Computational Methods

Machine learning for animal behavior

Deploying collars on entire social groups gives us a massive amount of data. Before we can gain insights into animal behavior, we first have to identify events of interest, such as vocalizations and behaviors.

Traditional manual approaches to labeling data are time-consuming and expensive, and rapidly become infeasible for large datasets.

Machine learning has the potential to allow us to process large amounts of data automatically - but adapting machine learning approaches to perform reliably on real biological datasets is a nontrivial challenge. Taking advantage of both conventional approaches and more recent advances in deep learning, we are developing methods to make sense of our large, muilti-modal datasets.

Photo by Christian Ziegler

Self-supervised learning for bioacoustics

Bioacoustic datasets present some real challenges.

First, they are typically noisy. For instance, the meerkats we study are continuously digging in the ground for prey, resulting in a vast amount of broadband noise masking their vocalizations.

Second, they are sparse. The audio files recorded on our meerkat collars contain less than 1% vocalizations - the rest are other sounds which must be filtered out.

Third, they are (relatively) small. While our bioacoustic datasets are large by behavioral ecology standards, they are very small compared to datasets in other domains of machine learning such as image and speech recognition.

To address these challenges, we developed animal2vec, a large transformer model and self-supervised training scheme. In a pre-training phase, animal2vec first learns a good representation of bioacoustic data (called an embedding) from large amounts of unlabeled data. It can then be fine-tuned to specific bioacoustic tasks, such as identifying the onsets and offsets of calls and classifying them into types.

Along with animal2vec, we released MeerKAT (Meerkat Kalahari Audio Transcripts), the largest labelled bioacoustic dataset on a terrestrial mammal to date. We hope that this dataset, and others we plan to release soon, will stimulate future research in machine learning tailored to the domain of bioacoustics.

Schäfer-Zimmermann et al. 2025

Using acceleration to detect behaviors

Illustration by Pranav Minasandra

We'd also like to know what our study animals are doing. To identify behaviors from collar data, we use accelerometers - small sensors that records the acceleration over time along three perpendicular axes.

By training a machine learning model such as a random forest classifier, we can use accelerometer data to reliably detect each animal's basic behavioral state - such as resting, walking, or running - at each moment in time.

Sequences of behavioral states give us insight into animals' daily activity patterns, social interactions, and decision-making rules.

By analyzing accelerometer data from collars deployed on multiple species we also recently revealed common patterns in the statistical structure of their behavioral sequences, hinting at common underlying behavioral algorithms.

Page updated

Google Sites

Report abuse