Long Time Scales and Hierarchical Structure in Drosophila Behavior Open Access

Overman, Katherine (Summer 2022)

Permanent URL: https://etd.library.emory.edu/concern/etds/bk128c180?locale=en


  Animal behavior is fundamentally a biological process with complex dynamics. However, to fully understand this process, an interdisciplinary approach is essential. In this thesis, we used methods from physics and machine learning to study behavior. We aim to quantify Drosophila melanogaster fruit flies' stereotyped behaviors, to build a model that is capable of reproducing complex, hidden behavioral features, specifically the long time scales and hierarchical structure that we see in our data, and to offer explanations for what internal processes are modulating these features, both in general and in the context of aging. The data we are using are hour long recordings of fruit flies imaged from above in a featureless dish that restrict their ability to fly but otherwise allows them to move freely. Using methods from Berman (2014), we are able to take these recordings, identify the set of stereotyped behaviors and embed them into a 2D plan known as the behavioral map. This indirectly gives us a time series of behavioral states, and we have two different frameworks of this data that we use.


  The first consists of 300 fruit flies, half male and half female, with ages ranging from 0 to 70 days, which is the typical lifespan of these flies. With this data set, we measure behavioral changes as a function of age and find that the changes differ between the males and females. As youths, the female flies are very active, performing mostly locomotive behaviors or behaviors that require lots of movement. As they age, they become more lethargic and perform more idle or ``lazy'' behaviors. On the other hand, male flies are lazy as youths, perform more active behaviors in mid-life when it is optimal mating time, and finally return to laziness in their late life. We find that many of the hidden dynamics that we measure from data, show no change. This includes the entropy of the behavioral map, the stereotypy (or repeatability) of the behaviors, the long time scale dynamics, and the hierarchical structure in the data. However, we also find that a major contributor to the change in behavior is a change in the amount of energy available to the flies, or rather, that the flies have a changing energy budget that modulates their behavior.


  The second data set has 59 young, male fruit flies. We use the time series from this data to train a recurrent neural network so that it is capable of reproducing the data and the data's hidden dynamics. It has been shown by previous work that the internal states of an RNN act like a dynamical system with fixed points, and that the interactions between the fixed points lend some clarity on how neural networks make their computations. By doing so with our network, we find that the fixed points of our network behave as a multi-well plane with basins containing one or more behavioral states. The plane shifts over time which changes the barriers between our basins, and we show that this paradigm is capable of modulating the dynamics in our data.


  In conclusion, this thesis serves to offer explanations for the types of internal mechanisms that are necessary to produce the dynamics that are inherent to behavioral data across long time scales, even as long as a fruit fly's lifetime. Common behavioral models are currently incapable of doing this, thus we also propose neural networks are a better class of model to be considered for future behavioral studies. These models could be crucial to unlocking what is modulating behavior within an organism. 

Table of Contents

Introduction Defining Animal Behavior Measuring Animal Behavior Behavioral Spaces Modeling Animal Behavior Recurrent Neural Networks Time Scales and Hierarchy Behavioral Effects of Aging Internal Recurrent Neural Network Dynamics Basin Hopping Model Arrhenius Behavior Application to Our Network Thesis Outline Measuring the repertoire of age-related behavioral changes in Drosophila melanogaster Introduction Results Experiments and behavioral densities Quantifying behavioral changes with age Estimated Energy Consumption Alters with Age Complexity of the Behavioral Repertoire Long Time Scales and Hierarchical Structure in Behavior with Age Stereotypy Discussion Materials and methods Acknowledgements Supplemental Figures Using Recurrent Neural Networks to Model Hidden Dynamics of Animal Behavior Introduction Results Model Validation Fixed Points Dynamics Basin Dynamics Discussion Methods Conclusion

About this Dissertation

Rights statement
  • Permission granted by the author to include this thesis or dissertation in this repository. All rights reserved by the author. Please contact the author for information regarding the reproduction and use of this thesis or dissertation.
  • English
Research Field
Committee Chair / Thesis Advisor
Committee Members
Last modified

Primary PDF

Supplemental Files