Semi-Supervised Learning for Connectionist Networks

At the computational level, language is often assumed to require both supervised and unsupervised learning. Although we have a certain understanding of these computational processes both biologically and behaviorally, our understanding of the environmental conditions under which language learning takes place falls short. I examine the semi-supervised learning paradigm as the most accurate computational description of the environmental conditions of lexical acquisition during language development. This paradigm is assessed for task learning and generalization and I argue that its real ecological validity and occasional improvements in performance over supervised learning make it an ideal candidate for modeling of language acquisition and other learning problems