Revealing Invisible Landscapes
Daniel Steinbock
Stanford University
| pag. 1
Outline
• High-level introduction
• Theory behind Particle Flow Networks
• A simple example
| pag. 2
Landscape Models
• Intuitive metaphor for complex systems
• Non-homogeneous state space
| pag. 3
Physical Systems
• Minimize on potential energy surface
• e.g. ball rolling in a bowl, water flowing downhill
| pag. 4
Genetic Systems
•
•
•
•
•
| pag. 5
Populations of living organisms
Adapting on fitness landscapes
Height of landscape = fitness
Random variation and selection
Populations move uphill as more successful
variants reproduce quicker
Social Systems
•
•
•
•
| pag. 6
Diversity of social landscapes
Seeking positive outcomes
Plans, choices, ideas => mental landscapes
Searching for people on social fitness
landscapes
Landscape Search
• Generalize random search of ordered
landscapes
• Natural landscapes are structured such that
random searches are successful
| pag. 7
Particle Flow Networks
• State space formalized as network
• Particle = random walker
• Particle swarm takes a statistical sample of
possible random walk paths
• Traces out characteristic landscape of the
underlying state space
| pag. 8
Typical Application
• Social network structure is explicitly known
• Social landscape is implicit
• Emergent product of network dynamics
| pag. 9
e.g. Trust Network
• Trustworthiness is a function of how many
people trust you and the trustworthiness of those
people
• Definition is recursive, emergent
• To calculate it for one person we need to
calculate it for the whole network
| pag. 10
Summary
• Landscape model of complex systems
• Random walks work well on naturally ordered
landscapes
• Particle flow networks simulate random walk dynamics
and reveal the landscape when only the network
structure is known
• Applicable to analysis and simulation of social networks
and other complex systems
| pag. 11
Thanks
• Thank you ECCO
• Thank you Francis & Marko
• images & audio licensed under Creative
Commons
• Send questions and comments to
daniel@sonic.net
| pag. 12
Particle-Flow Networks for
Individual and Collective
Intelligence
Marko Rodriguez, Francis Heylighen, Daniel Steinbock
| pag. 13
Outline of the Presentation
1. What is a particle-flow network?
2. General paradigm for simulating intelligence.
3. How do particle-flow networks apply to
individual-intelligence?
4. How do particle-flow networks apply to
collective-intelligence systems?
| pag. 14
Particle-Flow Networks
Part 1
| pag. 15
Particle-Flow Networks
• Network: as defined by a
set of nodes and
directed-edges.
• Particle-flow: discrete
particles which travel
through the network
performing certain
elementary functions.
| pag. 16
Particle-Flow Networks
• Edge-weight: refers to
the probability that a
particle at that node will
take that outgoing edge
at time step
t + 1.
| pag. 17
Particle-Flow Networks
• Energy Content: the amount of
energy currently in the particle
• Decay-Scalar: the percentage of
energy lost each time-step
• Initial-Node: the node which created
the particle
• Current-Node: the current location
of the particle
• Path-Length: the amount of edges
the particle has traversed
| pag. 18
Particle-Flow Networks
• Particle-storage: refers
the amount of particles in
a node at time step t.
• Flow-amount: refers to
the amount of energy that
has flowed through a
node over the period
[0, 1, …, t ]
| pag. 19
Particle-Flow Networks
• Attractivity (a.k.a.-sink):
refers to the probability
that a node will hold a
particle at time step t. An
attractivity value of 100%
means that none of the
particles that reach the
node ever leaves it.
| pag. 20
Particle-Flow Networks
• Programmically each particle is endowed with its
own send() and recv() function.
| pag. 21
public void send(Node currentNode){
pathLength++;
energyContent = energyContent * decayScalar
currentNode.flow = currentNode.flow + energyContent;
currentNode.storage = currentNode.storage++;
}
public void recv(Node currentNode){
if(RANDOM > currentNode.activity) {
// view currentNode.outgoingEdges and make a hop
currentNode.storage = currentNode.storage--;
}
}
General Paradigm for
Intelligence
Part 2
| pag. 22
Cognition
• Capability to infer from experienced to as yet not
experienced phenomena.
– Prediction, anticipation
– Imagination, conception
– Planning, problem-solving, decision-making
• General form: input -> output
– input = problem, condition, perception, present
information...
