Seminar on Computational Learning and Adaptation

  Relations between Concept Learning and Attribute Learning

Dorrit Billman
CSLI Computational Learning Laboratory
Visiting Scholar, Psychology
Stanford University

How does the structure of attributes influence concept learning and the structure of concepts influence attribute learning? Much research, both psychological and computational, has investigated learning new concepts, while much less has looked at the possibility or nature of learning new attributes. Representation of instances is typically assumed to be fixed, and representation of concepts to be built from this fixed vocabulary. We consider what it would mean to learn a new attribute, why it might be useful, how to assess whether attribute learning has taken place, and whether certain types of human concept learning do indeed produce formation of new attributes.

I will present experiments on the role of 'consistent contrast' in human concept learning and attribute formation. Consistent Contrast, is the relation among a set of contrasting concepts when the same attributes are consistently informative across the contrast set. For example, animal concepts contrast consistently the extent that people believe multiple, contrasting types of animals (monkeys, bears, horses, etc) have informative values on the same attributes (types of diets, vocalization, etc)-- monkeys eat bannana and chatter; bears eat honey and growl, etc. One set of experiments shows that concepts which contrast consistently are learned faster and support wider inductions than alternative concept structures. A second set of experiments suggests that people form new attributes, by reorganizing features into alternative values of a new attribute. Our data suggests that people form new attributes 'in order' to make newly learned concepts contrast consistently. We compare attribute formation to other types of representation change discussed in the computational and psychological literature, and consider why representation change of several types may be both valuable and difficult.

Return to the seminar schedule