anytime algorithms for learning anytime classifiers saher ... - Technion
anytime algorithms for learning anytime classifiers saher ... - Technion
anytime algorithms for learning anytime classifiers saher ... - Technion
Create successful ePaper yourself
Turn your PDF publications into a flip-book with our unique Google optimized e-Paper software.
<strong>Technion</strong> - Computer Science Department - Ph.D. Thesis PHD-2008-12 - 2008<br />
Learning<br />
Classification<br />
$$<br />
$$$$<br />
$$<br />
$$$$<br />
Training Data<br />
- +<br />
-<br />
+ -<br />
- -<br />
+<br />
+ - -<br />
- - + +<br />
-<br />
+<br />
+<br />
-<br />
contract<br />
<strong>learning</strong><br />
pre-contract<br />
classification<br />
contract<br />
classification<br />
New case<br />
?<br />
Anytime<br />
Learner<br />
Anycost<br />
Classifier classification $$$$<br />
label<br />
interruptible<br />
<strong>learning</strong><br />
interruptible<br />
Figure 2.1: Overview of our proposed framework. The training examples are provided<br />
to an <strong>anytime</strong> learner that can trade computation speed <strong>for</strong> better <strong>classifiers</strong>.<br />
The <strong>learning</strong> time, ρ l , can be preallocated (contract <strong>learning</strong>), or not (interruptible<br />
<strong>learning</strong>). Similarly, the maximal classification budget, ρ c , may be provided to<br />
the learner (pre-contract classification), become available right be<strong>for</strong>e classification<br />
(contract classification), or be unknown, in which case the classifier should utilize<br />
additional resources until interrupted (interruptible classification).<br />
While minimizing the total cost of classification is appropriate <strong>for</strong> many tasks,<br />
many environments involve strict constraints on the testing costs that a learner<br />
can use. To operate in such environments, the classifier needs to do its best<br />
without exceeding the bounds. There<strong>for</strong>e, <strong>classifiers</strong> that focus on reducing the<br />
total cost may be inappropriate. Like the bound on the <strong>learning</strong> process, the<br />
bounds on classification costs can be preallocated and given to the learner (precontract<br />
classification), known to the classifier but not to the learner (contract<br />
classification), or determined during classification (interruptible classification).<br />
Note that unlike cost-sesitive learners that attempt to minimize the total cost of<br />
classification, bounded classification does not require MC and ρ c to be provided<br />
in the same currency. This allows us to operate in environments where it is not<br />
possible to convert from one scale to another, <strong>for</strong> example time to money or<br />
money to risk.<br />
Figure 2.1 gives an overview of the different possible scenarios <strong>for</strong> <strong>learning</strong><br />
and classification. More <strong>for</strong>mally, let E be the set of training examples, A be<br />
the set of attributes, AC the attribute costs, and M be the misclassification cost<br />
matrix, and h the learned hypothesis. Further, let ρ l be the allocation <strong>for</strong> <strong>learning</strong><br />
resources and ρ c denote the limit on classification resources <strong>for</strong> classifying a case<br />
15<br />
$$