1. Grounding compositional symbols:
no composition without discrimination?
Alberto Greco,
Elena Carrea
University of Genoa, Italy
greco@unige.it
elena.carrea@unige.it
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
2. Overview
What is Symbol Grounding?
Symbol Grounding and Compositionality
Why is Compositionality relevant to
Cognitive Sciences?
Purposes of the Study
Experiment
Provisional conclusions
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
3. The Symbol Grounding problem
Searle (1980) argued against the
symbolic model of mind with the
mental experiment of the
Chinese Room.
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
4. The Symbol Grounding problem
Can we say
that this
man
understands
Chinese?
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
5. The Symbol Grounding problem
Harnad’s Merry-go-round:
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
6. The Symbol Grounding problem
Harnad’s Merry-go-round:
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
7. The Symbol Grounding problem
Harnad’s Merry-go-round:
This problem is similar to having to learn
Chinese with only a Chinese/Chinese dictionary.
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
8. The Symbol Grounding problem
How can symbolic systems
be linked to the world?
Can a purely symbolic model
be grounded only in other
meaningless symbols?
This is the Symbol Grounding problem
(Harnad 1990).
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
9. The Symbol Grounding problem
Harnad’s solution:
grounding experience
must be sensorimotor…
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
10. The Symbol Grounding problem
Symbols (words) acquire their
meaning only when they are
associated with sensorimotor
grounding representations.
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
11. Overview
What is Symbol Grounding?
Symbol Grounding and Compositionality
Why is Compositionality relevant to
Cognitive Sciences?
Purposes of the Study
Experiment
Provisional conclusions
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
12. Compositionality
In Harnad's view, grounding has been
conceived as compositional:
from sensorimotor toil
to symbolic theft,
thanks to propositions.
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
13. Compositionality
An example:
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
14. Compositionality
An example:
“a zebra is like a
horse with stripes”…
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
15. Compositionality
What is compositionality?
A definition of compositionality from a conference on this
topic*:
“Compositionality is a key feature of structured
representational systems, be they linguistic, mental or
neuronal.
A system of representations is compositional just in case the
semantic values of complex representations are determined by
the semantic values of their parts.”
* Compositionality, Concepts, and Cognition:
An Interdisciplinary Conference in Cognitive Science,
Düsseldorf, Germany, February 28 to March 3, 2004
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
16. Compositionality
Systematic combination of meaningful
components,
According to syntactical rules.
Meaning of complex expressions,
Productivity (linguistic and conceptual).
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
17. Overview
What is Symbol Grounding?
Symbol Grounding and Compositionality
Why is Compositionality relevant to
Cognitive Sciences?
Purposes of the Study
Experiment
Provisional conclusions
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
18. Compositionality
Why is it relevant to Cognitive Sciences?
Formal languages of mathematics, logic, and
computer science are (usually) considered as
compositional.
Linguistics: are natural languages compositional?
Philosophy: is whole meaning really coming out from
part meaning? What about context?
Psychology: is representation compositional? Is
language of thought compositional? (from cognitivism to
ecological, embodiment, re-enacting theories).
Cognitive and neural modelling (AI, connectionism,
robotics): no easy compositionality in neural networks;
binding problem; Can analog features be represented
compositionally?
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
19. Overview
What is Symbol Grounding?
Symbol Grounding and Compositionality
Why is Compositionality relevant to
Cognitive Sciences?
Purposes of the Study
Experiment
Provisional conclusions
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
20. Purposes of the Study
Key-question:
What is complex symbolic
composition based on?
…a corresponding
composition of grounding
sensorimotor
representations?
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
21. Purposes of the Study
Key-question (in other words):
Are structured grounding
representations needed, to express
meanings that have
a complex structure?
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
22. Purposes of the Study
Key-question (in other words):
Are structured grounding
representations needed, to express
meanings that have
a complex structure?
How can we study this empirically?
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
23. Purposes of the Study
How could we empirically test whether
grounding representations
have been established?
Starting from scratch…
from nonsense, to meaning…
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
24. Purposes of the Study
Our solution:
association task between
nonsense symbols (words)
and target perceptual or motor
stimulus patterns.
nonsense words perceptual or motor
patterns
(arbitrary, not yet
grounded symbols)
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
25. Purposes of the Study
nonsense words perceptual or motor
patterns
Our assumption:
successful learning of this association
shows that symbols have been grounded,
and that a corresponding grounding
representation has been established.
