Misclassification and Margin based SVM Active Learning Algorithm for Audio Event Detection

Audio event detection has become a hot research due to its wide applications in many fields, such as multimedia retrieval etc., the detection needs large amounts of labeled samples to train the audio event models, but in real life, the labeled samples are expensive to obtain, the shortage of such labeled samples is a big obstacle. Active learning is an efficient way to deal with the problem of insufficient labeled samples. The most popular support vector machines active learning is the margin based sampling (MBS), which is to query the sample closest to the current hyperplane, but when the current hyperplane is far away from the true hyperplane, the sample closest to the current hyperplane is not so informative, querying such samples would have a much slower adjustment of the hyperplane. In order to accelerate the adjustment, this paper proposes the misclassification and margin based sampling (MMBS) active learning algorithm. In order to query more informative samples, MMBS selects samples based on misclassified samples’ KL divergence in the first few iterations, after that, considering the lower misclassification confidence and the outlier problem, it switches to MBS. Experiments show that compared to MBS and representative sampling (RepS), MMBS can get the highest detection performance under the same human annotation workload.