Transcript Modeling event perception in infancy

Modeling event perception in infancy
Arthur Franz
Frankfurt Institute for Advanced Studies
http://fias.uni-frankfurt.de
FIAS, 2008-4-29
Arthur Franz
1
Motivation
•
Where does knowledge come from?
 Study “simple” systems: infants
•
Perception and conception of spatial events seem to be
crucial
•
E.g.: occlusion, launching, object unity and, permanence,
continuity, object solidity, support,… ”naïve physics”
•
Hypothesis: most of them can be learned from purely
statistical properties of visual input.
Arthur Franz
2
The habituation paradigm
How do people investigate what infants know?
 Habituation paradigm
Example: Perception of object unity
rod movement baseline
test
displays
Arthur Franz
Mean looking time (sec)
habituation
displays
habituation
habituation
test
test
3
How can we model this?
Input example for the object unity experiment
7 x 7 pixel retina
BACKGROUND
FOREGROUND
We build a network that learns to represent occluded
objects.
Arthur Franz
4
Input coding
Assumption:
Neurons tuned to
velocity AND disparity
In MT?
Arthur Franz
5
Neural network
Got it?
Arthur Franz
6
Calculation details
Learning: backpropagation of error with real inputs and outputs.
Objective function:
Arthur Franz
7
Pre-training
Pre-training corresponds to the infant’s visual experience
with the world
 Varying the pre-training time allows for modeling
infant’s of various ages!
Pretraining with random
moving or stationary rectangles
Arthur Franz
8
What the network “imagines”
Arthur Franz
real inputs
full inputs
real outputs
imagined outputs
9
Relation to infant experiments
In infant experiments the looking time is measured.
Looking time ~ attention, novelty ~ habituation (“tiring”) of
certain neurons in the infant’s brain.
New stimulus => other neurons get active => dishabituation
Dishabituation in the model = difference between the
hidden layer activity during habituation and the activity
during a test stimulus.
Arthur Franz
10
Modeling object unity (1)
full inputs
rod movement
baseline
habituation
displays
complete rod
broken rod
test
displays
Rod movement => preference for broken rod
Baseline
Arthur Franz
=> no preference
11
Modeling object unity (2)
rod occlusion
complete rod broken rod
control
control
habituation
displays
test
displays
Rod occlusion => preference for broken rod (age effect!)
Complete rod control => pref. for broken rod
Broken rod control => pref. for complete rod
Arthur Franz
12
Modeling object unity (3)
full inputs
rod occlusion
rod pieces
habituation
displays
test
displays
Rod occlusion => preference for broken rod
Rod pieces
Arthur Franz
=> preference for broken rod
13
Modeling object unity (4)
rod moves
rod & block move
block moves
no movement
habituation
displays
test
displays
Result: after long pre-training the network shows
a preference for the broken rod in each condition!
=> Age effect, see adult data
Arthur Franz
14
Modeling object unity (5)
rod-polygon
baseline
complete
broken
habituation
displays
test
displays
rod-polygon => preference for broken rod
Baseline
Arthur Franz
=> no preference
15
Adult data
Arthur Franz
16
Modeling perception of occluded
trajectories (1)
habituation
thick
4-month-olds
occluder
thin
occluder
continuous
test
discontinuous
test
Arthur Franz
17
Modeling perception of occluded
trajectories (2)
long pre-training
exp. condition
baseline
short pre-training
exp. condition
baseline
habituation
displays
test
displays
Exp. condition => preference for
discontinuous display
Exp. condition => preference for
continuous display
Baseline => no preference
Arthur Franz
18
Modeling perception of occluded
trajectories (3)
thin
thick
preference
2 mo 4 mo
6 mo
Pre-training time / 1000
• Natural explanation for data
• Model explains how and why preferences change
• Object permanence develops in the network!
Arthur Franz
19
Summary
• The neural network provides a model of infant’s perception of
occluded objects, object unity and object permanence.
• In total 9 fundamental experiments from 2 different laboratories
have been explained.
• The network is a developmental model and can reveal the
mechanisms of change. Especially, the how and why questions can be
adressed.
• It demonstrates that much of infants’ perception can be learned
and explained solely on the basis of statistical regularities of raw
visual input. No innate principles or modules need to be postulated.
Arthur Franz
20
Drawbacks and open questions
•
Backpropagation of error => Andrea’s network?
•
Dishabituationmeasurment is done only with first hidden layer.
What not the second? Why not the whole network? Habituation
with intrinsic plasticity?
•
Stimuli are “flat” on the screen in the lab => no bottom-up
disparity-based separation possible!
•
Evidence for neurons tuned to both velocity AND disparity?
•
In some experiments the prediction error is more suitable as a
dishabituation measure. How to combine?
•
Imagined outputs are noisy. The calculation of the full inputs is too
“constructed”.
Arthur Franz
21
Future work
•
Predictions of the model
•
Elaborate the relation of this model to the
experimenters verbal accounts
•
Include other event categories into pre-training
(blocked motion, launching). Many other experiments
can then be explained. continuity, solidity, object
permanence,…
Arthur Franz
22
Thank you!
Arthur Franz
23
References
• Kellman, Spelke (1983). Perception of Partly Occluded
Objects in Infancy. Cognitive Psychology, 15, 483-524
• S.P. Johnson, J.G. Bremner, A. Slater, U. Mason, K.
Foster, A. Cheshire (2003). Infants' perception of
object trajectories. Child Development, 74, 94-108
Arthur Franz
24
Talk feedback
• Hidden+context layer doing everything?
• Feedback from im. Layer to hidden/context?
• Bayesian / optimization approach
• Disparity cells not present before 4 months
• Modeling with Kalman filters
Arthur Franz
25