Swarm Intelligence: A Reading Note
Chapter 3: Connections Between People
The previous sections explored structures for processing representations, but to endow machines with intelligence, they must possess the ability to learn and adapt. In the application of artificial intelligence, symbolic learning involves adding representations, whereas connectionist learning involves changing the weights of connections or the structure of the network. Machines can modify their parameters in response to feedback from the external environment, a process known as Reinforcement Learning. This mirrors the theory of operant conditioning in humans, where individuals learn and modify their behavior based on rewards and punishments. However, earlier behaviorism only considered stimulus-response without delving into how cognition changes. With advancements in biomedicine, many studies discovered the biological basis of learning mechanisms. Long-term potentiation in neurons changes the strength of neural connections, and the plasticity of neurons can alter neural structures. These findings resonate with the feasibility of neural network operations, where changing connections is a crucial part of the learning process.
The connections within a group can be represented by a graph, composed of nodes and edges, as illustrated below. Hopfield (1982) invented a neural network based on this graph structure. Nodes represent binary values, and the connections between them have weights. Nodes update their stored values based on inputs from connected nodes and corresponding weights until a stable state is reached. This model can simulate the mental states of humans. Each node represents a belief, attitude, or behavioral pattern, and the connections represent our cognitive structure. When input values and weights conflict, the neural network continually updates its values and cannot reach a stable state. This requires a change in weights to balance the network, similar to how we resolve cognitive dissonance by altering our cognition to achieve harmony.
Moreover, in culture and society, interactions between people can also be represented as a graph. Humans learn and adapt to their environment through social learning. According to Bandura's (1977) social learning theory, human learning is an interactive process between the individual and the social environment, leading to socialization. This theory emphasizes that human behavior is influenced not only by environmental factors but also by personal cognition of the environment. Social psychological research shows that human behavior is deeply influenced by social schemas. We understand others' beliefs through social interactions and revise our own beliefs. Social impact theory (Latané, 1981) attempts to quantify the degree of social impact on an individual with three indicators: the strength of the influencer, the proximity of the influencer, and the number of influencers. In the context of a Hopfield neural network, a node's value is influenced by the magnitude of connected nodes, the distance between nodes, and the number of connected nodes. Importantly, even if two points are not directly connected, changes in values can be transmitted to distant nodes through the network. If we view these values as mental states, our beliefs, attitudes, and thoughts influence each other and propagate through the network, potentially leading to a consensus. However, most beliefs in society still conflict, causing constant societal change.
But why do some ideas spread rapidly and become mainstream societal consensus, while others fade away in the course of social change? Neural network structures alone cannot fully explain the cultural propagation of beliefs. Next, we will introduce another random system—evolution—and use evolutionary theory to explain the changes in social culture.
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.