Reflections on my Ph.D. journey

I defended my Ph.D. thesis the 27th of January 2023 (a while back!), finally becoming a so-called doctor. It has been a weird Ph.D., which has been basically remote (thanks covid!) except for the last year and a few months of my first year. There are a lot of things that I have missed in comparison to a conventional Ph.D. But still, there are a few things that I think I have learnt during this period, and that maybe can be handy for aspiring Ph.D. students.

Own your work

I had to start a new line of research in my Ph.D. lab, something they had not worked on before: digital pathology and multi-modal data fusion. My lab had previous experience working with gene expression data (specifically RNA-Seq) and had developed a few methods for feature selection and created an R library for running a whole machine learning RNA-Seq pipeline It is pretty outdated, but it is available here.

That meant that my supervisors had zero expertise in this area; thus, I had to learn everything from scratch by myself! However, I think this was a great (sometimes irritating) advantage. That made me realize one of the important aspects of doing a Ph.D.: you need to own your own work. Your supervisors need to be your guiding light, they need to get you to focus on important (and publishable) problems. But you need to be the one that has control over your problem. You are not going to become an expert overnight, but slowly, you need to become better than your supervisors in that specific area you are working on.

Read (all) papers (you can) well

In the field of medical AI, keeping up with the rapid pace of new advancements is nearly impossible. This challenge extends to all AI research areas. Don’t be discouraged if you feel the world is moving forward while you’re standing still. You’re not. When I started in digital pathology, there were far fewer publications. Today, we’re witnessing an explosion of research and foundation models in AI digital pathology.

Reading papers is essential during your Ph.D., but how you read them is more crucial than how many. Scientific manuscripts follow specific, often unspoken rules, tones, and structures. Reading papers with an eye towards how ideas are expressed, framed, and positioned will improve your own scientific writing. Modern science involves an element of marketing; effectively selling your idea enhances its appeal and, consequently, its chances of publication.

Have a good organization

I’ve never been the most organized person. In research, you run numerous experiments, often with variations in hyperparameters, data splits, or random seeds. Having a solid system for organization and experiment control is vital for clarity. Without it, you’ll waste days deciphering your own work.

Early in my Ph.D., I used Tensorboard. At Stanford, I discovered Weights & Biases (wandb), and it was a game-changer. It makes tracking experiments, labeling them, and staying organized incredibly easy. I highly recommend it!

Another valuable tip I’ve discovered is to organize all meaningful results into a presentation format, as if you were preparing for a conference talk. The process of sorting and presenting your findings forces you to critically evaluate your work, often revealing errors or gaps that might otherwise go unnoticed.