H.J. Terry Suh,
RLG Short Talk, Spring 2024
PhD Process as an Optimization Problem
(Graduate Descent)
The hard question in life
Why do I exist?
What should I live for?
What should I do?
The hard question in grad school
Why do I exist in grad school?
What should I live for in grad school?
What should I do in grad school?
The hard question in grad school
1. Do world-class research
2. Grow as a researcher
3. Prepare for a job
Why do I exist in grad school?
What should I live for in grad school?
What should I do in grad school?
but there are inevitable tradeoffs....
Grad school as an optimization problem
Maximize what?
The cost function is not clear at all.
Job Prospects
Maximize chances of landing a job
The sparse-reward problem
1st year
2nd year
3rd year
4th year
5th year
I don't know what I should do
Didn't get a good job
I'm not doing anything
I didn't publish anything
I'm not motivated
The sparse-reward problem
Maximize chances of landing a job
1. Too Sparse of a reward to optimize for.
2. The purpose of grad school is not just getting a good job.
Intellectual Curiosity
Maximize time spent on doing what I find interesting
The sparse-reward problem
1st year
2nd year
3rd year
4th year
5th year
Gonna follow the next immediate interesting thing
Come 5th year
I was....
- an honest academic
- followed fundamental curiosities
- but don't have much to show for it.
But I believe people will recognize my intellectual curiosity.
Comparing Researchers
How could I possibly compare different researchers just based on what they find interesting?
No control over interests
You have no control over what you're interested in
It's almost unfair how much your interests dictate your status in the field.
No control over interests
And you can't afford to have no control over your PhD
(Beyond some factors)
Interests as Constraints
Maximize ??
subject to You have to find it interesting.
1. Know your interests. Usually hiding in plain sight
2. Make sure they are not too narrow.
3. But if you did work that you didn't find interesting, you likely failed your PhD (it's a feasibility problem)
Maximize Fame
subject to You have to find it interesting.
Fame
Maximize Fame
subject to You have to find it interesting.
Fame
1. Become a K-Pop Star
2. Open a Youtube Channel
Maximize Metrics
subject to You have to find it interesting.
Metrics
1. Number of Papers
2. Citations
3. H-Index
4. Awards
Quadratic Gaming (Most Basic)
1st year
2nd year
3rd year
4th year
5th year
While True:
- Publish a paper
- Cite every single paper you've published
Citations: 1 + 2 + 3 + 4 + 5 + ... = n*(n+1)/2
H-Index : n
Number of papers: n
Diffusion
RL
Comformal
Prediction
Safety
Tactile
Tactile
Factor
Graphs
Combinatorial Gaming (More Advanced)
Diffusion
RL
Comformal
Prediction
Safety
Tactile
Tactile
Factor
Graphs
Mix & Match
Combinatorial Gaming (More Advanced)
Exponential Gaming (Most Advanced)
RL
What if I tweaked this a bit?
What if I tweaked the previous tweak a bit?
Crisis in Value Assignment
"Using citation metrics as part of academic recruitment decisions leads to an increase in self-citations", LSE
Crisis in Value Assignment
How many citations does this paper generate in total?
Publication has cost
# Citations
# Authors
Do More Metrics actually pay off?
" You can't estimate your odds of getting a faculty job from common quantitative metrics ", Jeremy Fox
So how do we treat metrics?
Maximize ??
subject to You have to find it interesting.
# of papers > n
# of citations > n
1. Too low of a metric is a red flag.
2. Too high of a metric is sometimes a red flag.
Depth
Maximize Depth
subject to You have to find it interesting.
# of papers > n
# of citations > n
Think Deep!
Depth
But What Exactly is Depth??
Perceived Depth
ReLU networks work well because we know piecewise linear feedback is good, and deep learning lets us find these pieces automatically.
That's not deep, we need to understand more.
Using neural tangent kernels we prove that for feedback linearizable nonlinear systems we achieve exponential stability with rate of O(7/3log(log(n))^2)
Perceived Depth
Perceived Depth is not Real Depth.
Don't be fooled by what "looks" deep.
Perceived Depth vs. Impact Tradeoff
Perceived depth
Impact
Staying here is a tricky game
What exactly is depth?
1. Robust to Questions
If you did deep research, you are....
Q: What about this case and that case?
A: I've thought about that, and my answer is....
What exactly is depth?
1. Robust to Questions
2. Parsimonious in your explanations
If you did deep research, you are....
If you can't explain it to a six year old, you don't understand it yourself - Richard Feynman
What exactly is depth?
1. Robust to Questions
2. Parsimonious in your explanations
3. Recognize universal connections to your research
If you did deep research, you are....
Depth
Maximize ??
subject to You have to find it interesting.
# of papers > n
# of citations > n
You have to be sufficiently deep.
Cost or constraint?
In my opinion, there's no real middle ground - constraint.
Impact
Maximize Impact
subject to You have to find it interesting.
# of papers > n
# of citations > n
You have to be sufficiently deep.
Impact
But What Exactly is Impact??
What makes some research impactful while others are forgotten?
Impact: The "Engineering" Factor
1. Solve an important problem
From Bill Freeman's slides on how to do research
"Important" means...
1. Important
2. There are enough people who think it's important
Impact: The "Engineering" Factor
Some Hypotheses
1. "Solved an unsolved problem" >>>>
"Solved a solved problem drastically better" >>>
"Solved a solved problem slightly bettter".
2. "Solved a problem" >>>>
"Solved a problem with some method"
Impact: The "Surprise" Factor
2. Disprove a commonly held belief
Impact: The "Novelty" Factor
3. Provide a New Way of Thinking about the Problem
But show why this perspective is useful / encompassing other views
Impact with Communication
Quality of the Paper
Ability / Effort
to Communicate
From personal talk with Jeff Lipton
IMPACT
Impact with Communication
If you had a chance to sell your work, why don't you do it? - Yunzhu
1. Go give talks, conferences, find connections and talk to people
2. You can only have impact if people know your work
3. Social Media?
Impact with Communication
1. Spend time on writing really well?
2. Toy examples / illustrative figures
3. Explain to 6 year old
Why Impact?
Maximize Impact
subject to You have to find it interesting.
# of papers > n
# of citations > n
You have to be sufficiently deep.
Back to Life
Will to Power
Ubermensch
Eternal Recurrence
All beings exist to exert their own interpretation and value in this world.
If your life repeats in a cycle forever, would you be content?
Someone who gives meaning to his own life without holding himself to other worldly values, proud in eternal recurrence.
An Uber-Scholar
Will to Power
Uber-Scholar
Eternal Recurrence
Academics exist to exert their own academic ideals, visions, and interpretations in topics of their interest.
If your academic career repeats forever, would you be content?
Someone who ultimately believes in their own vision of research without holding themselves to the values of other
scholars.
The Modern Scholar
What we think they are
How they actually are
It is now a scholar's "duty" to care about how their idea makes impact in this world
A Very Important Question
It is now a scholar's "duty" to care about how their idea makes impact in this world
But how do you know "what to make impact with?"
How do you develop an academic vision?
Conclusion
Maximize Impact
subject to You have to find it interesting.
# of papers > n
# of citations > n
You have to be sufficiently deep.
PhD Process as Optimization
By Terry Suh
PhD Process as Optimization
- 202