École Nationale Supérieure de
Techniques Avancées

Résumé

In this report, we describe an approach to behaviour quantication in Drosophila larvas. In the context of neuroscience and nervous system analysis, behaviour quantication is an important building block, preced- ing identication and analysis of the neurons responsible for the execution and coordination of behaviours. Here, we attempted to use unsupervised learning techniques to nd a dierentiable mapping of video recordings of animals onto a latent space in which similar recordings would be close to one antoher, and even more, separable in a set of discrete behaviours. To achieve this goal, we implemented a structured inference model from deep learning literature known as VaDE. In the end, we did not manage to create a well clustered latent space, but we did obtain a continuous latent space showing interesting structure. This demonstrates the di- culty of the task of dening discrete behaviours for the Drosophila larva and shows that a continuous description of behaviour is viable. We de- scribe exhaustively the modications attempted at various levels of the modelling process (data processing, network architecture, loss function) to obtain a clustered latent space, and propose some thoughts on why our particular approach failed. We also describe the software developped dur- ing this process, which was designed with generalizability and reusability in mind.