Experiment 11

Local learning and clustering

Note: The EM algorithm for clustering with random initialization seems to perform better than we disclaimed in the paper, if the number of training examples in each cluster is not too small. See page 1 and page 2. We now feel that initialization from random clustering is better and more convincing in terms of cluster discovery.

See also Experiment 17, which allows local shifts of templates.

In experiment 11, we first perform local learning, and select the starting templates. Then we perform clustering from the starting templates.

(1) data, codes, and readme for local learning and clustering (November 2009)

Number of elements in each template is 50. Number of neighbors for local learning is 10. Number of clusters is set at 3. eps1 eps2 eps3

(2) another example (November 2009)

Partial view: Number of elements in each template is 50. Number of neighbors for local learning is 40. Number of clusters is set at 2. eps1 eps2

Full view of all the training images

Number of elements in each template is 50. Number of neighbors for local learning is 40. Number of clusters is set at 2. eps1 eps2

(3) another example (November 2009)

Number of elements in each template is 60. Number of neighbors for local learning is 20. Number of clusters is set at 8. eps1 eps2 eps3 eps4 eps5 eps6 eps7 eps8

Back to active basis homepage