I quit Computer Engineering in the University of Sherbrooke during COVID, so I'm lucky enough to not have graduated young. Why am I lucky? Because with experience, I was able to choose my field more carefully. People who know me know that I have a broad range of interests in computer science. But I chose Machine Learning.

Here are the reasons. Keep in mind that this isn't coming from someone with 3 PhDs who has read all of the recent publication in computer science journals. This is coming from someone who has fiddled with their computer quite a bit and who is nearly done with a bachelor's degree.

Machine Learning and Data Science are emergent fields of research

My goal was always to do research. I want to be one of those scientists in a laboratory surrounded by PhDs and focus my time on a single, very hard problem, as opposed to the industry where many small problems have to be fixed. I had never had a clue what to do research about. Seeing as how Machine Learning scientists publish many articles and are among the most cited these days, just from this, Machine Learning is a fairly obvious field of research. But this is not enough to warrant dedicating my life to it.

Most problems in Computer Science are already mostly solved

Compilers were the first hard problem to be considered solved. I also consider operating systems to be solved. They tried to build Plan9 as a successor to UNIX, but it never caught on. Instead, taking UNIX and building on top of it seemed to work well enough, the base architecture is more than good enough for our world. Every aspect of operating systems is nearly solved and the amount of experts is diminishing. File systems? With ZFS, you don't need to worry about losing data, about reaching size limits, about RAID. Most of the development is in optimizing the algorithms and adapting it to modern day challenges. Networking? Between TCP/IP and more performant proprietary protocols built on top of UDP, we can consider this solved. The only challenges basically remain electrical engineering ones, keeping up with the exponential growth of data in the networks. Memory management? From the time it was separated into pages it's been good enough, the remaining work was on how to handle and restrict these pages.

Algorithms have been made to solve a wide variety of issues and are divided into classes. Techniques exist to make them faster. State of the art algorithms are built upon existing algorithms and do low-level trickery to reduce the exponent of a polynomial by 0.05 if we're lucky. Conceiving new algorithms is basically just playing legos with existing algorithms and then applying well-known optimization techniques.

The remaining fields, for me, are bioinformatics, information security and, you guessed it, machine learning. Many of the advancements in bioinformatics are related to machine learning.

Why not information security?

I don't know much about this. For me, keeping my software and libraries up to date, using TLS for everything and using GPG on my emails is good enough. I know how to configure a firewall, and it seems like the best thing to do is give others the least amount of access possible. Not a very interesting subject in my opinion.

What is an interesting subject of research in information security is cryptography. But I see it as one big can of worms full of headaches.

Machine Learning is fun!

Machine Learning has given me a lot of headaches. But it's also the most fun I've had with a computer since I've discovered UNIX. It's a weird child of probability theory and linear algebra. What's fun about it? It's very hard to get a grip on how neural networks work! You build a base architecture by using what you know works well for this specific problem, train the model, see how it behaves, whether its loss goes down and by how much, and if that doesn't work you try something else. You may have to do research to find why it doesn't work. And if it does work, you try to make it better, and you may have to do research to make it better. The end result is a machine that can tell you which letter is written on this piece of paper.

With a more advanced model, you can do more advanced things. With things like a recurrent neural network and/or a transformer architecture, mixing it with already built models, you can have it read to you the text that is written in cursive on this piece of paper. How cool is that?

Machine Learning is the future

It's in constant development. Soon we may see models that learn to behave like humans. I don't think this is possible with large language models as we know them, but if we keep going it will become possible. Machine Learning engineers are currently trying to rebuild human brains in binary. It's no different from playing God. Now we're far from consciousness, but if we can have something like an LLM that knows that it's wrong when it is wrong, mixed in with reinforcement learning based on the emotion of the responses of its user, we'll have walked a long way.

This is something I want a part in.

The pitfalls

Machine Learning is like bruteforcing intelligence. Machine Learning costs a lot of energy. Machine Learning is making hardware costs spike because they're being reserved by big data centers. While these are big problems, as someone interested in research, I see them as solutions waiting to be found.

Remember the size and energy costs of the average computer 50 years ago. I don't, but you might. In Computer Science, efficiency is a sign of performance.