Swarm Intelligence: A Reading Note
Chapter 3: Connections Between People
The previous sections explored structures for processing representations. But intelligence requires more than structure; it requires the capacity to learn. In symbolic AI, learning means adding new representations. In connectionist AI, learning means changing connection weights or network architecture. Both can be driven by reinforcement learning, where machines modify parameters in response to environmental feedback. This mirrors operant conditioning: organisms reshape behavior through reward and punishment. Early behaviorism, however, stopped at the stimulus-response boundary without examining how internal cognition changes. Biomedical research filled that gap. Long-term potentiation alters the strength of synaptic connections; neural plasticity restructures the network itself. These biological mechanisms validate the connectionist approach, where modifying connections constitutes the core of learning.
The connections within a group can be represented as a graph (nodes linked by weighted edges). Hopfield (1982) built a neural network on this structure. Each node stores a binary value and updates it based on inputs from connected nodes, scaled by connection weights, until the network settles into a stable state. This model can simulate human mental states. Treat each node as a belief, attitude, or behavioral pattern, and the connections as cognitive structure. When inputs and weights conflict, the network oscillates and cannot stabilize. Restoring equilibrium requires changing weights, the same operation by which we resolve cognitive dissonance: altering cognition until internal states cohere.
This graph structure scales beyond the individual. In culture and society, interactions between people form networks, and humans learn by navigating them. Bandura's (1977) social learning theory frames learning as a reciprocal process between individual and social environment, producing socialization. Behavior reflects not only environmental pressures but the individual's cognition of those pressures (the social schemas that filter perception and guide interpretation). Latané's (1981) social impact theory quantifies this influence through three variables: the strength of the influencer, the proximity of the influencer, and the number of influencers. Mapped onto a Hopfield network, a node's value is shaped by the magnitude of connected nodes, the distance between them, and the count of connections. Even nodes without direct links transmit influence through intermediaries; changes propagate across the network. If we read node values as mental states, then beliefs, attitudes, and thoughts influence one another and cascade through social structure, sometimes converging toward consensus. Yet most beliefs in a society conflict, and that persistent tension drives continual social change.
Why do some ideas spread rapidly and crystallize into consensus while others dissolve? Network structure alone cannot explain the cultural propagation of beliefs. The next section introduces a second random system, evolution, and uses evolutionary theory to account for the dynamics of social and cultural change.
References:
Hopfield, J. (1982). Neural networks and physical systems with emergent collective computational abilities. Proceedings of the National Academy of Sciences of the United States of America, 79, 2554-8.
Bandura, A. (1977). Social learning theory. Oxford, England: Prentice-Hall.
Latané, B. (1981). The psychology of social impact. American Psychologist, 36(4), 343-356.