Reflections for a gratifying Master’s in AI
I am a second-year Master student in Artificial Intelligence at Vrije Universiteit Amsterdam. Before, I did a CS bachelor in Portugal. My main goal with this post is to talk about some caveats on doing a MSc in AI.
It can benefit anyone interested in Artificial Intelligence (AI). More specifically, it is targeted at Bachelor students interested in studying within a university setting despite the majority of the advice applying to self-learning.
I attempt to summarize some of the most interesting lessons I take away from the last 2 years, exploring the AI field, with some advice in the end. As a disclaimer, it might be a bit more target for people interested in the technical and fundamental areas of Artificial Intelligence.
Many people want to apply Machine Learning to their area of studies or application. Often the logical thing is to consider doing a MSc in Artificial Intelligence. But is it really required?
Is this Master right for you?
In my opinion, it really depends on your aspirations and background. If you just want to apply AI models to a specific type of data, maybe you could start by doing the FastAI courses. Their approach is very practical and covers all the basics you need. The path of online self-learning is cheaper and works in the majority of the cases.
On the other hand, if you want to learn about the foundations of modern Artificial Intelligence and Machine Learning (ML). About how it connects to Cognitive science or how to apply it to cutting edge research, doing a Master’s in AI is a straightforward path. I and the majority of my classmates share this mindset. Personally, since I took a course in classical AI during my bachelors, I knew I wanted to learn more about the possibilities of combining AI / ML and science to help to advance our understanding of the Universe.
It is okay to have multiple ‘Aha’ moments
When you are introduced to a new topic in Artificial Intelligence, it is common to daydream about the capabilities of those tools. About how it can be combined with previous knowledge you have to solve open problems. That’s when an ‘Aha’ moment is likely to occur! It is exciting to dream about all the possibilities and how they can potentially revolutionize technology. Even though in the end you will likely not solve any major issues in AI, it is important to nurture these thoughts. Mostly because they keep your motivation high in order to explore and learn about new concepts. I certainly had some of these moments, and the first one was regarding combining Knowledge Reasoning with Machine Learning. It encapsulates what is known as Neural-Symbolic AI [1]. I even emailed a professor about the potential of these ideas.
The point is to try to practice thinking about how AI tools you learned can be applied to problems you care about. But do it critically since often there are some unforeseen obstacles at the beginning.
Being comfortable communicating is the only way
Whether you are interested in continuing in academia or getting an industry job, communication is a very marketable skill. This is something I realized during my master. I was more or less comfortable writing in programming languages and when struggling, StackOverflow would do the trick. The same was not true for mathematics where being comfortable translating your thinking in this universal language is key to deepening your “AI game”. Additionally, soft communication skills such as writing and speaking showed to have a greater impact than I could ever imagine.
The problem is that unless you find good reasons for how it can benefit you, you are likely to postpone practising them. It is often not easy but my approach is to focus on formalizing my opinion in questions I care about. Either in an opinion based format or explaining technical concepts to some audience.
Everything is Applied Mathematics
This should not be shocking to you but engineering is nothing more, nothing less than applications of mathematical principles. The mathematics behind AI is an important aspect and the one I’ve struggled with the most. It depends on your aspirations but to really understand AI you need to combine a lot of practice with real problems and an understanding of the tools you are using. For the latter, there is no other way than understanding the mathematics. Here I tried to apply the maxim “fake it until you make it”, but in maths, no such thing exists. I thought I could just jump into the master math level and it would work out. The closest analogy I find is to decide and go surfing on the Nazaré Supertubos without even being comfortable surfing with 2-meter waves … I guess you can predict the final result.
This is not to make you afraid, since I certainly learned and discovered more from courses I wasn’t prepared for or even failed during the master’s in AI. My recommendation, here, is to be honest with your level of “mathematical maturity” and start focusing on the gaps you have; it will most certainly pay off in the end.
Advice for newcomers to the field
- Start by reading a lot of posts, youtube talks, and books to have a rough idea of all the possible ways people are contributing to this field. Warning: This will take you a long time but don’t rush, it’s likely to be the most rewarding part.
- Watch Andrew Ng’s lecture on research, job opportunities and skills in AI
- If you haven’t done any academic writing or research, learn how to formalize experiments. This is important, more than your technical skills, it is the way you expose your work that matters.
- Once you figure out how you want to proceed in the field, start seeking ways to improve on the required skills. Normally, there is a considerable amount of online resources since AI is the hot topic right now.
- More broadly, take time to explore, try and make mistakes because, in the end, you decided to study to gain knowledge.
In the end, not being ready is normal. Don’t stress too much about it and remember that a master is an amazing time to explore your interests ;-)
Updated at 04.07.21 — thanks to Serafim, Peter and Luca for their feedback.
[1] For example, this paper https://arxiv.org/abs/1905.06088
