🎙️ Podcast Link 🎙️

When is having a good awareness of the “big picture” in a PhD or general research not helpful?

When you need to focus in on the “little picture”, to preserve your sanity, lower your sky high stress levels, and make some concrete, tangible progress forwards.

When things get tough in research, researchers, including PhD students, can have a (natural) tendency to try to focus on trying to achieve too many things at once, with the natural result of achieving nothing at all, and getting ever more frustrated as a result.

For research fields like robotics, computer vision and machine learning: yes the eventual aim of many PhDs is to create a new method or set of techniques that improves the state-of-the-art in some meaningful way, including demonstration across many datasets / domains / robots and in the case of “helper” pre-processing / post-processing techniques, potentially across many different techniques.

But when you’re struggling, trying to tick all those boxes simultaneously: that novelty box, that generality box – both for domains and for techniques – is completely overwhelming, and usually counterproductive to both your sanity and your progress.

What I often do with students in these situations is reset: dial back the ambition, come up with a simple proof-of-concept of the proposed ideas, often using old statistical or ancient machine learning techniques, and demonstrate that it makes a meaningful difference on .

That won’t in of itself be the research they need, the contributions they need, or enable them to publish (probably), but it will be a massive confidence booster and provide them with a rock solid, well understood foundation to build on, rather than floundering in a sea of too many conflicting goals.

#PhD #research #reset #careeradvice #academia #university

Full Video Notes

One of the ironies of doing a PhD is that those who are best prepared for and understand what is needed to do a PhD are often the ones who are the most stressed out when things inevitably go off the rails at some stage during their PhD.

To explain this, let me give an example of a PhD working in an area like computer vision or somewhere similar in computer science. Typically they will be trying to create a new technique or method that is able to outperform the current methods, has some new knowledge contribution or new technique contribution, and is something that they will hopefully eventually publish and that will form a key part of their PhD.

Now of course things do not always go as planned. What can often happen is that the research can go off track, the code cannot work, and the system that they are programming or coding may become so unwieldy that they do not really understand what is happening or why things are going wrong. Things can feel like they get out of control very very quickly.

In those situations, one of the things that you do not want to be doing as a supervisor and as a student is trying to do all of the things that your PhD will eventually require all at once, right then and there. The pressure will be immense, and actually doing that is grossly unrealistic. Even when things are going well it is still very very challenging.

One of the things that we often do in situations like this is work with the student to step things back a little bit and try to get them focusing on just one core thing. What they will inevitably say, especially if they are a student who understands the PhD process and has done their preparation before starting the PhD, is something like, “Oh, but what you are proposing to do, I probably will not be able to publish that. It does not seem particularly novel. I am worried about whether it will meet the bar for sufficient research contribution and generation of knowledge for my PhD.”

The key aim at this point is to totally take that off the table. You point out that at the moment the student is understandably struggling, and that most students, almost all students, will struggle at some stage, often multiple times during their PhD. You reassure them that this is really an exercise to get them focused and to get some very clear, concrete, tangible wins on the board.

Just because the specific thing that you are going to do right now is not necessarily going to directly produce a publication or some groundbreaking research, that does not mean it is not a very important step toward doing so.

In the computer vision field or robotics field, typically a student may be working on creating a new technique and trying to make that technique work better at some specific task. It could be depth estimation, localization, semantic segmentation, or general scene understanding.

One of the eventual aims of this work is to demonstrate that the method being proposed works generally and across the board, that it is not a fluke and not overfitting to one specific data set. What you eventually want to do is demonstrate that the system works and produces compelling performance benefits across a number of standard benchmark data sets. This is very much the conventional way things are done.

If you are proposing a pre-processing or post-processing method that is supposedly agnostic and can be applied to any system, then you will also need to demonstrate that it helps not just one technique but a whole range of techniques that you could apply it to.

If you are already struggling and overwhelmed, the pressure of having to demonstrate a system across many data sets and show that it benefits many existing techniques is overwhelming. It is way too much to tackle at that point in time.

So what you might do instead is step it back several steps. You might try to create something very simple. It could use 20-year-old statistical methods or basic machine learning methods. It does not have to be flashy or very modern. You demonstrate some key simple concept on one technique and on one data set, and you suppress all those thoughts in the back of your head saying that it will not have enough novelty or that you will not be able to publish it.

That is not the point of this exercise. The point is to do something where you understand completely what is happening, in terms of the data sets, the techniques, and the results. You generate a crisp, compelling result.

The result itself may not be state of the art compared to all the other methods out there, but in terms of the core improvement or benefit you are trying to demonstrate, it will demonstrate that very clearly.

Once you have your confidence up, and your confidence will grow even with this small win, you can then work with your supervisor to build on that foundation. You add sophistication to your technique, expand it, and increase its complexity as needed, but you do so building on a simple foundation that you understand completely.

You also build on it by starting to explore, demonstrate, and evaluate it on multiple data sets, multiple domains, or in robotics on multiple robots. The key is that you are starting from a simple, very well understood foundation where your first, simplistic take has already demonstrated compelling performance advantages.

This approach is generally much easier than trying to solve everything at once. Novelty, demonstration on lots of techniques, demonstration on lots of data sets simultaneously, especially when you are working with a system that is particularly complex or unwieldy, is extremely difficult.

Inevitably it will not work. That is research. That is the nature of research. It can become very overwhelming if you are trying to tick off all the boxes you know you eventually have to tick off for your PhD all at once.

Stepping back, simplifying things, and creating something that is compelling, albeit simplistic and limited, is a great way to break out of this stressful cycle of frustration and feeling like you are getting nowhere, and to get back on track with your PhD.