Circuit Dynamics and Computational Neuroscience I.1.l Deep and machine learning Monday AM + Wednesday AM

2476 - Disentangling the Latent Structure of Mouse Behavior by Learning a Spatiotemporal Manifold


Abstract Body

Aim: Modern neuroscience calls for robust behavioral quantification in order to interpret underlying brain activity. Recently, pose-estimation tools enabled efficient tracking of animal body-parts via supervised deep learning. While such tools provide a continuous representation of the animal body motion, the extraction of underlying discrete states which are hypothesized to underlie behavior remains a challenge. Unsupervised behavior quantification has the power to unravel subtle differences that are undetectable by a human experimenter. ​

Methods: Here, we present Variational Animal Motion Embedding (VAME), a probabilistic machine learning framework for spatiotemporal clustering of behavioral signals obtained from pose-estimation tools. We propose that these continuous signals can be grouped into discrete states via clustering of the latent vector obtained from a recurrent neural network autoencoder. Moreover, we chose to establish our machine learning model within the framework of variational autoencoders. This allows the model to learn a complex distribution of the data and to generalize well to previously unseen data.

Results: We investigated behavioral differences between non-transgenic and transgenic mice model model of Beta-amyloidosis. We show that our method robustly identifies differences in the distribution of behavioral modules and is able to predict the phenotype of individual mice based on this distribution. ​

Conclusion: We argue that VAME forms a state-of-the-art approach to robust behavior quantification, allowing for clustering of spatiotemporal signals. Alzheimer’s disease is of major interest and finiding body pose signatures as biomarkers of the disease model would be a major leap towards early detection of the disease.