Simplicity and informativeness in semantic category systems

Carr, Smith, Culbertson, & Kirby (2020) doi:10.1016/j.cognition.2020.104289 [pdf, bib, osf, git]

Parameters

Bottleneck b

Exposures ξ

Noise ε

Model results show the mean and 95% confidence interval of 100 chains. The simplicity and informativeness priors use w = 1; the strong informativeness prior uses w = 500.

Typical chain

show training data

Simplicity

Informativeness

Strong informativeness