Low-dimensional learned feature spaces quantify individual and group differences in vocal repertoires.

Area of Science:

• Computational ethology
• Machine learning in behavioral science
• Bioacoustics

Background:

• Analyzing large-scale behavioral data, especially vocalizations, is challenging.
• Traditional methods using handpicked features have limitations, missing variability and complicating analysis.
• Need for advanced analytical techniques to capture the complexity of vocal behavior.

Purpose of the Study:

• To apply variational autoencoder (VAE), an unsupervised learning method, for vocal behavior quantification.
• To compare VAE-derived features against handpicked features in analysis tasks.
• To explore VAE's utility in analyzing vocalizations from laboratory mice and zebra finches.

Main Methods:

• Utilized a variational autoencoder (VAE), a type of unsupervised machine learning model.
• Trained the VAE on vocalization data from laboratory mice and zebra finches.
• Compared the performance of VAE-learned features with traditional handpicked features.

Main Results:

• The VAE model produced a parsimonious representation that outperformed handpicked features.
• Enabled high-resolution measurement of vocal variability in zebra finches (tens of milliseconds).
• Challenged the common belief of distinct clusters in mouse ultrasonic vocalizations.
• Captured tutor-pupil birdsong similarity with superior fidelity compared to prior methods.

Conclusions:

• Unsupervised learning approaches, like VAE, are powerful tools for quantifying complex vocal behavior.
• VAE offers a more comprehensive and accurate method for analyzing high-dimensional bioacoustic data.
• This approach advances the field of computational ethology and bioacoustics analysis.