My PhD Journey in AI + ML

while balancing YouTube on the side

This is a special AI Coffee Break: Recently, I’ve defended my PhD dissertation successfully and after sharing pictures of this academic milestone, many of you have asked about my PhD journey, the work I did, and the challenges I faced.

So, I thought it would be fun to share my experiences with you. I'll take you through my entire PhD journey—from why I decided to pursue a PhD, to the extra duties I had to fulfil, like teaching. Spoiler alert: there was no unhealthy amount of coffee involved. Well, except for Ms. Coffee Bean.

I’ll also explain to you my PhD hat 🎓, give you a brief rundown of the research that went into my thesis, and explain why I decided to start this YouTube channel during my PhD.

« Enjoy this post in video format 👇! »

I hope that by sharing my story, I can provide some insights and inspiration for anyone considering or currently pursuing a PhD. But remember, this is just my experience: a single data point in the vast universe of PhD adventures out there.

So, grab a cup of coffee ☕ and let’s start! First, let me tell you a little bit about why I decided to pursue a PhD and how I got started on this journey.

Why do a PhD?

My decision to start a PhD hit me while I was wrapping up my master's degree in Physics. Funny enough, my master’s thesis wasn’t your typical physics project. I was actually doing machine learning in image processing, and I found myself utterly fascinated by the possibilities of what ML and AI can do. I’ve always had academic ambitions, so diving deeper into research felt like the natural next step. But I cannot give you a more noble motive in the grand scheme of things, because let’s be real—life’s big decisions are often a mix of timing, place, and the people who are next to you and inspire you along the way.

Now, about my research topic: My interest in image-related AI and ML sparked a curiosity to understand more about humans: What exactly makes people snap a photo? And what words do they use to describe those images? Here’s where it gets interesting: to truly get to the heart of this, I realized I needed to dive into language. After all, language and vision are intertwined—they both describe our actions and observations of the world. And this was before large language models (LLMs) were all the rage.

The next part of the puzzle was to choose the location. My partner had already started a PhD in Heidelberg, so naturally, Heidelberg became the only option for me. After some searching, I found a professor in the Computational Linguistics department who was brave—or perhaps crazy—enough to embark on a research journey with me to combine image and text understanding – which is called multimodal understanding.

This professor specialized in natural language processing, while I brought my expertise in image processing to the table. Together, we set out to explore deep learning models that can use both vision and language. It was a perfect blend of our skills and interests, and thus began my PhD adventure.

PhD Timeline

I started my PhD in 2019, and it took me just a bit more than 5 years to complete. Before I could even really start, I had 6 months to prove that I could work in natural language processing, or short: NLP; and I was encouraged to take one master’s level course to see how it all works there. The reason was that I had studied physics for bachelor and masters. It was not enough that I had done all my research projects in physics on computational and ML topics, and also that I did a computer science bachelor too, while studying physics. The problem was that I had not done any text processing before, just image processing. So, I was technically a newcomer to the field and had to learn NLP. It was quite amusing, actually—I ended up explaining the math and neural network bits to the more seasoned NLP students, because this was right when NLP was being revolutionized by transformer neural networks. Before, NLP leaned heavily on linguistics, but now it was all about neural algorithms which require quite a bit of linear algebra, calculus and probability theory knowledge, so things I knew from physics. As you can see from the linguistic knowledge that went into ChatGPT: NLP of today requires much less hard-core linguistics than it used to.

Then, all 5 years were a mix of doing research, teaching my own courses independently, and going to a conference each year after publishing a paper there.

My whole PhD – by choice – was quite unstructured, in the sense that I needed to write a research proposal, yes: but in the realm of AI and ML where everything moves so fast, I knew that from the proposal, only the broad interests and research questions will last, because the methods and the datasets will certainly change. And, they did change. When I started my PhD, image and language networks were CNNs and LSTMs concatenating their vectors. Then suddenly came transformer encoders, and now transformer decoders are all the craze.

Collaborations

Anyway, I did research; writing my own papers which I knew would be part of my PhD thesis, but also other papers in larger collaborations that helped build up my skills. For starting collaborations, it was important to talk to people at conferences and meet them. It does not help to be shy, even I know: it can be scary at times to approach a person that you know and admire because you read all their papers (I met Yann LeCun at the HLF 2022!).

