Multi-timescale representation of animal behavior Público

Abstract

This dissertation focuses on developing computational methods to identify and characterize long-timescale dynamics in behavioral data across multiple spatiotemporal scales. Utilizing the dynamical nature of Recurrent Neural Networks (RNNs), we formulate novel approaches to learn dynamical models from large timeseries datasets of animal postures. We introduce LIDAR, a framework for training RNNs on extensive timeseries data while maintaining temporal coherence through statefulness. Applying this to human gait data, we generate quantitative gait signatures that encode individual and group-specific locomotor patterns, with potential clinical applications in diagnosis and therapy development. We extend our methodology to create multi-timescale representations of behavior using hierarchical RNN models. We apply this approach to a dataset containing simulataneous neural and behavioral recordings, allowing us to explore neural correlates underlying behavior at multiple timescales.

To address limitations in existing methods for reconstructing state-space dynamics, we propose a novel encoder-decoder RNN architecture capable of identifying long-timescale non-stationarities in time series data generated from a modified Lorenz system. Our work demonstrates the utility of RNNs in approximating dynamical systems from vast behavioral datasets, intentionally leveraging these overparameterized models to learn underlying dynamical timescales. These approaches offer promising avenues for disentangling the hierarchical organization of behavioral patterns, characterizing long-timescale physiological states, and understanding neural dynamics underlying complex behaviors. 

Table of Contents

1 Introduction 1

1.1 A brief history of ethology 3

1.2 Tracking Posture 5

1.3 Postural Dynamics 10

1.4 Takens’ Embedding Theorem and ensemble methods for state-space reconstruction 15

1.5 Thesis Outline 20

2 Using Recurrent Neural Networks to mimic dynamical systems 22

2.1 Introduction 22

2.2 Recurrent Neural Networks 25

2.3 LIDAR: a python package to train RNNs statefully 29

2.4 Characterizing human gait dynamics using RNNs 30

2.4.1 Introduction 30

2.4.2 Dataset 34

2.4.3 Model training and selection 36

2.4.4 Measuring gait signatures from RNNs 38

2.4.5 Gait signatures are individual and group-specific 40

2.4.6 Biomechanical interpretability of gait signatures 42

2.4.7 Using RNNs to generalize to new gait speeds 43

2.5 Building multi-timescale representation of animal behavior using RNNs 45

2.5.1 Datasets 46

2.5.2 Building representations at a single timescale 49

2.5.3 Building representations at coarsened timescales 52

2.5.4 Linking behavioral and neural representations across timescales 54

2.6 Conclusion 56

3 Unsupervised learning algorithms reveal sex biases in baseline and stress adaptive behavior 62

3.1 Introduction 62

3.2 Methods 64

3.2.1 Animals 64

3.2.2 Stress 65

3.2.3 Surgeries 65

3.2.4 Behavioral Assessments 65

3.2.5 Chemogenetic Manipulation 66

3.2.6 Posture Tracking 66

3.2.7 Quantitative identification of behaviors 67

3.2.8 Training classifiers across assays 68

3.2.9 Calculating statistical upregulation between two assays 68

3.3 Results 68

3.3.1 Sex biases in the exploration of an open field 68

3.3.2 Stress reverses baseline sex differences 69

3.3.3 Chemogenetic activation of susceptibility circuit drives similar upregulations in both sexes 73

3.4 Discussion 75

4 State-space reconstruction of dynamical systems with long timescalenon-stationarities 82

4.1 Introduction 82

4.2 The Lorenz System 83

4.3 Modified Lorenz System 87

4.4 Reconstructing state-space dynamics using RNNs 92

4.5 Conclusion 93

5 Discussion 98

Bibliography 102

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.

School	Laney Graduate School
Department	Physics
Degree	Ph.D.
Submission	Dissertation
Language	English
Research Field	Physics, Theory Physics, General Biophysics, General
Palavra-chave	Physics Dynamical Systems Animal Behavior Long-timescale dynamics Physics of behavior Behavioral Analysis Recurrent Neural Networks
Committee Chair / Thesis Advisor	Berman, Gordon, Emory University
Committee Members	Rieser, Jennifer, Emory University Nemenman, Ilya, Emory University Weeks, Eric, Emory University Liu, Robert, Emory University

Última modificação

Primary PDF

Thumbnail	Title	Date Uploaded	Actions
	Multi-timescale representation of animal behavior ()	2024-10-31 09:14:15 -0400	Press to Select an action Download

Supplemental Files

Thumbnail	Title	Date Uploaded	Actions