words, memes and collective intelligence: the emergence of distributed cognition
DESCRIPTION
Talk presented by Francis Heylighen at the Language and Complexity Symposium in Barcelona (2010)TRANSCRIPT
Francis HeylighenEvolution, Complexity and Cognition group
Vrije Universiteit Brussel
Words, Memes and Collective Intelligence:
the emergence of distributed cognition
• reductionism emergence→
– whole is more than the sum of its parts• determinism uncertainty→
– we cannot predict exactly what will happen• materialism organization→
– no matter-mind duality• reversibility evolution→
– systems change irreversibly towards more adaptive and complex organization
From Newton to Complexity
• Collection of interacting agents
• Characterized by self-organization
– global coordination and adaptation– emerging from local actions
• examples– Societies, groups, communities– Internet, markets– Ecosystems, organisms
Complex Adaptive Systems
• Goal-directed individual– tries to maximize "fitness" or "utility"
• Senses local conditions– compares sensation with goal to establish possible
discrepancy• decides about appropriate action• performs action• monitors effect of action
– performs if necessary new action to correct remaining discrepancies
Agent
• Survival and success requires solving problems– this implies intelligence
• Perception: gathering information• information processing• problem-solving
– searching for best strategy to reach goal• decision-making
– selecting the most appropriate action
Agents Exhibit Cognition
• Agents have only local information• Agents limited in their capacity for processing
information• The environment is complex and unpredictable
– non-linearity may give rise to chaos• Therefore, agents cannot predict the long-term
effects of their actions– they must rely to some degree on trial-and-error– They must constantly learn and adapt their actions
Bounded rationality
• complex systems spontaneously order themselves– no internal or external agent is in control of the process– the process is distributed over all participating agents– the resulting order is robust under perturbations
• agents "align" the one with the other– Local alignment propagates to neighbours
• global order emerges from local interactions– small changes may be amplified– until they spread over the system as a whole
Self-organization
Example: Magnetization
• Co-ordination– arrangement of agents or actions
– minimization of conflict or friction
– retention and amplification of cooperation or synergy• Basic mechanism
– reinforcement of synergetic interactions– suppression of frictional interactions
• Self-reinforcing growth of the synergetic assembly– speeds up process– but may amplify biases
Self-organization
• Eventually, actions and agents become coordinated
• Emergence of a "superagent"– coordinated, goal-directed collective
– characterized by collective intelligence: group can solve more problems than its members
• Distributed cognition– different agents contribute different results at different
times and places– results are integrated– together, they solve the global problem
Emergence of distributed cognition
Distributed Cognition
• Together we can be more intelligent than individually– E.g. ants, bees, termites– But also more stupid
• James Surowiecki: – "The Wisdom of Crowds”– Summary of requirements for
collective intelligence
Collective Intelligence
QuickTimeª and aTIFF (Uncompressed) decompressor
are needed to see this picture.
QuickTimeª and aTIFF (Uncompressed) decompressor
are needed to see this picture.
Requirements for Collective Intelligence
• Diversity
–the more varied the individual experiences, the more comprehensive and balanced their collective knowledge
• Aggregation
–individual contributions must be synthesized into global solution
• Independence
–no individuals or subgroups should be able to bias the result
• Decentralization (distribution)
–Different individuals should be able to work on different parts of the problem
–Division of Labor
Group discussion
QuickTimeª and aTIFF (Uncompressed) decompressor
are needed to see this picture.
• Talking Heads simulation: emergence of language– Steels 1999
• Talking Nets simulation: group communication– Van Overwalle & Heylighen 2006
• Memetic transmission experiment– Lyons & Kashima 2003
• Happiness discussion experiment– Heylighen & Biebaut (unpublished)
Some studies of emergent distributed cognition
• Development of shared standards– agents must agree about doing things the same way
• so as to avoid confusion or conflict
– for example• driving on the same side of the road
• using the same conventions
• using compatible technological formats
– e.g. vocabulary, grammar, morals, norms, traffic rules• Integration of complementary knowledge
Types of cognitive coordination
• Division of labor– who does what?– parallel coordination– allows greater specialization
• Workflow– who does what when?– sequential ordering– partially solved problem must be passed on to next in
line
Types of cognitive coordination 2
Parallel and Sequential Coordination
QuickTimeª and aNone decompressor
are needed to see this picture.roofing plastering
laying
electricity
plumbing
paintingparallel
sequential
• Agents present their "position"• Agents are influenced by the position they hear
– by shifting their own position closer• After shifting they express their new position• This goes on for several rounds• Suppose more agents have position A
– → everybody moves a little closer to position A than to rival positions
– in the next round, A will be expressed even more compared to its rivals
– → further movement towards A– eventually, everybody moves towards A
Emergence of Consensus
• homogeneization– largest group tends to grow because of positive
feedback– dissenters tend to give in (conformist pressure)
• partitioning in distinct subgroups– minorities communicate more among each other than
with the majority– e.g. under geographical or social separation– they will reinforce the majority opinion
Results of self-organization
Example: Magnetization
Example: Magnetization
QuickTimeª and aTIFF (Uncompressed) decompressor
are needed to see this picture.
