My PhD
TL;DRA short-story about why I decided to do a PhD on the subject of explainable multi-agent reinforcement learning. I detail how I weighted this decision and how I created my proposal. I also try to depict what I plan to do for making the best of my PhD.
# Foreword
I have started my PhD this September (2024) and wanted to take time to reflect on my decision process and ambition. Keep in mind that the first part will be biased towards “doing a PhD” even though I try to distill an objective point of view.
Epistemic StatusAs I am only starting a PhD, I am not sharing experience about my PhD but rather about the process I went trough deciding whether to do a PhD. My distillation of PhD knowledge should thus be seen as a condensed of what I read and understood from interactions with other people (colleagues, friends and relatives).
# Why Doing a PhD?
Because You Can…
First doing a PhD is an opportunity that not everyone has the chance to take. It is a natural way to continue breeding knowledge after a Master’s at the university and remain a selective pathway. What’s more, managing to secure a PhD funding also means that you secured an employment for 3 years (in France).
Then, the PhD is a diploma internationally revered, that can open a lot of doors in academia but also in industry. While it’s true that an industry PhD leads to industry and an academia PhD leads to academia, it mostly is a matter of will. Will doing your PhD you’ll mix with other PhDs, both in industry and academia, and you can breed collaborations which later-on can open other doors.
Especially, if you hesitate between a job (linked to R&D) and a PhD, I think it is safe to look for an industry PhD. In most companies you’ll start and end as a regular employee, and when you are already part of the company (even as a PhD) you can evolve in the company more easily.
The Opportunity Cost
If you choose to do a PhD, you might miss other opportunities. You might miss a great job opportunity leading to a high salary or great technical expertise. You might miss creating your startup early. You might also miss daily life perks as doing a PhD can take over some of your free time. In my opinion a great job or a startup creation can wait 3 years, and my later point is more about your expectations and work-life balance.
ReverseThere is also an opportunity cost for not doing a PhD. This internationally revered badge is often the entry ticket to high-value position in big R&D teams. And while it’s technically possible to do it later few will take the plunge and accept this high-demanding and stressful experience later-on.
While I expect not so much PhDs to have already created a startup when beginning their journey (unlike me), I think it is common build one after a PhD. I also think that your PhD knowledge is a crucial Edge (c.f. Creating a Startup), e.g. you could be using your research subject to create a startup. You have 3 years to think this through!
Hot TakeA PhD is just one more experience, and as such it is beneficial even you fail or give up. Obviously, I expect most academia folks to disagree but I am ready to double down on it.
Which is why I think beginning a PhD not necessarily blocks you for 3 years. It shouldn’t prevent you from doing other enterprise, should you find them more exciting or promising for your sake.
The Intrinsic Motivation
In my opinion the intrinsic motivation overrules the previous conditions:
- Don’t do a PhD if you have no intrinsic motivation, even if you can.
- Having an intrinsic motivation can help waver or forget the opportunity cost.
UncertaintyIf you are unsure about doing a PhD maybe you should try it. And I really mean try, as I think dropping from a PhD is also valuable. You should see the PhD as an experience where you will test your fit for research, which you can only scratch the surface during an internship.
The motivation doesn’t necessarily need to be research on its own but could be diverse. You might care about a particular technology or subject that you want to dig. You might want to push the common knowledge even further. You might want to develop tools, softwares or reflections to help a particular field like health.
# Proposal Creation Process
Now that you’re convinced about doing a PhD, let me tell you about the process to get one.
DIY!
If you can, do it yourself! This is a great opportunity to engage with the subject you want/need to explore and as the saying goes, “Fail fast!”. The sooner you discover a subject, the sooner you’ll be able to pivot should you find it less exciting or interesting.
Every exploration you’ll make or article you’ll read will be beneficial for building your scientific knowledge, even if peripheral to your “actual” subject. And by actual I mean the subject you’ll end up writing about. I personally already made a shift in my subject (from the proposal), focusing on a niche subject. As the more you dig, the more you discover, and inevitably you’ll need to focus on a subset of the initial subject that you might not even know existed before starting your PhD.
With my former university (ENS Lyon) I had the opportunity to apply for funding and chose the subject. Yet, because of various causes (gap year and others (that I might detail in a forthcoming story)), I found my supervisor late w.r.t. the application deadline and was assigned, mostly by myself, to do the proposal. I thought it as a double-edge sword, it was a way to explore a subject I cared about and a way to convince my supervisor support my application.
