Political Cognition and Connectionism: Modeling Political Reasoning Daniel Schneider, Stanford University

Political Cognition and Connectionism:
Modeling Political Reasoning
Daniel Schneider, Stanford University
“Connectionism” and “Parallel Distributed Processing” describe a
group of computational models which have been applied to many
cognitive-psychological processes. This poster explains some general
concepts of connectionism, illustrates these concepts with a simple
example, provides ideas for further research and points to relevant
introductory literature.
Only a few studies have applied computational approaches to
political cognition, for example:
A Simple Political Science Example
Further Steps & Applications
Holbrook et al. (2001) proposed a ‘Asymmetric Nonlinear Model’
(ANM) of attitudes towards candidates. Citizens integrate positive
and negative beliefs into an attitude according to this equation:
A = α1
(F)m
+ α2
(U)n
+I
(A=Attitude; F=Favorable B.; U=Unfavorable B.; I=Initial Impression;
m, n, α1, α2 = parameters)
• McPhee (1963): a brief report on computer simulations of voters’
decisions in election campaigns
• Taber and Steenbergen (1995): a framework for computer
simulations and models of electoral behavior
• Models of spreading activation processes of candidate evaluation
(Boynton & Lodge, 1994; McGraw & Steenbergen, 1995; Taber &
Timpone, 1994)
• Survey responses and attitude activation (Boynton, 1995)
• Hanges, Lord, and Dickson (2000): a connectionist approach
the relationship between political leadership and culture.
Holbrook et al. estimated the parameters based on survey data:
A = 19.66(F).36 – 12.27(U).61 + 54.83
The initial impression has a slight positive offset, initial beliefs are
more relevant, and citizens seem to focus on flaws (see Lau, 1982).
A very simple connectionist approach was used to model this
relationship:
F1
F2
F3
F4
F5
Favorable
Belief
Favorable
Belief
Favorable
Belief
Favorable
Belief
Favorable
Belief
Candidate
Parallel processing: simultaneous activation of nodes which gives
rise to perception, pattern representation or decisions, rather than serial
processing of information.
Negative
U1
U2
U3
U4
U5
Unfavorable
Belief
Unfavorable
Belief
Unfavorable
Belief
Unfavorable
Belief
Unfavorable
Belief
Fig 2.: Recurrent-Network of Candidate Evaluation
The implementation has three steps:
1.Pre-Trial Training: By simultaneously activating unfavorable
beliefs with the negative-node, the network learns which nodes
represent negative beliefs (the same for positive beliefs)
Input from unit i to j:
2.Training: The network then learns about the specific respondents
beliefs: connections between beliefs and the candidate node are
trained; to replicate different respondents/conditions, different
numbers of positive and negative beliefs are activated
Netinput for unit i:
out1
W in0out1
Win0out0 Win1out0
Win1out1
3.Test: The candidate node is activated and the results on the
positive and negative node are compared: stronger positive node,
in0
in1
‘Delta Rule’ learning:
positive evaluation; stronger negative node, negative evaluation
(50 different ‘respondents’, random order, noise, and starting
∆wij = [ai(desired)-ai(obtained)] aj ε
weights).
Pre-trial training modeled negativity with increased activation for
Activation (a) is created by a function (e.g., linear,
Fig 1.: A Simple
negative beliefs. Positive offset is modeled by one initial training
sigmoid, binary,…) of the netinput.
Feed-Forward Network
connection between the positive-node and the candidate-node.
out0/out1:
Output Nodes
Hebbian learning (not shown here) uses
A replication of the non-linear regression by Holbrook et al. was
in0/in1:
Input
Nodes
correlations in activations (‘fire together, wire
Wij:
Connections run on the data points generated by the model:
together’) and does not require a ‘desired’
.76 – .11(U).61 + .028
weights
A
=
.05(F)
activation.
from i to j
General usefulness of the model is confirmed (scale is irrelevant).
netinputi=Σjajwij
• media priming effects: understanding the role of accessibility
• stronger connection between political psychology, social
psychology and investigations in neuro-foundations of political
behavior.
Advantages of Connectionist Models
• Constraint satisfaction
Figure 1 shows a simple feed-forward network with two input nodes,
two output nodes and connections between all nodes (see McLeod et
al. 1998). In a feed-forward network information is passed on in only
one direction, no back-propagation takes place.
out0
• online vs. memory-based impression formation: modular
approach with separate memory and online-tally-systems and a
connectionist ‘handler’ that can move information from one
system to the other
• Memory access by content: activation of any node will send
activation to other nodes (McClelland, 1981)
(all nodes are interconnect, some arrows are omitted)
inputij=ajwij
• some possible future applications:
• Distributed representations are neurologically plausible, fault
tolerant, and damage resistant (‘graceful degradation)
Positive
Distributed processing: non-localized representation of information:
not individual nodes (or units) represent information, but information is
stored in connections between nodes (which in turn are uninterpretable
on their own).
• does not replicate all features of the ANM; some features are not
tested in the Holbrook et al. (2001) paper (e.g., order of presentation
of facts rather than unordered recall) or not tested experimentally.
(see McLeod et al., 1998)
+
What is Connectionism or Parallel Distributed
Processing (PDP)?
• this example directly implements specific characteristics of the
ANM (e.g., stronger pre-training trials for negative beliefs) to achieve
the desired outcome
• No distinction between ‘processing’ and ‘memory’ (interesting
opportunities for a model of OL/MB candidate evaluation)
Recommended Software & Reading
Software
• FIT2 by van Overwalle: http://www.vub.ac.be/FIT/
• PDP++: http://www.cnbc.cmu.edu/Resources/PDP++/
• Old PDP tools (DOS):
http://www.cnbc.cmu.edu/~jlm/distrib/pdp-DOS.zip
• New PDP tools (MATLAB): in development
http://psych.stanford.edu/~jlm/
Suggested Readings
• Rumelhart & McClelland (1986): Parallel Distributed Processing
(Volume I); McClelland & Rumelhart (1986): Parallel Distributed
Processing (Volume II) – classical ‘bible’
• McLeod, Plunkett, Rolls (1989): Introduction to Connectionist
Modelling of Cognitive Processes
• Smith (1996): What Do Connectionism and Social Psychology Offer
Each Other? (JPSP)
• van Overwalle (2007): Social Connectionism: A Reader and
Handbook for Simulations - collection of papers using FIT2
• O’Reilly & Munakata (2000): Computational Explorations in
Cognitive Neuroscience – broad overview, includes software.