– output = solution, action, interpretation, expectation...
| pag. 23
Knowledge
• Knowledge = collection of “if…then” rules
– A -> B, B -> C, B -> D, D -> E, E -> A, E -> F, …
– Connections can be: deductive, abductive, semantic,
causal, probabilistic, associative...
– Examples
• banana -> fruit
• drop stone -> stone falls
• winter -> snow
• dog -> cat
| pag. 24
Knowledge Network
• Rules determine a weighted network
– Nodes A, B, C... = concepts, categories, distinctions
– Links A -> B = expectancy of B, given A
– Weights = degree of expectation or conditional
probability
• Learned through experience
| pag. 25
Cognitive processing
• Making inferences with complex inputs
• Input = Different nodes are activated to different
degrees
– Activation propagates along links
• Activating new nodes
– activations combine and interact
• Output = nodes in which most activation settles
| pag. 26
Individual Intelligence
Part 3
| pag. 27
Individual Intelligence
• Intelligence = problem-solving ability
• Intelligence (quantitative) = efficiency with which
network finds good solutions
– Spreading activation is a very demanding process
– Activation propagates along links
• activations diffuse, combine and interact
– Energy dissipates
• -> not enough may remain to activate best solution
| pag. 28
Individual Intelligence
• IQ test: measure of fluid intelligence
• Example questions:
– Which one is most like the first word?
• love : death, hate*, beginning, family
– Which word of the second list best fits in the first list?
• touch, taste, smell, see : cry, swim, climb, hear*
– Which of the following is least like the others?
• dog, car*, bird, fish
| pag. 29
WordScore: an IQ simulator
• Uses particle flow network to solve test
–
–
–
–
Network = based on Word Association data
Given words = Initial nodes
Potential solution words = Sinks
Answer = sink that collects most/least particles
• Gets about 75% correct = 3 x better than chance
– About average IQ for 12 year-old?
– Improves/worsens depending on parameter settings
• Decay rate, number of particles, link strengths...
| pag. 30
Demonstration
| pag. 31
Collective Intelligence
• Part 4
| pag. 32
Collective Intelligence
• Collective Intelligence = distributing problemsolving over many individuals
– selecting right person to tackle each (sub)problem
• Network representation
– Nodes= individuals
– Links = trust or knowledge relationships
– Flow = propagating questions to the right individuals
| pag. 33
Homophilic Networks
• In word-networks edges connect similar words
(similarity by association).
• In social-networks edges connect similar people.
– Friendship networks: amicability
– Trust networks: opinions/perspectives
– Co-authorship networks: expertise
| pag. 34
The Collective Mental Map
• A collective of individuals creates a footprint of
activity which can be used as a map of the
community.
• Particle-swarms allow you to search that map to
interact with individuals.
– Provide them user specific information
• Problems or Solutions
– Provide them decision-making influence
• Problem-Solving Influence
| pag. 35
Opinion-Based Representative
Decision-Making System
• Given a particular opinion poll, if all individuals of the
society participate in the decision-making process, the
result is X.
• Given any subset of the group, is the decision derived
by this subset still X?
SOLUTION: A method to holographically represent the
collective’s decision-making behavior within any subset
of the collective.
| pag. 36
Full Participation
= Active Voter (A = 100%)
Decision = (0.8 + 0.5 + 0.8 + 0.9) / 4 = 0.75
Goal is to achieve this value as
voter participation wanes.
| pag. 37
Waning Participation
Decision = (0.5 + 0.9) / 2 = 0.7
Error = 0.05 = | 0.7 – 0.75 |
| pag. 38
Simulation on a population of
1,000
0.25
Decision Error
0.2
0.15
0.1
0.05
0
100
95
90
85
80
75
70
65
60
55
50
45
Active Population
| pag. 39
40
35
30
25
20
15
10
5
Trust-Networks
edge(i,j) = 1 - │opinion(i) – opinion(j) │
| pag. 40
Trust-Networks
Two members of the community are
voicing their opinion on a particular topic.
