| Resumo: | In previous studies, a Support Vector Machine (SVM) was trained with a universe of data consisting of 3284 samples, which was divided into two sets, for training and validation, with approximately the same number of elements. Then put into practice a strategy of active learning having been obtained at the end an optimum detector for those sets. At this stage the objective was to apply the above active learning strategy, but now subjecting the detector for continuous recording on a system of sliding window, i.e., simulating a real environment detection. To prevent the unsustainable growth of training and validation sets, algorithms to reduce those sets were experimented, based on the presumed existence of redundant elements. Tests performed showed that the classifier performance improves over time when compared with the performance of the same classifier without applying the active learning strategy, given the same data. With another algorithm applied to reduce the validation set, tests showed an improvement of performance similar to that on previous experience, but without loss of performance regarding the original validation set.
|
|---|