Finding Your Thesis Statement & Story
After a few years in a PhD program, working on whatever research project your advisor had funding for, or whatever project seemed interesting or fun at the time, it can be daunting to find a unified story or thesis statement that accurately, concisely, and interestingly summarizes your work. In this post, I will describe how I was able to find a cohesive story that unified the disparate strands of work I’d completed, and helped me focus my thesis proposal into an interesting final project.
Background— My Work So Far
My thesis work began in reinforcement learning, using differentiable decision trees to try to make RL agents more sample-efficient by leveraging human knowledge for warm-starts. Then I developed a method that discretized those differentiable decision trees into discrete trees, making interpretable models that humans could inspect. After that, I took a hard-pivot into language-guided reinforcement learning, and influence-functions as a tool to explain properties of word-embedding models (long story for how those projects came to be). Then, on an internship, I did some really cool and interesting work on personalized federated learning (which I was very proud of and wanted to include), and finally I had recently completed a large-scale user study on human perceptions of explainability. So this was a lot of topics and papers across an absolute smorgasbord of subjects.
Step 1: Group Your Projects
The first and easiest piece of advice I have to offer is to group your projects into meaningful areas. Ignore the in-depth mechanics of your projects and just think about the high level problems, insights, or contributions. Are there natural clusters that emerge? Can you find common themes across some of the papers? Again, you don’t need to find a silver bullet for all of your work, just try to reduce the number of subjects that you need to unify.
Step 2: Summarize Your Work
For me, this meant talking to as many people as possible about my work. I would give people a summary of everything I had done, and eventually I got sick of saying the same things over and over. I found ways to take shortcuts or group projects, and eventually found a useful few-sentence summary of my work that conveyed what I most cared about in my completed projects. For me, this was:
“I’ve done a bit of human-in-the-loop RL, trying to integrate human expertise into RL agents to improve their generalization or their sample efficiency, I’ve done some personalization in my time at Apple, and I’ve worked on human-centric interpretability/explainability, both on contributing new methods and on understanding how humans use those methods. For my proposal, I’m planning to bring those all together by learning to personalize explanations of a task-centric agent alongside a human in a user-study.”
Step 3: Re-Examine Your Work
After you have this summary that ties your work into a clean narrative, you may need to revisit your completed projects and re-frame them in this new light. Originally, my differentiable decision tree project was more about sample-efficiency than about the human in the loop. While I kept almost all of my original results, diagrams, experiments, and slides, my script changed to focus more on how this was an interactive learning approach that leveraged human expertise and put the human front-and-center. While this wasn’t exactly the original pitch, it allowed me to present a cleaner and more cohesive overview of my work that wasn’t as jarring for my audience.
As part of this step, also consider revisiting your proposed project! If you find a clean story that unifies your work, it may also reveal a natural capstone project to bring everything together. Follow that thread and consider how you might complete the narrative with your final project. If you see a natural end-product, your committee likely will do, and they’ll be expecting you to go there.
Step 4: Find a Thesis Statement
This is maybe the hardest part, but, as I’ve mentioned previously, it doesn’t need to be a single sentence. So don’t stress too hard. My thesis statement was three parts, one for each body/cluster of work that I had completed. My thesis statement slide read:
Interactive and explainable machine learning yields improved experiences for human users of intelligent systems.
1. Machine learning with human expertise improves task performance as measured by success rates and reward.
2. Personalized machine learning improves task performance for large heterogeneous populations of users.
3. Machine learning with explainability improves human perceptions of intelligent agents and enhances user compliance with agent suggestions.
One overarching “vision” statement, and 3 testable statements that I would prove in my talk.
Conclusion
Finding a story for your work is hard and does not come naturally or easily to many. Give yourself plenty of time, talk to people both in and out of your field, and try to view your work from a new angle. It’s almost always possible to find a satisfying narrative, it just takes work, creativity, and re-framing to see your projects in the right light. Good luck!