Learning biases

Why biases (constraints)?

• Quine's Problem: finding the referent
• Making generalizations on the basis of few examples
• Category (meaning) space, positive and negative evidence, and constraints
  

Kinds of constraints

• Constraints on inter-lexicon relationships: Mutual Exclusivity
  • How it could help
  • The problem of taxonomies
  • The problem of bilinguals
• Constraints on kinds of categories (meanings)
  • Whole object
  • Taxonomic vs. thematic relations
  • "Dimensional" biases: shape, material

The Shape and Material Biases

• The experiments
• How they could help
• Where do they come from?
• Should we expect it in all languages?
• Second-order (case-based) learning in the Shape Bias experiment
  1. Store a lot of episodes in LTM, including utterances with sounds and referents.
  2. (Training object presented.) This is a dax. Create an episode for the new sound and referent.
  3. (Test objects presented.) Show me the dax. Create episodes for each test object.
  4. To see if any of the test episodes make "better sense", find their mean similarity with episodes in LTM.
  5. If this doesn't work, create a tentative "mini-category" for each pair of test episode and training episode, a blend that includes shared features only. Replace "coincidental" values with variables so we will be looking for similarity not with this category but with categories of this type.
  6. Find the mean similarity of the mini-category with pairs of episodes in long-term memory. Pick the referent in the test episode for the the mini-category with the highest similarity.

© 2006. Indiana University and Michael Gasser. www.indiana.edu/~gasser/Q450/biases.html 2006-10-10