Embodiment

1
Embodiment

• presented byLee McCauley

2
Overview

• Defining Embodiment
• Implications for embodied cognition
• An experimental example

3
Intelligence is determined by the dynamics of
interaction with the world

• (Brooks, 1991)

4
Definitions of Embodiment

• A system X is embodied in an environment E if
  perturbatory channels exist between the two.
  That is, X is embodied in E if for every time t
  at which both X and E exist, some subset of Es
  possible states have the capacity to perturb Xs
  state, and some subset of Xs possible states
  have the capacity to perturb Es state. (Quick
  et al. 1999)

5
Situatedness

• The idea that the physical world influences the
  behavior of an agent
• In other words, the environment has a direct
  effect on the internal structures of the agent

6
Structural Coupling

• There is a difference between the relationships
  of a systems components and the components
  themselves
• A set of relational criteria define the
  organization of a class of systems (Maturana
  Varela, 1980)
• The components that make up the organization
  define the structure of a system (Maturana
  Varela, 1980)

7
Structural Coupling (cont.)

• Structural coupling is, therefore, the mutual
  synchronization of environment-agent
  perturbations
• Example a fly on a painting by Rembrandt
• The fly is not structurally coupled to the full
  culturally dependant properties of the painting

8
Implications

• Just as a clock maker cannot replace parts
  randomly and expect a clock to function
  correctly, perturbations of the states of an
  agent and environment must be in sync or they are
  not structurally coupled
• For complex agents, this requires complete
  self-adaptation

9
Is physical embodiment necessary?

• Sharkey and Ziemke (2000) would argue that it is
• Autopoietic follows its own agenda (Sharkey and
  Ziemke infer radical implications for this)
• Allopoietic system that is controlled externally
• The implication is that an embodied system must
  not only be autopoietic cognitively, but also
  physically

10
Is physical embodiment necessary?

• Sharkey and Ziemkes argument is really a
  practical argument rather than a theoretical one
• Is physical embodiment necessary? No, but it
  helps
• What their argument reiterates, however, is the
  notion of self-steering its own development

11
Embodiment defined (by Riegler, 2002)

• A system is embodied if it has gained competence
  within the environment in which it has
  developed.
• The fact that most or all current artificial
  intelligence programs do not exhibit embodiment
  has to do with their explicit design rather than
  with the space they are habitating.

12
Applying embodiment to the symbol grounding
problem

• First, a clarification of what a symbol is
  (Peirce)

Interpretant
A Semiotic Symbol
Representamen (not necessarily physical)
Object
13
Symbol clarification

• A sign is called a symbol if the form in relation
  to its referent is either arbitrary or
  conventionalized, so that the relationship must
  be learned (Pierce)

14
Meaning of meaning

• A semiotic symbols meaning arises in its
  interpretation. As such the meaning arises from
  the process of semiosis, which is the interaction
  between form, meaning and referent. This means
  that the meaning depends on how the semiotic
  symbol is constructed and with what function.
  (Vogt, 2002)

15
Advantages of Semiotic Symbols

• They are, by definition, grounded
• It already bears the symbols meaning with
  respect to reality
• Lakoff (1987) argues that such structures are
  meaningful to begin with
• A semiotic symbol is situated and embodied
• It should be constructed (learned) through the
  interaction of the agent within its environment

16
Advantages (cont.)

• The semiotic symbols are not static
• Each symbol is (re-)constructed every time it is
  used

17
Robots learning language

• Vogt (2002) describes an experiment based on a
  previous version by Steels Vogt (1997)
• Using a language guessing game, the robots must
  come up with their own lexicon to describe
  objects that they sense in the environment
• The agents (robots) will be constructing and
  using semiotic symbols

18
The Language Game Model

• Steels (1996) suggested that there are three
  mechanisms for language development
• Cultural interaction
• Agents can share parts of their vocabulary
• Individual adaptation
• Agents expand their lexicon when they encounter
  novel situations
• Each speech act is evaluated for effectiveness
  which can be used to strengthen or weaken
  form-meaning associations
• Self organization
• Structural reorganization emerges from the
  iterative combination of the previous mechanisms

19
The Language Game Model (cont.)

• There have been several versions of this
  experiment that look at different aspects of
  language development
• Lexicon formation (Steels, 1996)
• Lexicon dynamics (Steels and McIntyre 1999)
• Multiple word games (Van Looveren, 1999)
• Stochasticity (Steels and Kaplan, 1998)
• On several platforms
• In simulations (De Jong, 2000 Belpaeme, 2001)
• On immobile robots (Belpaeme et al., 1998 Steels
  and Vogt, 1997 Vogt, 2000)
• On mobile robots (Steels and Vogt, 1997 Vogt,
  2000)

20
The experiment

• The players (robots) engage in a series of
  guessing games
• The hearer tries to guess what referent the
  speaker is naming

21
Steps of the experiment (game)

1. One robot assumes the role of speaker while the
   other assumes the role of hearer
2. Both robots start sensing their surroundings
3. Both robots preprocess the sensory data
4. The speaker selects one referent as the topic
5. The robots categorize the preprocessed data
6. The speaker produces an utterance and the hearer
   tries to interpret the utterance (assigns a
   meaning to it)
7. If the meaning matches one of the categorized
   referents, then an action can be taken by the
   hearer to acknowledge (like pointing)
8. Successes evaluated in the feedback which is sent
   to both robots so they can adapt their ontology
   of categories and lexicon

22
Whats the connection to symbol grounding?

• The guessing game is analogous to the process of
  semiosis
• The robots construct semiotic symbols
• Meaning arises from the sensing, segmentation and
  categorization of the referents
• When the guessing game is successful, then the
  semiotic symbols are created and used
  appropriately

23
Sensing
24
The Lexicon

• A lexicon L is a set of form-meaning associations
  FMi
• An entry in the lexicon is FMi Fi, Mi, si
• Fi is the form which is made up of a set of
  consonant-vowel strings
• Mi is the meaning represented by category
• si is the association score

25
Results

• Background
• Random chance suggests that speaker and hearer
  will select the same referent 23.5 of the time
• The speaker may select a topic that the hearer
  did not detect
• Potential understandability is around 80

26
Results for real
27
Results for real (cont.)
28
(No Transcript)
29
An interesting point

• An important result that the experiment reveals
  is that semiotic symbols need not be categorized
  the same under different circumstances. As the
  semiotic landscape shows, theres no
  one-to-one-to-one relation between a referent,
  meaning, and form this relation is rather
  one-to-many-to-one. In different situations, the
  robots detect the referents differently. Yet
  they are able to identify them invariably at the
  form level. In the process of arriving at such
  invariant identification, the co-evolution of
  form and meaning reveals to be extremely
  important.

30
Conclusions

• Embodiment requires structural coupling
• An embodied system does not have to be physical
• Symbols are not just labels, but require a
  semiotic process of self-guided construction that
  is synchronized with the agents interaction with
  its environment
• A one-to-many-to-one relationship between form,
  meaning, and referent does not sacrifice
  performance