Pitfalls of Using Feature-Based Classification for Mouse Ultrasonic Vocalizations
* Presenting author
Abstract:
House mice communicate through ultrasonic vocalizations (USVs), which are above the range of human hearing (>20 kHz), and have been classified into >10 syllable types. Feature-based classifiers are mainly used to study USVs, and yet these have been shown to be outperformed by image-based models in studies of other types of acoustic signals. In this study, we focused on the pitfalls of feature-based classification of USVs comparing them with a state-of-the-art image-based classification method (CNNs). Results show that the limitation of feature-based methods can be attributed to two main reasons: 1) features are defined by the frequency track (FT) of USVs, and as a result, any error in determining the FT has a profound impact on the extracted features and thus the performance of the classifier. 2) The features are based on human-determined thresholds which may not be met in case of any trivial variations in calculated FTs. We will show the exact results of the numerical experiments.