Shedding light on social learning

Culture involves the origination and transmission of ideas, but the conditions in which culture can emerge and evolve are unclear. We constructed and studied a highly simplified neural-network model of these processes. In this model ideas originate by individual learning from the environment and are transmitted by communication between individuals. Individuals (or "agents") comprise a single neuron which receives structured data from the environment via plastic synaptic connections. The data are generated in the simplest possible way: linear mixing of independently fluctuating sources and the goal of learning is to unmix the data. To make this problem tractable we assume that at least one of the sources fluctuates in a nonGaussian manner. Linear mixing creates structure in the data, and agents attempt to learn (from the data and possibly from other individuals) synaptic weights that will unmix, i.e., to "understand" the agent's world. For a variety of reasons even this goal can be difficult for a single agent to achieve; we studied one particular type of difficulty (created by imperfection in synaptic plasticity), though our conclusions should carry over to many other types of difficulty. We previously studied whether a small population of communicating agents, learning from each other, could more easily learn unmixing coefficients than isolated individuals, learning only from their environment. We found, unsurprisingly, that if agents learn indiscriminately from any other agent (whether or not they have learned good solutions), communication does not enhance understanding. Here we extend the model slightly, by allowing successful learners to be more effective teachers, and find that now a population of agents can learn more effectively than isolated individuals. We suggest that a key factor in the onset of culture might be the development of selective learning.