Encoding the Structure of Animal Motion Dynamics using Variational Auto-Encoding

Abstract

Quantifying and detecting the hierarchical organization of behavior presents a significant challenge in neuroscience. Recent advancements in markerless pose estimation have enabled the spatiotemporal tracking of behavioral dynamics. However, there is a pressing need for robust and reliable technical approaches that can unveil the underlying structure within these data and segment behaviour into hierarchically organized motifs. In this thesis, I propose an unsupervised probabilistic deep learning framework called Variational Embeddings of Animal Motion (VAME) that addresses these challenges. By leveraging VAME, I can identify the behavioural structure from deep variational embeddings of animal motion, providing a powerful tool for behavioural analysis. To demonstrate the framework’s effectiveness, I utilize a mouse model of beta amyloidosis as a use case. The results demonstrate that VAME not only identifies discrete behavioral motifs, but it also captures a hierarchical representation of how these motifs are utilized. This hierarchical representation allows for the grouping of motifs into communities, revealing intricate behavioral patterns that were previously overlooked by human visual observation. Remarkably, VAME detects differences in community-specific motif usage among individual mouse cohorts that were previously undetectable without the framework’s aid. Importantly, the proposed approach offers robust segmentation of animal motion, making it applicable to a wide range of experimental setups, models, and conditions. It eliminates the need for supervised or a-priori human interference, which greatly enhances its versatility and efficiency. VAME’s unsupervised nature also alleviates the burden of manual annotation and subjective biases associated with traditional methods of behavioral analysis. In summary, this work presents a significant advancement in the field of behavioral neuroscience by providing a powerful and unsupervised framework for uncovering the hierarchical organization of behavior. VAME’s ability to identify discrete motifs, capture their hierarchical representation, and detect differences in motif usage among communities paves the way for deeper insights into the complex dynamics of animal behavior.