Disentangling the Latent Structure of Mouse Behavior by Learning a Spatiotemporal Manifold
1 DZNE Bonn
Naturalistic behavior is highly complex and dynamic.
Modern neuroscience calls for robust behavioral quantification and
analysis methods in order to interpret underlying brain activity.
Here, we present a spatiotemporal machine learning framework to
disentangle the latent structure of behavioral dynamics recorded during head-fixed
and unrestrained experimental conditions. Our method builds on top of state-of-the-art markerless body part tracking algorithms. It uses a variational recurrent neural network to identify temporal dependencies in the multivariate marker time series. These dependencies are capture in the latent space of the model from which we can cluster the animals behavior into similar motifs. Based on this clustering we find differences in behavior in distinct mice phenotype which have not been naively observable otherwise. To underline the versatility and robustness of our approach we demonstrate how the discovered motifs correlate to population neuronal activity recorded in the mouse CA1 region.