Vowel harmony, a phonological process whereby adjacent vowels share values of a phonological feature, has raised important challenges for generative phonology, particularly Optimality Theory (OT) (Prince and Smolensky, 1993/2004), a theory of linguistic typology in which output forms are computed in parallel from an infinite candidate set. The parallel nature of computations in OT, as well as the unconstrained candidate set for possible outputs poses challenges for a theory of vowel harmony, which applies in a local fashion, such that vowels share the same feature value as their nearest neighbor. Particularly, standard theories of vowel harmony in OT predict the existence of pathological vowel harmony processes that are unconstrained by locality, producing patterns that are never found in natural language. Building on the work of Turbidity Theory (Goldrick, 2001), this dissertation proposes Turbid Spreading, a theory of representations for harmony that provides a solution to the ‘myopia’ generalization in OT. Representations for features are both rich as well as constrained, making it possible to account for several aspects of vowel harmony (e.g., non-participating vowels and epenthetic vowels) without over-predicting. Evidence for the completeness of the predicted typology is provided using computational methods (i.e., finite-state machines). The cognitive bases for the typological restrictions on vowel harmony typology are verified in a series of 12 experiments using the artificial grammar learning paradigm in adults. In these experiments, English speakers are exposed to mini versions of vowel harmony languages, followed by a forced-choice comprehension test. This test contains novel items as well as items from the training set. In particular, several novel test items include novel representations (e.g., novel vowels), which have been specifically held out from training to test. This ‘poverty of the stimulus’ method (Wilson, 2006) makes it possible to test learners' bi inferences towards ambiguous stimuli. The results of these experiments suggest that learners' biases conform to the cross-linguistic typology of harmony languages. Learners are biased to learn harmony patterns that are frequently occurring and phonetically natural, but biased against rare or non-existing patterns. These findings support the hypothesis that typological restrictions are grounded in learning biases.