Nelson, Jonathan
Coauthors(s): Javier R. Movellan Cognitive science department and INC, UCSD
UCSD
Cognitive science

cogsci.UCSD.edu/~jnelson

Inference by means of uncertainty

Does a model's good fit with particular behavioral data imply that another model could not fit those data as well? One way to test a model of concept learning is to compare people's beliefs, given certain information, with the predictions of the model. For instance, one can ask whether a model effectively matches subjects' beliefs about what novel objects are most likely to be consistent with an unknown true concept, given some examples of objects that are consistent with that concept (e.g., Tenenbaum, 2000). But the best means of assessing a model's fidelity to people's beliefs are not always clear. We show how the assumption that people's queries are reasonable -- for instance, that people don't ask questions for which they already have unequivocal answers -- can provide insight into, and improve the modeling of, people's beliefs.

---

close this window