A number of asymmetries in lexical memory emerge when monolinguals and early bilinguals are compared to (relatively) late second language (L2) learners. Their study promises to provide insight into the internal processes that both support and ultimately limit L2 learner achievement. Generally, theory building in L2 and bilingual lexical memory has relied on ontological models: models that posit that lexical access involves inferences made about representations. This article contrasts ontological models of bilingual and L2 lexical memory with a neural network model, the ART2-IN. Testing an ART2-IN after training it on real language samples shows that asymmetries in lexical access may have roots in the effects that different paths to bilingualism have on perceptual memory systems. The ART2-IN models both early Spanish-English bilingualism and the adult acquisition of English by a native Spanish speaker, and makes direct contact with important characteristics of L2 learning. The ART2-IN is different from ontological models in assigning causation: Ontological models assign the cause of learner asymmetries to internal structures and mechanisms, while the ART2-IN distributes cause over the learner and the learning environment. (Contains 7 notes, 6 tables and 3 figures.)