Abstract
<p>Relying on the rapid development of computer technology, machine learning is increasingly being applied to studies of animal behaviors. Video tracking enables observation of behaviors in an automatic, reliable and consistent way for an extended period. In Drosophila melanogaster, the fruit fly, various machine learning techniques have proven to be powerful tools in studies of locomotion tracking and overall behavior classification. In our study, we applied machine learning to investigate two important behaviors of flies: grooming and sleep. Despite being universally existing in almost all animals, the regulation of grooming is poorly understood. In the first application, based on a k-nearest-neighbors algorithm, we developed a high-throughput platform that allows automated long-term detection of grooming in flies. Our data show that flies spend ~13% of their waking time grooming, driven largely by two major internal programs: one controls the timing and the other controls the duration of the grooming behavior. Our quantitative approach presents the opportunity for further dissection of mechanisms controlling long-term grooming in flies. In another of our studies, we applied a convolution neural network model to facilitate our study of sleep in flies. Drosophila has been widely applied as an ideal model organism in the study of sleep behavior. In most studies, a prolonged period (>5 minutes) of behavioral immobility is qualified as a period of sleep. However, such a criterion is potentially problematic. According to the widely accepted criteria to define sleep in all animals, we tried to improve and complete current definition of sleep in flies by seeking for fly-specific postures of sleep. In our work, based on a convolutional neural network, a model that has been successfully applied in various tasks in the field of machine vision, we built a system that potentially enables to distinguish postures accompanied with different stages of rest and sleep.</p>