iterated learning and the cultural ratchet
DESCRIPTION
my presentation for cogsci 2009, as of 7/27TRANSCRIPT
The Cultural RatchetIterated Learing and variations
ResultsWhat does this mean?
Iterated Learning and the Cultural Ratchet
Aaron Beppu Tom Griffiths
Department of PsychologyUniversity of California Berkeley
Cognitive Science, 2009
Aaron Beppu, Tom Griffiths Iterated Learning and the Cultural Ratchet
The Cultural RatchetIterated Learing and variations
ResultsWhat does this mean?
Outline
1 The Cultural Ratchet
2 Iterated Learing and variations
3 Results
4 What does this mean?
Aaron Beppu, Tom Griffiths Iterated Learning and the Cultural Ratchet
The Cultural RatchetIterated Learing and variations
ResultsWhat does this mean?
What is the Cultural Ratchet?
The process of cumulative cultural evolution requiresnot only creative invention but also . . . faithful socialtransmission that can work as a ratchet to preventslippage backward – so that the newly invented artifactor practice preserves its new and improved form atleast somewhat faithfully until a further modification orimprovement comes along.
– Michael Tomasello, The Cultural Origins of Human Cognitionwe do thisother animals don’t
Aaron Beppu, Tom Griffiths Iterated Learning and the Cultural Ratchet
The Cultural RatchetIterated Learing and variations
ResultsWhat does this mean?
What are some existing explanations?
We can try to explain this by focusing on individual teachersand learners :
We’re sophisticated learners (e.g. Tomasello 2001)We learn strategies and goals from each otherWe can follow each others attentionWe can faithfully reproduce behaviors we learn from others
We’re also good teachers (e.g. Gergeley & Csibra 2005)We understand which things are hard to learnTeachers call learner’s attention to the most relevant partsHuman learners benefit from an enriched learningenvironment
Aaron Beppu, Tom Griffiths Iterated Learning and the Cultural Ratchet
The Cultural RatchetIterated Learing and variations
ResultsWhat does this mean?
Iterated Learning
Modeling Cultural Evolution
hypothesis
data data data
hypothesis
...
Learners are ordered into a sequence, and each agent learnsfrom the output of the previous agent
Can be used to study change in culture, (e.g. evolution oflanguage Kirby 2001), or to examine people’s biases (e.g.Griffiths, Christian & Kalish, 2006)
Aaron Beppu, Tom Griffiths Iterated Learning and the Cultural Ratchet
The Cultural RatchetIterated Learing and variations
ResultsWhat does this mean?
Iterated Learning
Analyzing Iterated Learning
Specifics:
Agents are Bayesian, and all have the same prior. Bayes’ Rule :
p(h|d) =p(h)p(d |h)∑h′ p(h′)p(d |h′)
The data that a given agent passes on is sample according tohis posterior beliefs. p(di |di−1) =
∑h p(di |h)p(h|di−1)
The pure iterated learning model converges to the prior beliefsof its agents.
Aaron Beppu, Tom Griffiths Iterated Learning and the Cultural Ratchet
The Cultural RatchetIterated Learing and variations
ResultsWhat does this mean?
Iterated Learning
Cumulative Cultural Evolution
Cumulative cultural evolution will require constantly bringing innew data from the world, in addition to some informationpassed from one generation to the next. We vary what kind ofinformation is passed between agents.
...hypothesishypothesis
??? ??? ???
new data new data
Aaron Beppu, Tom Griffiths Iterated Learning and the Cultural Ratchet
The Cultural RatchetIterated Learing and variations
ResultsWhat does this mean?
Iterated Learning
Observational Learning
...data data
hypothesishypothesis
data...
new data new data
In the “Mixed Data” case, each agent receives new data fromthe world, as well as data points sampled from the posteriordistribution of the previous agent. This is similar toobservational learning, achieved by observing the behavior ofanother.In this case, convergence is to
∑d∗ p(h|d∗)p∗(d∗).
Aaron Beppu, Tom Griffiths Iterated Learning and the Cultural Ratchet
The Cultural RatchetIterated Learing and variations
ResultsWhat does this mean?
Iterated Learning
Communicating Theories
...theory theory
hypothesishypothesis
theory
new data new data
In the “Posterior Passing” case, each agent receives new datafrom the world. In addition, rather than receiving samples fromthe previous person’s posterior distribution, each agent nowreceives the whole posterior distribution. Communicatingtheories might approximate how humans teach/learn.This chain converges to the truth.
Aaron Beppu, Tom Griffiths Iterated Learning and the Cultural Ratchet
The Cultural RatchetIterated Learing and variations
ResultsWhat does this mean?
Iterated Learning
Function learning is hard
Why function learning?
Strong bias towards positive linear functions; non-monotonicfunctions are hard (Busemeyer, Byun, DeLosh & McDaniel,1997)
The pure iterated learning case converges to a linear function(Kalish, Griffiths, Lewandowsky, 2007)
We can select a function that is too difficult for a single personto learn on their own, but which could plausibly be learned viacumulative cultural evolution.
By using a difficult to learn parabola, we can separate outgroups of people converging to the posterior from peopleconverging to the prior.
Aaron Beppu, Tom Griffiths Iterated Learning and the Cultural Ratchet
The Cultural RatchetIterated Learing and variations
ResultsWhat does this mean?
Apparatus
Aaron Beppu, Tom Griffiths Iterated Learning and the Cultural Ratchet
The Cultural RatchetIterated Learing and variations
ResultsWhat does this mean?
Examples from our results
Training Testing
Pure iterated learning
Aaron Beppu, Tom Griffiths Iterated Learning and the Cultural Ratchet
The Cultural RatchetIterated Learing and variations
ResultsWhat does this mean?
Training Testing
Mixed data
Pure iterated learning
Aaron Beppu, Tom Griffiths Iterated Learning and the Cultural Ratchet
The Cultural RatchetIterated Learing and variations
ResultsWhat does this mean?
Training Testing
Posterior passing
Mixed data
Pure iterated learning
Aaron Beppu, Tom Griffiths Iterated Learning and the Cultural Ratchet
The Cultural RatchetIterated Learing and variations
ResultsWhat does this mean?
0 2 4 6 80
0.1
0.2
0.3
0.4
0.5
0.6
Generation
RM
SE
0 2 4 6 80
0.1
0.2
0.3
0.4
0.5
0.6
Generation
RM
SE
pure iterated learningmixed dataposterior passing
Left : Cumulative average error (RMSE) for each chain for eachgeneration. Right : Same as left, but averaged by condition.
Aaron Beppu, Tom Griffiths Iterated Learning and the Cultural Ratchet
The Cultural RatchetIterated Learing and variations
ResultsWhat does this mean?
What do we walk away with?
Merely observing behavior of those around us isn’t enough tohave cumulative cultural evolution
We need to communicate theories and beliefs about the world.Language is one way we can do this.
This gives us an idea about the uniquely human characteristicsthat facilitate cultural evolution.
What next?
Can we adequately recover the posterior distributions from ourteachers merely from observing their actions?
What can we do with richer teacher/learner interactions?
How can we reconcile this inference based perspective withother pressures on cultural evolution?
Aaron Beppu, Tom Griffiths Iterated Learning and the Cultural Ratchet