But here's the thing: what's the worst that could happen? You might embarrass yourself, and they might ignore you. But guess what? They were already ignoring you because they didn’t know you existed! And the odds that the conversation will turn out all right and they will not end up hating you, is quite large.

Also, if you do not have the opportunity to meet someone in person, usually cold emailing works pretty well among researchers. Just bear in mind to email them about something they might find interesting. The worst that can happen is you not getting any response.

Teaching

I’ll talk about the research in a bit, but first I’ll tell you about the side-activity that came with my position at the institute, namely the teaching. I had the freedom (but also responsibility), to choose my own course and topics that I wanted to teach. So, I decided that in each winter term, I would teach a bachelor-level (undergrad) course about taking their first steps and designing their first experiments with simple machine learning methods. Since it was at the undergraduate level, I had to teach this course in German. Then, every summer term, I would teach a master-level course about deep learning methods that do not require annotated data but work unsupervised or with self-supervision. We started with PCA and clustering, and got into Variational Autoencoders, Generative Adversarial Networks, Transformers, State Space Models, Diffusion Models, and so on.

Teaching was a significant time commitment and one of the reasons my PhD took a bit over five years. With all the preparation and weekly classes, it could have easily turned into a time sink. But it taught me a crucial lesson that applies to YouTube as well: striving for perfection is a trap. The time needed to take a lecture or video from 80% to 100% perfect is disproportionately large and often not worth it. What does “100% perfect” even mean? I could create the perfect video for you, but it might not be perfect for the person next to you with a different background and different questions. So, I learned to let go of some of my perfectionism and actually get stuff done.

Now, why did I have to teach? In Germany, at least in STEM, PhD students typically receive a decent salary. Some students get funded by scholarships, which are prestigious but usually offer less money and benefits. Other students are paid through their professors’ grant money. I was an academic employee of the university, meaning I got a bit more money, but in return, I had to teach and perform some administrative duties for the institute. The big perk? Academic freedom to research any topic I wanted, unlike grant-funded students who are often tied to specific research topics.

Teaching turned out to be an invaluable experience, because it boosted my communication skills and it was actually the primer for me starting my YouTube channel, more on that a bit later in this video.

Thesis Writing

Okay, so, after 5 years of paper writing, teaching, and going to conferences from time to time, it was time to bring it to an end: writing the thesis. I managed to write it in two months. On the one hand, it felt natural, since paper writing and scientific writing is something that I had exercised already with almost 11 papers. But it was stressful, too, because I decided, perhaps a bit too late, to do extra experiments for each chapter to make them well-rounded. So, it was not only thesis writing that I was doing, but I was also running experiments and compiling them into a new paper and the thesis.

All in all, the lesson from this experience is: just start writing already! The hardest bit is putting a word on that blank document. Once that is done, the content will come together. Don’t forget to first decide on the structure you want to follow. And most importantly, keep writing even when you feel that the sentences and ideas are not perfect. You will improve them later in the next pass, because it is easier to edit content when all is put together, than it is to write something perfectly on the first try.

Despite the chaos of last-minute experiments, those two months of basically non-stop sitting at the computer were both stressful and peaceful. Peaceful because I could tell everyone I was in “deadline mode,” and they left me alone. I was fully in the “zone”. Plus, about one and a half months were during the semester break, so no teaching duties. Eventually, I handed in my thesis. A few months, some more teaching, and a bit of job interviewing later, it was time for the grand finale: my PhD defense.

PhD Defense ⚔️

The defense was a very comfortable discussion event that lasted around 75 minutes. I had to kick things off with a 15-minute presentation summarizing my entire thesis. So, the next time you find yourself grumbling about having to give a 10-minute presentation on a project you’ve worked on for a few weeks, spare a thought for me! I think keeping a presentation concise is a great skill to exercise, so I actually liked how this was organized.

