ping pong in church - mitweb.mit.edu/tger/www/posters/poster - ping pong in church.pdf · ping pong...
TRANSCRIPT
Ping Pong in ChurchProductive use of concepts in human probabilistic inference
Tobias Gerstenberg
Noah Goodmanvs.
(mh-query 1000 100 ;Monte Carlo Inference ;CONCEPTS (definedefine personstrengthpersonstrength (memmem (lambdalambda (person) (gaussian 10 3)))) (definedefine lazylazy (memmem (lambdalambda (person game) (flipflip 0.1)))) (definedefine (teamstrengthteamstrength team game) (sumsum (mapmap (lambdalambda (person) (ifif (lazy person game) (/ (personstrength person) 2) (personstrength person))) team))) (definedefine (winnerwinner team1 team2 game)
(ifif (< (teamstrength team1 game) (teamstrength team2 game)) 'team2 'team1)) ;QUERY (personstrength 'A) ;EVIDENCE (andand (= 'team1 (winner '(TG) '(NG) 1)) (= 'team1 (winner '(NG) '(AS) 2)) (= 'team1 (winner '(NG) '(BL) 3)) (lazy '(NG) 1) ;additional evidence, used in Experiment 2 ))
How do people make inferences from complexpatterns of evidence across diverse situations?
What does a computational model need in order to capture the abstract knowledge people use for everyday reasoning?
strength
matchteamwinner
being lazy
1
>
>
>
strong indirect
A
B
B
B
C
D
>
<
<
weakindirect
A
B
B
B
C
D
>
>
>
diverse
A
A
A
B
C
D
>
>
>
confounded
A
A
A
B
B
B
0
0.5
1
1.5
stre
ngth
data model
*
* *
*
* = lazy in experiment 2
Doubles
Thinking = probabilistic inference over compositionally structured representations
Thinking is compositional and productive.
A B C
A BC
AB
C
CB AB
AB
C
D
AA BB
A B
C
D
Ping Pong Logic ConnectionistNetwork
BayesianNetwork
Model
E
3
Singles Doubles
CommentatorSingles
2
0.75 0.8 0.85 0.9 0.95 10
2
4
6
8
10
12
14
# pa
rtic
ipan
ts
Pearson correlation
Individual Correlation
−1.5 −1 −0.5 0 0.5 1 1.5−1.5
−1
−0.5
0
0.5
1
1.5
Model Prediction
Empi
rical
Mea
ns
r = 0.989RMSE = 0.161
Overall Correlation
>
>
>
B
B
B
C
E
G
A
A
A
D
F
H
>
>
>
B
C
D
E
E
E
A
A
A
F
G
H
>
<
<
B
C
D
E
E
E
A
B
B
F
F
F
>
>
>
B
C
D
E
E
E
A
B
B
F
F
F
>
>
>
B
C
D
E
G
I
A
A
A
F
H
J
>
>
>
B
C
D
C
B
B
A
A
A
D
D
C
0.5
1
1.5
stre
ngth
data model
confoundedpartner
confoundedopponent
strongindirect
weakindirect
diverse roundrobin
4
personstrengthnormally distributedpersistent propertylazyp(lazy) = 10%not persistent
teamstrengthindividual strengthscombine additively
winnerteam with greaterstrength wins
0
0.5
1
1.5
stre
ngth
strong indirect
weakindirect
diverseconfounded irrelevant
• people can reason flexibly based on di�erent patterns and sources of evidence
• compositional concepts support productive exten-sions over novel situations and objects
• only a handful of concepts required to predict inferences in fourty di�erent situations
• understanding people’s intuitive theories through models based on probabilistic programs
before
after
model
University CollegeLondon
StanfordUniversity
Tobi
Noah
Gerstenberg, T. G., Goodman, N. D., Lagnado, D. A., & Tenenbaum, J. B. (2012). Noisy Newtons: Unifying process and dependency accounts of causal attribution. Cognitive Science Proceedings
Goodman, N. D., Mansinghka, V. K., Roy, D., Bonawitz, K., & Tenenbaum, J. B. (2008). Church: A language for generative models. Uncertainty in Arti�cial Intelligence
Goodman, N. D., & Tenenbaum, J. B. (in prep). The probabilistic language of thought.