Noticing causal properties of objects from sequence statistics
Anna Leshinskaya & Sharon L. Thompson-Schill
0.1
0
sentence acceptability test
sibbie
.02
.36
.43
static
effect
.10 .05
.04
.01
.35
rare
.25
.93
.02.25
.02
.36
.53
.70
.01
effect
.35
rare
frequent
.15
.04
.25
sequence presentation
.01
frequent
.15
250 trials/object; subjects told to pay close attention and try to predict what will happen
next (press a key when something unexpected occurs). Attention checks required subjects to identify the rare event.
2.4 s
1.2 s
*
0.7
7. Subjects can classify
a novel exemplar based
on common causal direction to the same
event under supervised
(labeled) conditions.
n = 30
0.6
0.5
0.4
0.3
0.2
0.1
0
*
causal acceptability
only in subjects who were accurate on non-causal questions (n =25)
2
3
4
unsure
5
definitely true
3
4
4.5
unsure
definitely true
*
light causer 2
3
2.5
2
1
5. Subjects accept causal interpretations of these contingencies
6. They conceptually distinguish object-event co-occurrence vs. object-dependent contingency structure.
Exp. 2
supervised category membership judgments
sibbie
cause
thale
effect
cause
effect
is this a sibbie or a thale?
cause
effect
vs.
frequent
cause
frequent
rare
effect
rare
frequent
cause
frequent
rare
frequent
rare
effect
cause
effect
rare
frequent
rare
more different
3.5
1.5
5
collected using a spatial sorting task
light causer 1
4
Thales cause confetti to appear.
2
*
5
effect to cause
caveat: this task included
feedback on contingency
(though not category)
learning; future work will
test this without.
similarity judgments
ligh
1. Subjects can identify which events are contingent on each other
from a continuous stream of events with minimal instruction.
2. They can bind different contingency structures to distinct objects.
3. These represenations are explicitly available.
4. And they are naturally directional (ie., structured).
shape to condition assignments counterbalanced
2
1
ctor
definitely true
t rea
unsure
5
effect
ligh
4
effect
1
definitely false
3
effect
ctor
2
1.5
effect
t rea
1
2
ligh
Light flashing around thales and their tilting are
strongly related.
cause
r2
unordered: cause & effect
2.5
cause
use
definitely true
cause
t ca
unsure
5
cause
ligh
4
object stimuli
light-reactors
light-causers
r1
definitely false
3
3
object-dependent structure
use
2
3.5
*
Exp. 3
unsupervised category induction
t ca
4
definitely true
*
n = 47
effect to cause
1
*
casue to effect
unsure
5
After the light flashes, thales tend to tilt.
.10 .05
.04
.01
4
ordered: cause to effect
definitely false
cause
.25
.93
.04
.25
3
only for object where frequent event was ambient, and for subjects who were
accurate on non-causal contingency and frequency judgments (n = 20)
.02
.53
.70
2
definitely false
contingency
.02 .49
1
1
.02
.02.25
After thales tilt, the light tends to flash
4.5
co-occurring (frequent)
event assignments varied across subjects; structure varied conditional on object identity
.43
ordered: effect to cause
definitely false
stimuli: objects
cause
5
Thales tilting causes the light to flash.
conditions: transition probabilities
static
tests explicit access and conceptual interpretation
1
thale
0.8
*
object by presentation order
mean response
Objects differ in terms
of two types of statistics: identity of the rare
event, and direction of
contingency between
two other events.
Object 2
0.2
cause to effect
3 object-based and 3 ambient
Object 1
0.3
co-occuring
stimuli: objects & events
Mean percent
accuracy =
0.65 with t(46)
= 4.19, p <
0.0001
0.4
frequency
Exp. 1
statistical property learning
Correlation between objects:
r(45) = -0.01
n = 47
0.6
unordered cause & effect
4. If statistics are assigned as properties of objects, they should enable
category formation in both supervised and unsupervised contexts (Gopnik
& Sobel 2000; Nazzi & Gopnik 2003; Kemp et al 2010). When these properties refer to the structure of events, they should be highly tolerant to sensory details of the objects and events, because they are based on higher-order relations rather than the sensory details. Can novel categories be
built purely on the basis of causal direction to an ambient event? Can
these categories generalize across the nature of the specific events involved?
*
0.5
1. Participants in such learning scenarios don't typically have explicit
access to this knowledge. Can they gain explicit access?
3. If statistical structure among events informs concept formation, then
might different kinds of statistics lead to different conceptual inferences?
*
0.7
which of the two videos is more typical for a thale?
Even naive learners are adept at inferring structure from raw streams of
events without top-down guidance (e.g., Hunt & Aslin, 2001; Saffran et al.,
1996; Turk-Browne, Jungé, & Scholl, 2005). Can this kind of mechanism
serve as a basis for building non-physical object properties?
2. Causal induction paradigms (e.g Cheng 1997) warn subjects they will be
making causal judgments and about which target event. Can causality
emerge as a property even when you're not looking for it?
0.8
mean response
Many properties of objects are not physical features: for example, turns on
the light, protects from rain, and enables communication are essential to concepts like light switch, umbrella, and telephone. How could we acquire such
property concepts, bottom-up, from experience?
strong transitions compared to weaker transitions to
test discrimination of relative contingency strength
0.9
Frequency
Questions
familiarity forced choice test
Contingency
University of Pennsylvania
vs.
cause
effect
frequent
rare
cause
effect
frequent
rare
light reactor 1
light reactor 2
more similar
within vs. between category distance: t(10) = -5.23, p = 0.0001
8. Subjects group objects by purely statistical properties.
9. They generalize over both object shape and the nature
of the causal/effect event.
Conclusions
Human adults can acquire
object categories based
purely on statistical properties in a bottom-up manner from sensory streams, and gain explicit
access to this knowledge. These categories can generalize over the
nature of the individual events. We are currently tracing the neural
mechanisms that allow such representations to be computed.
Cheng, P. W. (1997). From covariation to causation: A causal power theory. Psychological Review, 104(2), 367–405.
Gopnik, A., & Sobel, D. M. (2000). Detecting blickets: how young children use information about novel causal powers
in categorization and induction. Child Development, 71(5), 1205–1222.
Hunt, R., & Aslin, R. N. (2001). Statistical learning in a serial reaction time task: Access to separable statistical cues by individual
learners. Journal of Experimental Psychology. General, 130(4), 658–680.
Kemp, C., Tenenbaum, J. B., Niyogi, S., & Griffiths, T. L. (2010). A probabilistic model of theory formation. Cognition,
114(2), 165–96.
Nazzi, T., & Gopnik, A. (2003). Sorting and acting with objects in early childhood: An exploration of the use of causal cues. Cog
nitive Development, 18(3), 299–317.
Saffran, J. R., Aslin, R. N., & Newport, E. L. (1996). Statistical learning by eight-month-old infants. Science, 274(5294), 1926–1928
Turk-Browne, N. B., Jungé, J., & Scholl, B. J. (2005). The automaticity of visual statistical learning. Journal of Experimental Psy
chology. General, 134(4), 552–64.