Curious about why the model captures these results? Check out my extended abstract in the proceedings of SCiL 2025:
doi.org/10.7275/scil...

And to read more about PFA, check out this paper in Linguistic Inquiry:
doi.org/10.1162/ling...
6/6

I found that when you use PFA with a MaxEnt model and train it on Types I-VI, II starts off harder than IV, but becomes easier for the model later in learning. Since lab learning involves less exposure than natural language acquisition, this could explain the mismatch in M&P's (2014) results.
5/6

Probabilistic Feature Attention (PFA) can potentially explain this. PFA adds ambiguity to the learning process, randomly sampling which features a model can attend to (similar to dropout in neural networks). E.g., if the model doesn’t attend to [voice], it can’t distinguish between [t] & [d].
4/6

...And found that Type II phonotactic patterns were harder for participants to acquire than Type IV patterns. But when they looked at typological data, Type II phonological patterns were more common than Type IV. But why does typology seem to favor the more difficult pattern?
3/6

Shepard Types (I-VI below) have been used widely in category learning since they were introduced by Shepard et al. (1961). These assume a stimulus space with 3 features, 8 stimuli, and 6 possible ways to halve that space. Moreton & Pertsova (2014) adapted these to the domain of phonotactics...
2/6

Also saw this as a kid and was scarred for life

It's one of the unwritten rules of civilization: if a child pours air into a cup and offers you tea, you will sip it and say "thank you, it's delicious."

The amount of green food you could buy in grocery stores that summer was astounding.