Learning distributions as they come: Particle filter models for online distributional learning of phonetic categories
Human infants have the remarkable ability to learn any human language. One proposed mechanism for this ability is distributional learning, where learners infer the underlying cluster structure from unlabeled input. Computational models of distributional learning have historically been principled but psychologically-implausible computational-level models, or ad hoc but psychologically plausible algorithmic-level models. Approximate rational models like particle filters can potentially bridge this divide, and allow principled, but psychologically plausible models of distributional learning to be specified and evaluated. As a proof of concept, I evaluate one such particle filter model, applied to learning English voicing categories from distributions of voice-onset times (VOTs). I find that this model learns well, but behaves somewhat differently from the standard, unconstrained Gibbs sampler implementation of the underlying rational model.