GSO-2010: Richard A. Hullinger, John K. Kruschke, and Peter M. Todd
Evolution of Attention in Learning
A variety of phenomena in associative learning suggest that people and some animals are able to learn how to allocate attention across cues. The common view of attention is that it is a necessary byproduct of limited-capacity processing: since agents cannot attend to all possible stimuli, they are forced to use attention mechanisms to reduce processing demands. We believe that this is not the complete story and that the ability to allocate attention is evolutionarily adaptive in its own right.
The current work presents the results from evolutionary simulations designed to test this hypothesis. We use a genetic algorithm to evolve simple three-layer connectionist network architectures in environments with context-dependent relevance. In these environments, one binary cue serves as "context" and two other binary cues (the "focal" cues) alternate between being diagnostically relevant and completely irrelevant based on the value of the context cue. In such environments, agents must learn to use the context cue to indicate which of the focal cues should be attended to in order to behave correctly. In these environments an agent's fitness is measured by its ability to learn quickly and accurately across multiple context shifts.
Our simulations have shown four key results. First, we have shown that evolutionary pressure to learn quickly can produce architectures that store internal representations of the cues presented to the agents. Second, we have shown that while there are multiple architectures that can learn the regularities of context environments, agents with evolved attentional mechanisms learn about their environment faster than those without attentional mechanisms. Critically, these mechanisms evolve even when agents have sufficient capacity to process all available stimuli and suffer no penalty for agent size or complexity. Therefore, limited-capacity processing is not driving the evolution of attentional learning. Third, experiments with other environments, both with and without context-relevant cues, have shown that attentional structures are unlikely to evolve unless some information in the environment alternates between being diagnostic and irrelevant. Finally, we plot the relationship between the number of context switches that an agent experiences and the rate at which attentional mechanisms develop, and show an inverted-u shaped curve. Therefore, environments where the context changes too quickly or too slowly are less likely to result in the evolution of attentional mechanisms. The third and fourth results together indicate that the adaptive benefits of attentional mechanisms come from two sources. Shifting attention toward relevant cues promotes faster learning of the environment structure. Shifting attention away from cues that were relevant in one context but are irrelevant in the current context prevents previously learned associations on those cues from being interfered with.