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
26. Purposes of the Study
The experiment should allow us to
distinguish between two
possibilities:
blue circle
- compositional Grounding
REPresentations: separate GREPs
for each part (when a concept
GREPs
comes from the combination of
the parts );
- holistic GREPs: these combined blue-circle
concepts are best learned by
bringing them back to a single
unifying GREP. GREP
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
27. Purposes of the Study
Previous research on compositional grounding
with motor patterns
Greco & Caneva (2005) associated 3 pseudo-words to 3 features
(e.g. GAB DIN FIT or TANEC for “push right hand”).
compositional holistic
Greco & Caneva (2009) associated 2 pseudo-words to 2 features
(a particular nonsense motor pattern & hands up, down, fist)
(e.g. BASPI NOLE or TERPESOVA).
compositional holistic
And in other conditions (e.g. incremental learning) holistic again
…no consistent results
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
28. Purposes of the Study
Inconsistency in these results may be
explained by several factors:
• difficulty of learning tasks,
• different learning methods,
• words not balanced across conditions…
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
29. Purposes of the Study
We presupposed that:
• a single symbol is best supported by a single
Grounding REPresentation (GREP) when a
pattern tends to be represented holistically
• separate symbols (to be combined) are best
supported by separate GREPs when pattern
features tend to have separate representations
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
30. Purposes of the Study
We presupposed that:
• a single symbol is best supported by a single
Grounding REPresentation (GREP) when a
pattern tends to be represented holistically
• separate symbols (to be combined) are best
supported by separate GREPs when pattern
features tend to have separate representations
But we really don’t know much about the nature of
motor representations (do they tend to be
represented analytically or holistically?).
We need stimuli that have a
known representation…
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
31. Purposes of the Study
We presupposed that:
• a single symbol is best supported by a single
Grounding REPresentation (GREP) when a
pattern tends to be represented holistically
• separate symbols (to be combined) are best
supported by separate GREPs when pattern
features tend to have separate representations
But we really don’t know much about the nature of
motor representations (do they tend to be
represented analytically or holistically?).
We need stimuli that have a
known representation…
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
32. Overview
What is Symbol Grounding?
Symbol Grounding and Compositionality
Why is Compositionality relevant to
Cognitive Sciences?
Purposes of the Study
Experiment
Provisional conclusions
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
33. Experiment
From literature…
Handel & Imai (1972) showed different kinds
of properties of visual objects:
Integral Separable
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
34. Experiment
From literature…
Handel & Imai (1972) showed different kinds
of properties of visual objects:
Integral Separable
When attributes are
fused together, and are
perceived as one.
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
35. Experiment
From literature…
Handel & Imai (1972) showed different kinds
of properties of visual objects:
Integral Separable
When attributes are When attributes can be
fused together, and are easily perceived
perceived as one. separately.
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
36. Experiment
We repeated some classical experiments of
Handel & Imai to select good (integral and
separable) stimuli for our study.
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
39. Experiment
It is not an absolute separation;
integral and separable stimuli,
not features.
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
40. Experiment
INTEGRAL STIMULI:
We put the figures in jars to make the task more
engaging.
As integral stimuli, we used triangles with the same
blue which varied according to a matrix with 4
degrees of brightness and saturation.
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
41. Experiment
INTEGRAL STIMULI:
We put the figures in jars to make the task more
engaging.
As integral stimuli, we used triangles with the same
blue which varied according to a matrix with 4
degrees of brightness and saturation.
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
42. Experiment
INTEGRAL STIMULI:
We put the figures in jars to make the task more
engaging.
As integral stimuli, we used triangles with the same
blue which varied according to a matrix with 4
degrees of brightness and saturation.
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
43. Experiment
INTEGRAL STIMULI:
We put the figures in jars to make the task more
engaging.
As integral stimuli, we used triangles with the same
blue which varied according to a matrix with 4
degrees of brightness and saturation.
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
44. Experiment
INTEGRAL STIMULI:
We put the figures in jars to make the task more
engaging.
As integral stimuli, we used triangles with the same
blue which varied according to a matrix with 4
degrees of brightness and saturation.
An example:
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
45. Experiment
SEPARABLE STIMULI:
We used colored polygons as separable
stimuli.
We put into a matrix 4 shapes and 4 colors
as separable stimuli.
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
46. Experiment
SEPARABLE STIMULI:
We used colored polygons as separable
stimuli.
We put into a matrix 4 shapes and 4 colors
as separable stimuli.
An example:
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
47. Experiment
SEPARABLE STIMULI:
We used colored polygons as separable
stimuli.
We put into a matrix 4 shapes and 4 colors
as separable stimuli.
An example:
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
48. Experiment
SEPARABLE STIMULI:
We used colored polygons as separable
stimuli.
We put into a matrix 4 shapes and 4 colors
as separable stimuli.
An example:
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
49. Experiment
SEPARABLE STIMULI:
We used colored polygons as separable
stimuli.
