this fall, i'll be at the Recurse Center in New York City studying deep reinforcement learning + misinformation, democracy, and the integrity of our information ecosystem. i'm curious to see how this time will challenge me to be a better programmer, social scientist, technologist, and writer. i wrote about my motivation and what i'm hoping to learn here.
this is a tentative plan for my batch, with an understanding that it may change depending on momentum, inspiration, and learnings from my fellow Recursers. send me feedback via Twitter!
September 23 - 29
introduction to deep reinforcement learning
via Spinning Up in Deep RL & OpenAI Gym
implement solutions to these exercises
implement the vanilla policy gradient
implement proximal policy optimization
finish porting the lunar-lander gym environment to ocean-gym
ocean-gym is a fun side project where i port existing gym environments to be ocean-themed
in whale-lunar-lander, the agent will land a whale on the moon instead of a lunar lander
this has helped me to understand a bit more about what goes into an RL environment!
September 30 - October 6
applied deep reinforcement learning
last summer i researched smart cities with Sidewalk Labs, so i was very excited to speak with Kathy Jang, a researcher (and Recurser!) who works on traffic control with deep reinforcement learning via the Flow project out of Berkeley's Mobile Sensing Lab. i'm excited about real-world applications of deep reinforcement learning and plan to spend at least a week exploring what can be done in this space.
via the Flow project
work through these introductory tutorials
play with RLLib
build a small simulation
depending on the outcome of the first two weeks, i plan to either extend the exploration on deep RL into week 3 or transition into working on misinformation + NLP on Twitter immediately. either way, i will be spending 3-4 weeks studying misinformation.
weeks 3/4 - 6
October 7 - October 31
misinformation, natural language processing, and Twitter
via the Elections Integrity Hub datasets from Twitter
Russian troll bot vs. non-Russian troll bot
social network analysis + unsupervised learning on these actors to understand if + how they're linked together
test natural language embedding techniques to cluster tweets which are similar in content
scrape Twitter for other similar political data
train a classifier on this data to identify malicious actors and/or misleading tweets
test this out on today's context
test out the classifiers on current Twitter data in the context of the 2020 American election
test these methods on tweets surrounding the October 2019 Canadian federal election
aggregate relevant datasets for further research
read the “Securing American Elections” report from the Stanford Internet Observatory
i'm hopeful that this plan will serve as an ambitious-yet-malleable starting point for what will hopefully be six weeks of productive programming, research, and writing, among talented strangers who will turn out to be collaborators, mentors, and friends.