Deep Learning Methods and Methods in Drosophila melangaster Classification and Pose Estimation

Abstract

Systems neuroscience largely has two main goals: to understand how the brain represents information and how these representations drive behavior. With many of the recent methodological developments for controlling and measuring the activity of nervous systems in organisms such as the fruit fly, Drosophila melanogaster, which is useful for its powerful genetic toolbox, scientists are able to control and measure activity of nervous systems. However, there is still a lack of emphasis placed on quantitative tools for behavior, the primary output of the brain. One of the main goals of computational ethology is to model and quantify animal behavior through the use of machine learning and computer vision,- recent literature has attempted to computationally map, quantify, and model behavior. For example, there have been both supervised and unsupervised machine learning techniques to quantify behavior, which primarily focus on time series segmentation of animal behavior models. Instead, we approach behavior by extracting more useful features to represent animal pose, as has been done in mice, fish, and humans. In this thesis, we leverage deep learning to improve these algorithms for pose estimation in Drosophila melanogaster. We follow three separate steps to do so: first, we split videos of behavioral output into separate images of either fly or no fly, train and label the orientation of flies, and build various deep learning architectures to detect, classify, and estimate pose of Drosophila melanogaster