Human infants have the remarkable ability to learn any hu-man language. One proposed mechanism for this ability isdistributional learning, where learners infer the underlyingcluster structure from unlabeled input. Computational mod-els of distributional learning have historically been principledbut psychologically-implausible computational-level models,or ad hoc but psychologically plausible algorithmic-level mod-els. Approximate rational models like particle filters can po-tentially bridge this divide, and allow principled, but psycho-logically plausible models of distributional learning to be spec-ified and evaluated. As a proof of concept, I evaluate one suchparticle filter model, applied to learning English voicing cate-gories from distributions of voice-onset times (VOTs). I findthat this model learns well, but behaves somewhat differentlyfrom the standard, unconstrained Gibbs sampler implementa-tion of the underlying rational model.

In a noisy but structured world, memory can be improvedby enhancing limited stimulus-specific memory with statisti-cal information about the context. To do this, people have tolearn the statistical structure of their current environment. Wepresent a Sequential Monte Carlo (particle filter) model of howpeople track the statistical properties of the environment acrossmultiple contexts. This model approximates non-parametricBayesian clustering of percepts over time, capturing how peo-ple impute structure in their perceptual experience in order tomore efficiently encode that experience in memory. Each trialis treated as a draw from a context-specific distribution, wherethe number of contexts is unknown (and potentially infinite).The model maintains a finite set of hypotheses about how thepercepts encountered thus far are assigned to contexts, updat-ing these in parallel as each new percept comes in. We applythis model to a recall task where subjects had to recall the posi-tion of dots (Robbins, Hemmer, & Tang, 2014). Unbeknownstto subjects, each dot appeared in one of a few pre-defined re-gions on the screen. Our model captures subjects’ ability tolearn the inventory of contexts, the statistics of dot positionswithin each context, and the statistics of transitions betweencontexts—as reflected in both recall and prediction.

Speech perception is made much harder by variability betweentalkers. As a result, listeners need to adapt to each differenttalker’s particular acoustic cue distributions. Thinking of thisadaptation as a form of statistical inference, we explore the rolethat listeners’ prior expectations play in adapting to an unfa-miliar talker. Specifically, we test the hypothesis that listenerswill have a harder time adapting to talkers whose cue distribu-tions fall outside the range of normal variation across talkers.We also show that it is possible to infer listeners’ shared priorexpectations based on patterns of adaptation to different cuedistributions. This provides a potentially powerful tool for di-rectly probing listeners’ prior expectations about talkers thatdoes not rely on speech produced by many different talkers,which is costly to collect and annotate, and only indirectly re-lated to listeners’ subjective expectations.

Speech perception requires ongoing perceptual category learning. Each talker speaks differently, and listeners need to learn each talker’s particular acoustic cue distributions in order to comprehend speech robustly from multiple talkers. This phonetic adaptation is a semi-supervised learning problem, because sometimes a particular cue value occurs with information that labels the talker’s intended category for the listener, but other times no such labels are available. Previous work has shown that adaptation can occur in both purely supervised (all labeled) and purely unsupervised (all unlabeled) settings, but the interaction between them has not been investigated. We compare unsupervised with (semi-) supervised phonetic adaptation and find, surprisingly, that adult listeners do not take advantage of labeling information to adapt more quickly or effectively, even though the labels affect their categorization. This suggests that, like language acquisition, phonetic adaptation in adults is dominated by unsupervised, distributional learning.