Name
Abstract | ||
|---|---|---|
Manual generation of training examples for supervised learn- ing is an expensive process. One way to reduce this cost is to produce training instances that are highly informative. To this end, it would be beneficial to produce training instances inter- actively. Rather than provide a supervised learning algorithm with one complete set of training examples before learning commences, it would be better to produce each new training instance based on knowledge of which instances the learner would otherwise misclassify. Whenever the learner receives one or more new training examples, it should update its clas- sifier incrementally and, in real time, provide the teacher with feedback about its current performance. The feasibility of such an approach is demonstrated on a realistic image pixel classification task. Here, the number of training instances in- volved in building a classifier was reduced by several orders of magnitude, at no perceivable loss of classification accu- racy. |
Year | DOI | Venue |
|---|---|---|
1998 | 10.1142/S0218001499000112 | International Journal of Pattern Recognition and Artificial Intelligence |
Keywords | DocType | Volume |
interactively training pixel classifiers,real time,supervised learning | Conference | 13 |
Issue | ISSN | ISBN |
2 | 0218-0014 | 1-57735-051-0 |
Citations | PageRank | References |
3 | 0.60 | 18 |
Authors | ||
3 | ||
Name | Order | Citations | PageRank |
|---|---|---|---|
| Justus H. Piater | 1 | 543 | 61.56 |
| E. M. Riseman | 2 | 1402 | 458.95 |
| Paul E. Utgoff | 3 | 1118 | 377.09 |