heuristics.
• Heuristics are tailored to a specific set of
  problems.
• LISP data structures A very abstract, sparse and
  passive space in which to work.

27
Lenat (1984)To-do list

• Can we lean on the richness of a real-world
  domain?
• Can we make operations the basic units rather
  than static structures?
• Can we utilize well-defined informational
  criteria as opposed to a set of ad-hoc heuristics
  measures? (e.g. predictability of outcome of
  processes)
• Can we build a system that learns to conduct
  experiments on both the world and itself?

28
(No Transcript)