I have been working in the Dzirasa lab since the summer of 2018, using signal processing and machine learning on various forms of neural information to generate models of how the brain can be used to predict behavior.
Some of my past projects from previous labs include creating an image segmentation algorithm to more efficiently extract neural excitation data from calcium imaging videos, and analyzing behavioral data to explore mouse models of autism. You can see more details by clicking the stars at the top right of this page.
Advisers: Dr. Kafui Dzirasa and Dr. David Carlson
Summer 2018 – Spring 2019, Fall 2019 – present
Weekly hours: 10
Total hours: 800
Supervisor: Jake Benton
Email: jake.bentonject @duke.edu
Description: Our lab works to generate network-level models of psychiatric disorders such as depression, anxiety, and bipolar disorder that are consistent across mice. These models allow us to better understand the ways in which the brain works differently in individuals with a given disorder, and provide the opportunity to use techniques like optogenetics to interfere with these networks to potentially ameliorate the symptoms of the disorder. I’ve had several projects thus far, but all include using machine learning techniques to predict behavior from data collected from regions across the brain. Two of my primary projects are described below.
Using Neuron Firing Rates to Predict Social Behavior: This entails extracting information about high-frequency (>1000Hz) waveforms and sorting those waveforms to find individual neurons. From there, the individual firing rates are used as features in a classification problem to match information about brain activity to the mouse’s behavior. Besides further solidifying a link between the brain and behavior, the goal is to use this model to determine which areas of the brain contribute most highly to accurate predictions, and whether these regions differ in mouse models of psychiatric disorders such as autism or anxiety. Furthermore, success in this task would enable us to create a model of spike trains that can be generalized across mice, which has the potential to dramatically improve methods of understanding the brain.
Classifying Sleep States from EMG and Dorsal Hippocampus Data: This involves processing signals from the dorsal hippocampus of a brain as well as the trapezius and creating a three-dimensional state space that looks at information about the signal at each second. From there, second-long windows can be classified using a variety of techniques into sleep, awake, or REM behaviors. Once this has been done to establish ground truth, local fields from other regions of the brain can be used to train classifiers and ideally create a model of sleep that relies only on the brain and not on the trapezius muscle.
Over time, this project has shifted to require an entirely new algorithm. I’ve done some feature engineering to find the most useful ways to separate the data, and built a hand-initialized clustering method to apply mostly unsupervised clustering to the data to find brain states. I’m currently in the validation stage.
Connection to Reverse-Engineering the Brain: Once these models are generated, we can determine which parts of the brain are most crucial to their creation. This is a tangible link in reverse-engineering the connections between different regions of the brain that lead to different behavioral outcomes.