We put into a matrix 4 shapes and 4 colors
as separable stimuli.
An example:
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
50. Experiment
In our task meaningless labels were associated with
integral and separable stimuli.
We set also two conditions for the labels:
holistic
(a single word for the stimulus as a whole);
compositional
(two words that describe two features of the
stimulus).
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
51. Experiment
Four conditions:
compositional holistic label
label (1 word)
(2 words)
integral stimuli
P Q
separable
stimuli R S
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
53. Experiment
• Participants (28 undergraduate students) saw the
instructions on a PC screen.
• The instructions explained that the task was to learn and
remember the names of some jars.
STAGES:
1.familiarization with the stimuli (judging the similarity
between pairs of jars);
2.associative Learning (Ss. saw a jar and its name);
3.test (Ss. should click the correct name after seeing the jar).
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
55. Experiment
Results:
After familiarization (similarity judgments), we
found the same results reported in the literature.
(distance best measured by euclidean metric for
integral pairs, by city-block metric for separable
pairs).
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
56. Experiment
Expectations
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
57. Experiment
Results:
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
58. Experiment
Results (percentage of correct answers):
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
59. Overview
What is Symbol Grounding?
Symbol Grounding and Compositionality
Why is Compositionality relevant to
Cognitive Sciences?
Purposes of the Study
Experiment
Provisional conclusions
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
60. Provisional conclusions
How can we explain these results?
• Integral stimuli: the association with 2 words is
more difficult because of what we called
“competitive grounding”.
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
61. Provisional conclusions
How can we explain these results?
• Integral stimuli: the association with 2 words is
more difficult because of what we called
“competitive grounding”.
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
62. Provisional conclusions
How can we explain these results?
• Integral stimuli: the association with 2 words is
more difficult because of what we called
“competitive grounding”.
First the subject thinks that baspi-
BOTE is a certain blue.
In a second moment another blue is
called tispi-BOTE (because of the
same lighting).
This creates a competition between
the 2 representation.
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
63. Provisional conclusions
How can we explain these results?
• Integral stimuli: the association with 2 words is
more difficult because of what we called
“competitive grounding”.
First the subject thinks that baspi-
BOTE is a certain blue.
In a second moment another blue is
called tispi-BOTE (because of the
same lighting).
This creates a competition between
the 2 representation.
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
64. Provisional conclusions
How can we explain these results?
• Separable stimuli: learning with 2 words did not
show any extra-benefit because of the limited
number of words and stimuli to recall.
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
65. Provisional conclusions
How can we explain these results?
• Separable stimuli: learning with 2 words did not
show any extra-benefit because of the limited
number of words and stimuli to recall.
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
66. Provisional conclusions
How can we explain these results?
• Separable stimuli: learning with 2 words did not
show any extra-benefit because of the limited
number of words and stimuli to recall.
Compositionality probably shows its
convenience with a higher number of
words and stimuli.
With only 8 words it’s easier to learn a
single label.
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
67. Provisional conclusions
How can we explain these results?
• Separable stimuli: learning with 2 words did not
show any extra-benefit because of the limited
number of words and stimuli to recall.
Compositionality probably shows its
convenience with a higher number of
words and stimuli.
With only 8 words it’s easier to learn a
single label.
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
68. Provisional conclusions
How can we explain these results?
• Separable stimuli: learning with 2 words did not
show any extra-benefit because of the limited
number of words and stimuli to recall.
Compositionality probably shows its
convenience with a higher number of
words and stimuli.
With only 8 words it’s easier to learn a
single label.
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
69. Provisional conclusions
For this reason we are doing a new
experiment with 12 words and stimuli
to learn.
We are interested in studying if
compositionality becomes convenient
(with separable stimuli) beyond a
critical number of stimuli.
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
70. Provisional conclusions
Limitations:
Using meaningless words is not “ecological”;
It’s difficult to create words free from
participant’s personal association with common
words;
It’s difficult to neutralized the prior knowledge
(e.g. about geometric shapes).
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
71. Provisional conclusions
Benefits:
We set a paradigm for testing grounding
representation with a learning task;
We can test whether 2 labels of the
compositional condition really have 2 different
grounding representations.
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
72. Provisional conclusions
In the final step of the experiment we asked
participants to select among 4 jars the one
described by a particular sentence.
In this set 2 jars shared 1 attribute (color) but not
the other (shape).
In other words: 2 jars were BASPI but only 1 was
BASPI-NOLE.
Then we asked the opposite task (two NOLE but
only one BASPI).
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010
73. Provisional conclusions
By using this setting we could
exclude that the participants
memorized the compositional
sentence as a single word.
ESSCS 2010 London, 7^ July 2010
lunedì 26 luglio 2010