École Nationale Supérieure de
Techniques Avancées

Résumé

Studies done on Drosophila larvae help to draw connection between observed behaviour and underlying neural mechanisms and thus shed light on decision-making processes in animals. The analysis involves setting up experiments, performing neuronal manipulation, recording the observed behaviour and subsequent classifying of larvae movements into predefined types of actions. The classification step is needed to quantify the behavioral change. During these manipulations, different larvae genotypes are used, some of which have different behavioral dynamics and thus deviations in behaviour are observed. In addition, the team studies internal state changes, which also impacts behaviour dynamics and results in behaviour deviations. In recent studies some genetic lines of larvae showed important deviations from known types of behaviour. This has required to either redefine existing actions or to introduce new action types. Both options would necessitate to re-train the classifier already used in the laboratory, which would also include gathering and labeling new experimental data. This work focuses on finding ways to reproduce behaviour data artificially based on already collected real recordings to avoid labor-intensive setup of real experiments and processing of retrieved data. The proposed method helps to simulate larva behaviour and can be used to render realistic movement sequences with desired constraints put on larva and action features.