My talk was followed by the Q and A part of the exam: a lively conversation with my PhD committee that was asking me question after question. I felt confident during the entire conversation, I was feeling like I had the answers. After all, it was about what I had done in the last 5 years of my life. All in all, the defense went incredibly well. My only regret is that the faculty rules meant seating was very limited, so not everyone could attend. Even worse, I wasn’t allowed to record or stream the event on YouTube.

But 75 minutes later, I emerged from the exam room victorious, and my colleagues were waiting outside to present me with the PhD hat which they had made for me. A hat? Yes, this is common practice in Germany, let me show you the hat.

PhD hat 🎓

In Germany, it is common to receive a hat after a defense that was assembled by fellow PhD students in the lab. Here, creativity is important, as it should contain themes and motifs about my PhD time and about the personality of the PhD student.

So, for me, no wonder that the top of the hat is basically a YouTube video with important content in it. On the bottom-left, like usually, sits Ms. Coffee bean! I ate quite a lot of Nutella during my PhD, so it ended up in the video, too.

The chapters of the video are marked by figures from my most important PhD papers. The title of the video is the title of the thesis. Usually, here is the channel logo, so they put a merch sticker I have. Instead of the subscribe button, there is the CITE button, which is funny. I love sailing, so there is a sailing boat on the top?

And maybe you are wondering what these partially deconstructed PhD hats do hanging around are? I used a lot of Shapley values in the methods of my thesis, and they are all about all possible combinations to ablate model inputs, so the little hats represent all possible combinations to assemble a whole PhD hat.

I find it quite inspiring and funny, I just love it! Thanks to my wonderful colleagues for making it!

Research Work

I would like to give you a short taste, of what my research was about. You might find it a bit overspecialized, since a PhD of course is very specialized. But it’s ok, I forgive you, if you use just skip to the next part where I’ll talk about the challenges I encountered during my time as a PhD student and about why I decided to start a YouTube channel. It’s all summarized in the video below 👇, starting 14:35.

Now, to complete it all and get the PhD diploma, I just need to publish my PhD thesis (as a book), and when it is done, I will share it with you online, so you can read it if you’re interested. Btw, the papers that went into my thesis are all open on my Google Scholar profile 🎓.

Challenges

Now, let’s talk about the challenges I encountered during my PhD. When I tell my story, it might sound like it was all smooth sailing, but trust me, there were some thorny moments.

Staying Focused

One of the biggest challenges with intellectual work like this is procrastination. There’s no one forcing you to work at any given moment, and deadlines often seem too far away to be intimidating. Despite this, I managed to stay focused most of the time.

A huge help was my supervisor's weekly meetings. Even if I had no progress to show, we would have relaxed discussions about the research topic. These meetings kept me grounded and focused, especially in the early stages. Over time, I became more independent and didn't need as much guidance. Instead, these sessions became more about getting a reality check—letting her review my thoughts and ideas and getting valuable feedback.

Getting Paper Reviews

Oh, I just mentioned “reviewing.” Because, yes, during a PhD, you're expected to review papers. It’s a skill you can pick up fairly quickly. But the real challenge for me was receiving reviews. It’s heartbreaking when your well-crafted work gets superficial feedback or when reviewers don’t bother to understand the concepts. I even got a review once that was clearly written by ChatGPT!

To handle this, I had to learn a few things about the review system and not take everything personally. Sure, my paper was a significant investment of time and knowledge, but I am more than that paper. Sometimes, a reviewer might reject your work for reasons as trivial as not agreeing with a widely accepted definition you used.

I also realized how random the reviewing process can be in a rapidly growing field filled with more inexperienced researchers than seasoned experts. Review season is always an emotional rollercoaster, but I’ve learned to manage it much better over time.

Getting Scooped

Another challenge in the ever-exploding field of ML was the constant risk of getting scooped—someone else doing the same research and publishing it just a bit faster. This happened with my first research paper: while I was wrapping up my project, I saw the exact same idea pop up on arXiv, but reading the paper and seeing the thought process and the implementation, I found it less deep than what I did. At the time, I didn’t handle it well. I heavily compared my work to theirs, showing how I was approaching the problem with more depth, but the comparison led reviewers to dismiss my paper as just an incremental improvement.

