Fig. 9.

Results of Simulation 3 for the voiced VL category. The bottom panel shows that when the starting μ-values of the Gaussians were initially closer than the means in the data, the model always failed to learn a two-category solution (i.e. it over-generalized the dimension into a single category). As the starting values were moved further apart, the model became more likely to succeed. The top panel shows the μ-values over the course of learning for the different starting values. The point at which each line ends indicates the latest point during training at which one of the models for that starting value still maintained a two-category solution. All of the models that had two categories throughout training settled on μ-values near points further apart than the means in the dataset.