Adaptive Learning With Surrogate Assisted Training Models Using Limited Labeled Acoustic Sample Sequences
Guilherme Zucatelli, R. Coelho · 2021
In this paper, an adaptive learning solution based on surrogate models is investigated under reverberant scenarios. A new surrogate selection criteria is proposed, leading to a higher discrimination among models. The method is evaluated considering a classic source classification approach with ROC and AUC analysis. Furthermore, the Bhattacharrya distance is adopted to measure the separability of selected signals in the feature domain, whereas the sparse coding capability of each selected model is evaluated with the K-SVD. Results show that the proposed solution improves classification accuracy and class separability while providing a reduction on sparse coding reconstruction error for all scenarios. Further experiments with pH feature vector fusion improved the classification accuracy of adaptive learning solutions.