| pag. 41
Trust-Networks
100
100
100
100
| pag. 42
Trust-Networks
125
100
175
| pag. 43
Trust-Networks
150
Decision = (250 * 0.9) + (150 * 0.5) / 400 = 0.75
250
Error = 0.00 = | 0.75 – 0.75 |
| pag. 44
Simulation on a population of
1,000
0.25
Decision Error
0.2
0.15
0.1
K=0
K~3
0.05
Active Population
| pag. 45
1
4
7
10
13
16
19
22
25
28
31
34
37
40
43
46
49
52
55
58
61
64
67
70
73
76
79
82
85
88
91
94
97
100
0
Benefits of K~3 Network
• Since K~3 networks are the most optimal, individuals in
the group need only know 3 other individuals to create a
good model of the whole. Practical in terms of human
what we see in already existing social-networks.
• If K=N-1 was the most optimal network then this would
be less promising since everyone would need to relate to
everyone in the group. Fully connected social-networks
do not appear in nature.
| pag. 46
Demonstration
| pag. 47
The Problem Domain
• What about representative networks that are not
‘opinion’-based, but more ‘expert’-based?
• How should the network account for the context
of the problem?
• Should everyone have the same initial
distribution of particles for decision-making?
| pag. 48
Subset Mapping
• Whole to Subset Mapping: Modeling the whole of the
network within a subset of the whole.
| pag. 49
Subset Mapping
• Subset to Subset Mapping: identifying the most
representative nodes relative to a particular input subset.
The domain is the initial subset.
| pag. 50
Collective Peer-Review
Process
• Problem: Should manuscript X be published for the
community?
• Problem-Routing: Which members of the community
should review manuscript X?
• Problem-Solving Influence: Of those individuals what
is the relative influence each member should have?
[DEMONSTRATION]
• Solution: The communities decision on manuscript X.
• Solution-Routing: Which members of the community
would be interested in the published manuscript X.
| pag. 51
Peer-Review DecisionMaking
• Given an unpublished manuscript, determine the
amount of ‘influence’ each member of the
community should have regarding accepting or
rejecting the manuscript.
– Who should have more decision-making influence? A
Nobel Laureate in Chemistry or a less-renowned
computer-scientist?
• SOLUTION: Depends on the manuscript domain
| pag. 52
Co-Authorship Networks
• When two scientist coauthor a paper an edge
between them is
created within the
scientific communities
co-authorship network.
| pag. 53
Co-Authorship Networks
• Bollen: Digital-Libraries & Impact
Rating Methods
• Hussel: Co-Authorship Network
Visualization
• Nelson: Impact Rating Methods
• Luce: Digital-Library Architectures and
Measures
• Van de Sompel: OAI-PMH & CoAuthorship Networks
• Vemulapeli: Digital-Library
Architectures
• Marks: Co-Authorship Networks
• Liu: OAI-PMH & E-Print Architectures
| pag. 54
Demonstration
References
[1] Smith, J.
[2] Guy, L.
[3] Man, P.
| pag. 55
Results
Referee Name
References
[1] Smith, J.
[2] Guy, L.
[3] Man, P.
| pag. 56
FA
Influence
Recent Interests Related to Paper
Sompel, HV
0.09844
OAI-PMH and Co-Authorship Networks
Bollen, J.
0.08594
Digital-Libraries and Network-Based Impact Metrics
Carr, L.
0.08516
Digital-Libraries and Open Archive Services
Hall, W.
0.08066
Knowledge Management and Digital-Libraries
Rocha, L.M.
0.07892
Document Recommendation Systems
Lagoze, C.
0.05328
Digital-Library Architectures and Services
Harnad, S.
0.04883
Open Citation Linking and Digital-Library Architectures
Hitchcock, S.
0.04177
Electronic Journals and Citation Linking
Blake, M.
0.04156
OAI Repositories and Citation Linking
Jiao, Z.
0.03386
E-Print Services
Bergmark, D.
0.03262
Digital-Libraries and OAI-PMH
Miles-Board, T.
0.02049
Digital-Libraries
Davis, H.C.
0.01211
Digital-Libraries and Adaptive Linking
Roure, D.D.
0.01125
Dissemination of Scientific Information Services
French, J.C.
0.01081
Digital-Library Distributed Searching and Interfaces
Brody, T.
0.00986
OAI-PMH and Open Citation Linking
Conclusion
• Good life…
| pag. 57