Posner and Keele; Rosch et al.

1
Posner and Keele Rosch et al.
2
Posner and Keele Two Main Points

• Greatest generalization is to prototype.
• Given noisy examples of prototype, prototype is
  as well classified as examples.
• Prototype isnt only thing learned.
• Amount of variability in training examples
  influences representations.

3
How the Experiments Work

• Generate prototype.
• Then distortions.
• Train subject to know one level of distortions.
• Judge ability to classify
• Training set.
• More distorted ones
• Equally distorted but different.
• Prototypes.
• Matlab

4
From Palmeri and Flanery
5
http//www.msu.edu/course/psy/200/Burns/p200f03091
6_6.pdf
6
Conclusions of Experiment 3

• Train on distorted images.
• Performance as good on prototype as training set.
• Worse on less distorted images, of equal avg.
  distance to training set as prototype.
• Even worse on new of equal distortion as training
  set.
• Representation favors prototypes.

7
Conclusions of Experiment 12

• Train on a) less or b) more distorted versions.
• Test on even more distorted.
• Better if you trained on more distorted.
• Representation is more than prototypes.

8
Questions

• What is learned?
• Prototype?
• Not just prototype.
• Examples plus prototype.
• Distribution?
• Gaussian?
• Seems like examples mattered. Could distribution
  overfit the data?
• Examples Weighted distances.

9

• Is this domain sufficiently realistic?
• Are dots just dots?
• Maybe shape is important. This might mean
  forming links between dots, which noise disrupts.
  Perhaps exps 12 have more to do with set of
  shapes learned than variance in features.
• If real categories have structure, does this
  domain have too little structure?

10
Rosch

• Categories are in the world.
• Objects have structure, correlations of
  attributes.
• Basic categories are those with greatest cue
  validity.
• Cue predicts category.
• Superordinates dont have many attributes in
  common.
• Subordinates share many attributes with other
  subordinates.
• Why cue validity?

11
Experiments 1-4

• Main points
• Categories have shared properties
• Basic categories are where most of these first
  appear.
• Experiments
• Collect attribute lists.
• Example common attributes for fruit
  superordinate 3, basic 8.3, subordinate 9.5
• Document actions with objects.
• Similar pattern of attributes.
• Similarity of shape.
• Average shape.
• Basic categories have similar shape no evidence
  subordinates are less similar.

12
Experiments 5-12

• Main point
• Basic categories are psychologically real.
• Experiments
• 5 6 Prime with category name then
  compare/detect with quick exposure.
• 7 Category name is query compared to image, not
  prime.
• 1st task where performance is best for basic
  categories.
• 8 9 Children sort into basic categories
  earlier.
• 10 People name with basic categories.
• 11 Children learn these words first.
• 12 ASL is missing many subordinate and
  superordinate words.

13
Conclusions

• Is this belaboring the obvious?
• Well, it wasnt obvious.
• Convergence of evidence more convincing?
• Categorization is for inference.
• This seems like a very modern view.
• Different kinds of attributes common at basic
  level.
• This is what makes visual classification possible
  (if true).
• Are categories completely in the world?
• May depend on extent of knowledge.
• Context.