Active learning is great because it automatically selects samples for labeling, which makes labeling work less and purposeful. Previous work used margin (linear classifier) to determine the necessity for labeling a particular data point, while this paper makes the margin probabilistic, and select the margin that makes decision most uncertain as the data to be labeled (Then it is the same as minimum margin. Confused).
It is an online active learning. In each iteration, the parameters for the Gaussian pdf of the margin are updated. And a data point is selected to be labeled. After that, a new model is retrained and the process is repeated. But the drawback is the annotator needs to wait near the computer...
Reviewed by
zzb3886
as
- 2008-07-03 22:26:30
Search
Reviews
[Write a review of this article]
Find related articles from these CiteULike users