GotchaObviously, this depends a lot on your PhD funding and hosting team. I am glad I could do it, and as a matter of fact I turned down an industry PhD because I felt I wouldn’t be free to choose my subject (because of the company’s objectives).
For the interested folks, I give some technical details about my proposal (and also give the proposal) in the Roadmap section, spoiler it is called Interpretability for Multi-Agent Systems Safety.
On What Should You Compromise?
While I “brought” an opportunity with my application, I expect it not always to be the case. You should always ask if you can write it, but you might apply to funding already allocated to a particular PhD subject. In this case you might need to compromise and thus you should define your priorities among the following:
- The subjects you want to study.
- The places where you want to do your PhD.
- The people you want to work with.
- Your personal constraints.
My PrioritiesDuring my prospection, and in case I would not have gotten my funding I ranked it as the following. Working on Explainability and RL was a priority of mine as I think it really matters, see My Approach to AI Safety. I wanted to work with people caring about such problems. Even though you should be autonomous during a PhD, it’s good to have some support from your direction, even if it means giving you free reign. Finally, while the places didn’t really matter to me, Europe was easier for my personal life (3 years is a long time).
As the subject was my priority I fixed those even before looking for a supervisor, and thus turned down the industry PhD. Then, I looked for supervisors throughout Europe, but mostly in France. I reached out to dozens and interviewed with a few. In addition to a pre-PhD internship what convinced was that I could push a proposal. Among them was Nicolas Maudet, working on multi-agent systems and part of an XAI research group (EXPLICON), with whom I applied for my funding. Overall, I only applied to less than 5 PhD offers as I was fairly stubborn with the subject.
# So What’s the Plan?
You’ll now delve into the over-optimistic plans that I might try to hide in 3 years.
The Roadmap
During my gap year, and later on in my final year, I ended up passionated about XAI. I think that Machine Learning, and in particular Deep Learning, is really beautiful as an objective-driven science, which is also what makes it very successful. Yet, being able to understand how these algorithms solve the problems we couldn’t is the cherry on the cake.
I think that XAI really matters for AI in order to make it safe and reliable, as after the performances people are now looking for accountability. With new regulations, like the European AI Act, XAI is also a business opportunity which makes it very exciting.
Then, during my internships I explored interpretability applied to Reinforcement Learning. A subject I worked on was to interpret RL models able to play chess better than human. It was very fascinating, as many gamers I was thrilled by to work on AI beating humans at games. I later on discovered Multi-Agent Reinforcement Learning which seems to have a lot of research traction, but I don’t forget you AlphaZero.
Sub-ProjectsAs discussed in the following section I think it is important to have sub-projects. It helps broaden your vision and can prevent your subject from driving you crazy. I am thus planning to keep working on chess models occasionally.
I thus, in collaboration with my supervisor, decided to write my proposal on “Interpretability for Multi-Agent Systems Safety”, available here. At the time, I also included related works on LLMs (both LLM interpretability and LLM-based Multi-Agent Systems), as I was familiar with the field. Yet, I now only see it as a potential extension that I might only play around with.
The Daily Basis
I am convinced that the success of any enterprise lies in daily efforts. Making the objectives and efforts small and reachable waver the quotidian stress but also make the mountains more easy to climb!
In addition to small goals, it could also be interesting to divert your attention and energy to proxy goals. It could be about being involved in other projects or collaborations or about exploring a new research or technical field.
In addition to my main axes of research, I plan to work on the following:
- This blog: it helps me explore new subjects and master knowledge by digging more deeply.
- Teaching: I like teaching and it is a great opportunity to review foundational subjects.
- Projects on chess XAI: I feel I could do more in chess XAI.
- Projects on NLP XAI: Using my previous expertise to explore more.
Otherwise, I still want to take time with my hobbies:
- Working on my startup.
- Practising climbing and bouldering (indoor and outdoor).
- Learning foreign languages (I’ll focus on Japanese this year).
- Doing a podcast? (Really unlikely but who knows…).
# Resources
You can read my proposal about “Interpretability for Multi-Agent Systems Safety” here, to which correspond a public PhD offer (soon available) “Interpretability for Deep Multi-Agent Systems”.