• Majority becomes more radical after homogeneization– as agents’ position shifts towards the majority position, they
think of more arguments supporting that position– Therefore, everybody feels more confident to move further
in that direction• Results in “Groupthink” (Janis, 1972)
– Tendency to overlook important minority arguments, and make unrealistic collective decisions
• Examples– Failed Bay of Pigs invasion of Cuba by the US– Saddam’s non-existent "weapons of mass destruction”
Polarization
Polarization: simulation
QuickTimeª and aTIFF (Uncompressed) decompressor
are needed to see this picture.
• Homogeneization is good when– all positions are about equally valuable– but it is better to agree on one standard– e.g. common symbols or conventions
• Homogeneization is dangerous when– different positions have different strengths and
weaknessess– the majority position neglects important features
• in this case we need an aggregation of all positions– avoid non-linear amplification
Advantages and disadvantages
• individuals need to agree about how to call something
– e.g. when together solving a problem involving that thing
– first one proposes a long description
– the other one confirms he got it, or asks further questions
– when it is clear they both agree, they tend to shorten the description
• Examples
– the Garrod experiments: together move in a maze
– the Talking Nets simulation
Developing Common References
Developing a common reference
QuickTimeª and aTIFF (Uncompressed) decompressor
are needed to see this picture.
• agents need to develop a common vocabulary
– unambiguous collection of meaning → name mappings, e.g.
• square uxp→
• circle bakt→
• triangle hdi→
– minimize homonyms, synonyms, and disagreements• In the simulation, a shared lexicon emerges after many
agent-agent interaction– Interaction fails when misunderstanding, succeeds otherwise– Successful mappings get reinforced, unsuccessful ones
suppressed
Self-organization of a Lexicon
The Talking Heads experiment
QuickTimeª and aTIFF (Uncompressed) decompressor
are needed to see this picture.
Chinese Whispers
QuickTimeª and aTIFF (Uncompressed) decompressor
are needed to see this picture.
QuickTimeª and aTIFF (Uncompressed) decompressor
are needed to see this picture.
• Story told from A to B, B to C, etc.
– each transmission induces a variation of the story
– some elements are left out in the final version• "friction" within the communicative process
– some elements are amplified or added• "synergy" within the communicative process
• Natural selection of "fit memes"
– ideas adapted to the agents and their communication processes
– evolution of common knowledge• e.g. urban legends, common sense, shared culture, stereotypes, rumours…
Memetic transmission
• Simplicity– irrelevant details tend to be left out
• Coherence– ideas that do not fit existing knowledge left out
• Expressivity– ideas difficult to express in language left out
• Formality = context-independence– context-dependent meanings get lost
• Novelty– unexpected ideas are better remembered and transmitted
• Utility– Useful information is better remembered and transmitted
Memetic selection criteria
• Aggregating diverse experiences– a group of people knows more than a single individual
• Avoiding groupthink– conforming to others can amplify small deviations– To some degree counterbalanced by preference for
novelty• Group meetings may or may not exhibit
collective intelligence– CI can be boosted via Delphi procedure
•Expressing arguments independently and anonymously
Collective Intelligence
Group discussion
QuickTimeª and aTIFF (Uncompressed) decompressor
are needed to see this picture.
• Experiment set up by Heylighen & Biebaut– Data still being analysed
• Subjects get list of 20 factors contributing to happiness– must evaluate their respective importance
• Subject discuss these factors as a group• Three evaluation scores
– Subjects before the discussion
– Subjects after the discussion
– Experts in happiness research (without discussion)
The happiness discussion
QuickTimeª and aTIFF (Uncompressed) decompressor
are needed to see this picture.
• Comparison of scores before-after the discussion– 3 discussions using different methods
• Opinions became significantly more homogeneous– Consensus development, as expected
• Average opinion moved significantly closer to expert opinion– Evidence of collective intelligence
• Why was there Collective Intelligence, but not Groupthink?
Results happiness experiment
Some Hypotheses• Discussion forced subjects to
–Reflect more intensely about something they already knew
–Listen to alternative viewpoints
• List of factors made it difficult to “forget” important aspects
• Subjects knew each other ->
–Less need to establish common ground or mutual trust• A priori low friction
–More willingness to share novel information
• No final decision was demanded
–Free to brainstorm and remain independent
• Groups of individuals self-organize
• Hoped for results
– Homogeneization of positions within local community
• development of common words, rules, references, standards...
– Collective intelligence• synthesis of diverse knowledge
– Distributed cognition• Division of labor and workflow to coordinate information processing
• These mechanisms explain a great number of phenomena in communication, culture and language
– But need to be investigated in much more depth
Conclusion