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