Looking back, I realize the better approach would have been to acknowledge the simultaneous development and stand my ground, instead of framing my work as an improvement on someone else’s, when it was actually developed independently and I did not start from their idea, but from my own perspective.

A Global Pandemic

Then, of course, there was the COVID pandemic, which threw a wrench into everything. For two years, I was working entirely from home, juggling administrative duties and adapting my teaching methods. This sudden shift skyrocketed the time spent on these “side quests,” making it hard to maintain a good work-life balance. However, once the initial chaos settled, I taught myself the necessary skills for the online world, including video lectures and livestreams.

And that, my friends, is what directly led me to start my YouTube channel!

Starting YouTube 📺

During the pandemic, teaching became a whole new ball game. Reliable internet for live lectures? Not exactly a given in Germany. So, I had to pre-record my lectures and then hold interactive sessions afterwards. For the first time, I found myself standing in front of a microphone and camera, trying to act natural. It was tough, but over time, I got better—though there’s always room for improvement. Being forced to do it was a great way to overcome my shyness and try something I never thought I could do. And, to my surprise, those videos weren’t half bad!

It turned out, it was not ok with everyone if I put those lectures into the open internet, so I decided to do shorter form content, explaining nuggets of insight about concepts and papers I’ve come across. I was not brave enough to show my face back then, so I quickly drew Ms. Coffee Bean here and voila. The first video was out. Growth was slow at first—YouTube’s paradox is that you need viewers to get more viewers. But eventually, my videos started gaining traction. The first lesson was to just keep making videos. Even if the first videos were mostly watched by my friends and family, the more videos there were, the more other people stumbled upon them. The second lesson was to actively promote the videos on other social media platforms. As I was creating videos about papers, I could mention the authors in my posts, and they would sometimes share the videos on their accounts. That lead to noticeable spikes in viewership. And as the number of videos on my channel grew, people started to stick around and subscribe. I’m still extremely grateful for every single one of you!

YouTube quickly became a wonderful source of moral support. Seeing all the positive reactions and comments, especially during the pandemic, was incredibly uplifting. And of course, this continues to this day. I read every single comment, and it means the world to me.

As the channel grew, it became my online business card, showcasing my expertise to a wider audience. This visibility led to exciting job offers and collaboration opportunities. Sponsors also started reaching out, and one of the toughest things I did during my PhD, was setting up a business in Germany to have the legal means to receive any money and haggle my first contract with a sponsor. I knew that I could never have done this part without my partner, to whom goes a big thank you!

So, what began as a necessity during the pandemic evolved into a powerful tool for personal and professional growth, opening doors I never anticipated.

Tip(s) for Current and Future PhD Students

When it comes to advice for future PhD students, I could list the usual tips you’ll find in any blog or PhD journey story, but I want to share something unique from my experience: the value of having an online presence.

In the past, business cards were exchanged to share contact details and profiles. Today, face-to-face meetings are less common, and I often find myself searching for a researcher's online presence. Too often, I find nothing beyond their papers. Just a couple of blog posts, recorded talks, or videos can serve as an excellent online business card. Why? Because these snippets give a glimpse into who you are and what you do. It is like having already done an interview process with somebody and then I know whether I find a person interesting to work with or not.

I’m not saying everyone should become a YouTuber and churn out over 100 videos like I did. No, just a small online presence—a couple of blog posts or videos—can make a significant difference. The leap from 0 to 2 pieces of content is much more impactful than from 2 to 100!

Reflections 🌇

Or at least, this is what helped me be more noticed, build my communication skills – apart from all other skills a PhD taught me, such as, critical thinking, research methodologies, scientific writing, resilience, and so on.

So, would I do a PhD again if I had the choice? Oh, absolutely!

In conclusion, my PhD journey was a challenging yet incredibly rewarding experience. It taught me more than I could have ever imagined and opened doors I never knew existed. If you're considering or currently pursuing a PhD, remember that it's a marathon, not a sprint. Embrace the challenges, cherish the learning, and don’t be afraid to put yourself out there.

Thank you for staying trough until the end of this special AI Coffee Break piece. If you have any questions, comments, or want to share your own PhD experiences, drop them below in the comment section.