Representing Behavior with Recurrent Neural Networks Öffentlichkeit

Abstract

Behavior is a complex process that operates across many time and length scales, in many different contexts, and differentially to a variety of external and internal stimuli. The necessity to quantify behavior in a precise and meaningful manner, however, is growing as the advent of new technologies - optogenetics, connectomics, optical imaging techniques - in the world of neuroscience has led to an explosion of assembling and analyzing large swaths of neural data. Here we investigate recurrent neural networks (RNNs) as a model of the underlying dynamics of Drosophila melanogaster and find the behavioral representation it constructs similar to representations built from the previously published results of postural time series. This is a markedly different result from that of RNNs applied to rat models and we investigate the implications.

Table of Contents

Section 1 - Introduction 1 

1.1 Defining the problem 1

1.2 Traditional approaches to quantifying behavior 1

1.3 Stereotyped behaviors and modern approaches 2

1.4 Recurrent Neural Networks 3

1.5 RNNs to represent behavior 6

Section 2 - Data 7 

2.1 Fly recording experiments 8

2.2 Pose estimation with LEAP 8

Section 3 - Methods to fit data 8 

3.1 Data pre-processing 8

3.1.1 Converting to joint angle time series 8

3.1.2 Median-filtering the data 10

3.1.3 Splitting the data into train-validation sets 11

3.1.4 Standardizing the data 11

3.2 Building a dynamical model: feeding the data into an RNN 13

3.2.1 RNN Hyperparameters 13

3.2.2 1-layer RNN hyperparameters 14

3.2.3 2-layer RNN hyperparameters 14

3.2.4 3-layer RNN hyperparameters 14

Section 3 - Methods to fit data

3.3 Building a behavioral representation 14

3.3.1 Extracting hidden states from RNN 14

3.3.2 PCA on hidden states of RNN 15

3.3.3 Wavelet transform of principle components 16

3.3.4 Creating a behavior map with t-SNE 17

3.4 Identifying stereotyped behaviors 17

3.4.1 Peak-finding with Gaussian smoothing 17

3.4.2 Watershed transformation 18

3.4.3 Composite movies 19

Section 4 - Results 21 

4.1 Model errors 21

4.2 Joint angle behavior map 22

4.3 RNN-generated behavior maps 24

4.3.1 1-layer RNN 24

4.3.2 “Glitch” regions 25

4.3.3 Comparing different RNN architectures 26

4.4 Closed-loop networks 29

Section 5 - Discussion 31 

5.1 RNNs on flies versus rats 31

5.2 Future directions 33

Section 6 - Conclusions 33

About this Honors Thesis

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	Emory College
Department	Physics
Degree	B.S.
Submission	Honors Thesis
Language	English
Research Field	Biophysics, General Physics, Theory Computer Science
Stichwort	modeling behavior recurrent neural networks dimensionality reduction Drosophila melanogaster
Committee Chair / Thesis Advisor	Gordon Berman, Emory University
Committee Members	Laura Finzi, Emory University Ilya Nemenman, Emory University

Zuletzt geändert

Primary PDF

Thumbnail	Title	Date Uploaded	Actions
	Representing Behavior with Recurrent Neural Networks ()	2020-04-13 09:53:29 -0400	Press to Select an action Download

Supplemental Files

Thumbnail	Title	Date Uploaded	Actions