Mapping Behavioral and Neural Data with Variational Autoencoders Öffentlichkeit
Kang, Shinyoung (Spring 2021)
Published
Abstract
Animal behavior exists across many time and length scales, likely requiring commensurate multiscale neural activity to control it. Although there have been improvements in our ability to measure neural circuitry and animal behaviors over long timescales, the biggest challenge in measuring behavior has been finding precise methods of selecting meaningful quantitative representation for an animal’s behavior and fixing errors in the recordings. In this paper, we look at three-dimensional rodent’s kinematic data obtained through 20 sensors attached to a rodent. Previously, many methods such as principal component analysis attempted to reduce a complex animal’s postural data by selecting a select number of points representing the majority of the variance of the actions to create a meaningful representation. Instead of selecting a few points, we utilize variational autoencoders (VAEs), which not only reduce the dimension of the complex data but also can be used to generatively reconstruct the rodent’s posture allowing us to fix tracking errors in the recordings. The VAE fixes tracking errors by training on non-error data, thereby reducing noise in the postural data, a clearer identification of the stereotyped behaviors and generation of behavioral maps of those behaviors emerges. We look at behavioral maps created with different dimensional reduction methods, predicting future behavioral transitions from one behavior to another. Lastly, we analyze simultaneous behavioral and neural recording to see if we can identify certain neurons that fire during particular stereotyped behaviors such as grooming. This allows prediction of the behavior of an animal using neural data and opens the possibilities for quantitatively linking brain and behavior.
Table of Contents
Introduction………………………………………………………………………………………………………………..…………..….1
Results……………………………………………………………………………………………………………………………………….4
CAPTURE………………………………………………………………………………………………………….......………..…4
Variational Autoencoder Correction………………………………………………………………………............…….…4
Figure 1 – Error Rate of Sensors on the Rodent’s Original Postural Data………………….........................…...5
Figure 2 – Number of Sensors with Errors for the Rodent’s Original Postural Data……….........................…5
Figure 3 – A cartoon of how variational autoencoders work……………………………………….....................…6
Figure 4 – Training Loss for Variational Autoencoder…………………………..……………………...................…7
Figure 5 – Labeling of Sensors on the Rodent's Posture……………………………..…..………….................……7
Figure 6 – Error for Changes in Sensors……………………………………………………………………….............…8
Figure 7 – VAE Prediction for Rodent’s Posture Without Errors.…………………………….…...................……9
Figure 8 – Number of Errors Defined by Missing Values versus VAE.…………………………....................……9
Figure 9 – Changes in Error Rate Defined by VAE After the Prediction …………………….......................…..10
Behavioral Map Comparison…………………………………………………………………………...........….………….10
Figure 10 – General Pipeline for Behavioral Analysis………………………………..………….................…….…11
Figure 11 – PCA of the Interpolate filled Rodent Data…………………………………………….................……..13
Figure 12 – PCA of the VAE Predicted Rodent Data………………………………………..…….................……….14
Figure 13 – Behavioral Maps of Each of the Behavioral Methods……………………………......................……16
Figure 14 – Histogram of Velocities in the Embedded Space………………………………….....................…..…16
Figure 15 – Labeling Watershed Regions of the Behavioral Map………………………......................………….17
Transitional Matrix Analysis………………………………………………………………………..........…………………17
Figure 16 – Markov Transitional Matrix…………………………………………………………………..............….…19
Figure 17 – Transition rRates and Flux Plotted on Behavioral Map……………………………......................…21
Figure 18 – Optimal Trade-off Curves for Lags 1 to 100.........................………………………...........……..…21
Figure 19 – Informational Bottleneck Partitioning at Behavioral Space…………………….......................….22
Neural Analysis ……………………………………………………………………………………………….........……..…..22
Figure 20 – Density Plots for Firing at Front Leg Rubbing and Grooming Regions….….........................….23
Figure 21 – Density Plots for Firing at Front Leg Rubbing Regions……………………..…....................…..…24
Figure 22 – Density Plots for Firing at Lowered Position and Grooming Region……….........................…..24
Figure 23 – Density Plots for Firing at Rearing Regions…………………………………….………...................…25
Figure 24 – Density Plots for Firing at Lowered Position Leg Locomotion Regions……...........................…25
Figure 25 – Density Plots for Firing at Posterior Leg Locomotion Regions.………………........................…..26
Figure 26 – Density Plot for Firing at Idle Regions……………………………………………....................……..…27
Discussion………………………………………………………………………………………………………………...…….………..27
Limitations…………………………………………………………………………..……………………….......….…..……..27
Future Challenges……………………………………………………………………..…………………….........…………..28
Methods …………………………………………………………………………………………………..……………...………………30
Data………………………………………………………………………………………………….…………......……………..30
VAE……..……………………………………………………………………………………………………………......………..31
Figure 27 – VAE Layers….………………………………………………………………….…………….…..…..............…31
Figure 28 – Training VAE: Dataset Selection Pipeline ………………………………………………....................…33
PCA……………………………………………………………………………………………………………........…………….34
Morlet Wavelet Transform…………………………………………………………………………..……................…….35
t-SNE……………………………………………………………………………………………………………….......……..….35
Transitional Matrix……………………………………………………………………………………..………...........…….37
Predictive Informational Bottleneck………………………………………………………….……………................…37
References………………………………………………………………………………………………………..………......………….39
About this Honors Thesis
School | |
---|---|
Department | |
Degree | |
Submission | |
Language |
|
Research Field | |
Stichwort | |
Committee Chair / Thesis Advisor | |
Committee Members |
Primary PDF
Thumbnail | Title | Date Uploaded | Actions |
---|---|---|---|
|
Mapping Behavioral and Neural Data with Variational Autoencoders () | 2021-04-13 15:13:19 -0400 |
|
Supplemental Files
Thumbnail | Title | Date Uploaded | Actions |
---|---|---|---|
|
Rearing Position (Rodent in it's rearing position) | 2021-04-13 15:13:25 -0400 |
|
|
Front Leg Locomotion (Rodent performing front leg locomotion) | 2021-04-13 15:13:31 -0400 |
|
|
Idle State (Rodent not moving) | 2021-04-13 15:13:38 -0400 |
|