ABSTRACT
Populations of simple recurrent neural networks were subject to simulations of evolution where the selection criterion was the ability of a network to learn to recognize strings from context free grammars. After a number of generations, networks emerged that use the activation values of the units feeding their recurrent connections to represent the depth of embedding in a string. Networks inherited innate biases to accurately learn members of a class of related context-free grammars, and, while learning, passed through periods during which exposure to spurious input interfered with their subsequent ability to learn a grammar.
