Analyzing 50 kHz rats’ vocalizations using machine learning approaches

Abstract

Recent studies propose different categorization schemes for rats’ 50 kHz vocalizations. This study attempted to differentiate these categories based on spectrographic features extracted manually and using convolutional neural networks (CNN). In order to analyze the separability of the categories, we trained different classifiers on the extracted features and the best performance was achieved by a support vector machine (SVM) algorithm using the features derived from a CNN which yielded an accuracy of 63.67%. The results showed that many call categories have a high degree of overlap, suggesting that rats may also have a difficult time discriminating them. Next, we created a dendrogram using D-prime scores (a separability measure) generated from our SVM classifier. This dendrogram suggested a new grouping of calls into 7 different categories that are highly dissimilar. Finally, we trained another SVM model on the 7 new categories and achieved 77.8% accuracy.