Learning new speech categories 1 Running head: LEARNING NEW SPEECH CATEGORIES IN ADULTHOOD Success and failure of new speech category learning in adulthood: Consequences of learned Hebbian attractors in topographic maps

The influence of a native language on learning new speech sounds in adulthood is addressed using a network model in which speech categories are attractors implemented through interactive activation and Hebbian learning. The network has a representation layer that receives topographic projections from an input layer and has reciprocal excitatory connections with deeper layers. When applied to an experiment in which Japanese adults were trained to distinguish the English /r/-/l/ contrast (McCandliss et al., 2002), the model can account for many aspects of the experimental results, such as the time-course and outcome of the learning, how it varies as a function of feedback, the relative efficacy of adaptive and initially easy training stimuli versus nonadaptive and difficult stimuli, and the development of a discrimination peak at the acquired category boundary. The model is also able to capture some aspects of the individual differences in learning. Learning new speech categories 3 Success and failure of new speech category learning in adulthood: Consequences of learned Hebbian attractors in topographic maps One of the key issues a theory of perception must address is the effect of prior experience. Newborn infants initially have the ability to distinguish a rich variety of linguistic contrasts, but by about six months of age, their ability has begun to get attuned to their native language (Kuhl et al., 1992). This reshaping of the perceptual space presumably allows the developing infants to recognize their language sounds more effectively, and may contribute to warping of perceptual representations toward category prototypes (the perceptual magnet effect, Kuhl, 1991, 2000), and to the sharpening of sensitivity at category boundaries (the categorical perception effect, Liberman et al., 1957). However, it has been suggested that such reshaping may have the side effect of hindering the acquisition of language distinctions later in life (Flege, 1995). In this paper, we explore a mechanistic account of how first-language speech acquisition can influence the later acquisition of second-language speech sounds. Our exploration is guided by the assumption that speech learning is an instance of more general architectural and computational principles. We consider four such principles and evaluate their usefulness in addressing data from an experiment on the acquisition of a nonnative speech contrast in adulthood. Specifically, we use the principles to develop a computational model of the pattern of successes and failures in learning American English /r/ and /l/ by adult native speakers of Japanese (McCandliss et al., 2002). Models of perceptual learning often attempt to reproduce only the end state of learning, and frequently this does not Learning new speech categories 4 place sufficient constraints on the models (cf. Damper & Harnad, 2000, Edelman & Intrator, 2002). In order to impose additional constraints, we evaluate our model against several measures over the time course of learning and under different training conditions, and explore the extent to which the principles can account for a wide range of findings. The rest of the paper is organized as follows. First, we will give an overview of the R/L problem and of the McCandliss et al. (2002) results, pointing out aspects of the results that appear puzzling in the absence of an explicit mechanistic explanation. Next, we introduce and justify the principles underlying our model, and present an abstract and simplified version to illustrate its basic properties. Following that, we set out our model of the McCandliss et al. data and present the modeling results together with the experimental results. Finally, we evaluate the successes and limitations of the model and the consequent implications for the architectural principles.