anytime algorithms for learning anytime classifiers saher ... - Technion

Technion - Computer Science Department - Ph.D. Thesis PHD-2008-12 - 2008
Chapter 5
Anytime Learning of Anycost
Classifiers
Assume that a hardware manufacturer who uses a machine-learning based tool
for assuring the quality of produced chips. In realtime, each chip in the pipeline
is scanned and several features can be extracted from the image. The features
vary in their computation time. The manufacturer trains the component using
thousands of chips whose validity is known. Because the training is done offline,
the manufacturer can provide the values of all possible features, regardless of their
computation time. In realtime, however, the model must make a decision within
2 seconds. Therefore, for each chip, the classifier may use features whose total
computation time is at most 2 seconds. Alternatively, the manufacturer might
want to provide the classifier with a different maximal time, depending on the
case, or even configuring the classifier to utilize time until the next item arrives
and then querying it for a decision.
To act under these different resource-bounded classification scenarios, our
framework should produce predictive models that can control testing costs efficiently,
and should also be able to exploit additional learning resources in order
to improve the produced models.
For the first requirement, i.e., resource-bounded classification, a decision-tree
based classifier would make an ideal candidate. When classifying a new case,
decision trees ask only for the values of the tests on a single path from the root
to one of the leaves. Tests that do not appear on the actual path need not be
administered. Decision tree models are also considered attractive due to their
interpretability (Hastie et al., 2001), an important criterion for evaluating a
classifier (Craven, 1996), their simplicity of use, and their accuracy, which has
been shown to be competitive with other classifiers for several learning tasks.
Decision trees, however, cannot be used as is: when the classification budget
does not allow exploring the entire path, the tree cannot make a decision. This
99