Daphna Weinshall
(http://www.cs.huji.ac.il/~daphna/)
Wednesday 19th September 2012
Time: 2pm
4th Seminar Room
Alexandra House, 17 Queen Square, London, WC1N 3AR
Category Learning from Equivalence Constraints: theory, behavior and neural basis
I will talk about the process of comparing exemplars, both for improving performance by artificial (computer) classifiers as well as for explaining human category-learning strategies. We investigated human category learning from positive and negative equivalence constraints - being informed that two exemplars belong to the same category or to different categories, respectively. We found that people can use the two types of constraints effectively, but the nature of learning is different. It appears like the use of negative constraints enables a less intuitive but potentially more accurate categorization strategy. Developmental study showed that children use negative constraint effectively at a later age than positive constraints, while an f-MRI study showed different areas to be involved in the two learning processes. Computationally, we developed algorithms which learn a variety of distance functions between samples using such equivalence constraints. Distance functions can be useful for such applications as data retrieval or collaborative filtering. If time permits, I may discuss how distance functions can be used for the modeling of physiological data.