Abstract
This paper suggests a class of stochastic collective learning processes exhibiting very irregular behavior. In particular, there are multimodal long run distributions. Some of these modes may vanish as the population size increases. This may be thought of as "bubbles" persistent for a finite range of population sizes but disappearing in the limit. The limit distribution proves to be a discontinuous function of parameters determining the learning process. This gives rise to another type of "bubbles": limit outcomes corresponding to small perturbations of parameters are different. Since an agent's decision rule involves imitation of the majority choice in a random sample of other members of the population, the resulting collective dynamics exhibit "herding" or "epidemic" features.