Yordan Darakchiev leads the only program in Bulgaria where, in just a year, you can gain enough knowledge to start working in artificial intelligence. Machine learning is a hot topic, specialists are in demand like hot bread, and yet this series of courses remains one of a kind.
Having completed the program myself, I can vouch that it not only teaches you the basics of this promising field but also makes you outgrow yourself. It convinces you that mathematics can be interesting. That machine learning and artificial intelligence are not magic spells but important terms that need to be discussed, and used responsibly and knowledgeably.
How did Yordan, an astrophysicist by profession, come to create this venture, which is now giving so many people the fundamentals for the fifth time? How do its graduates go on to research institutes, to companies like Tesla and Meta, to promising startups, and even to ventures of their own? What does it take to succeed in the world of machine learning? How does this kind of technology help us understand the world we live in?
Which directions in artificial intelligence are the most exciting ones to watch for news?
– How do you explain machine learning to your grandmother?
– I ask her… can we talk about this some other time? Will you make me some tea? But joking aside, the definition depends a lot on the audience.
Machine learning is a way to get a computer to do something through examples, much like the way people learn. That's the definition I give most often, but most people you meet on the street would hardly be interested in machine learning.
– But they would ask about artificial intelligence—we use that phrase every day now.
– AI is a somewhat different phrase, one loaded with lots of problems, marketing ones included. The fundamental questions are the hardest to answer because they are abstract.
AI is an imitation of the human brain on a chip.
How that imitation is made depends on many things, above all on the abilities of whoever creates it. You have AI when you play chess against your computer, but also when Spotify recommends a song to you.
– And in the end, AI today overlaps with machine learning, its most popular line of development.
– Every kind of AI does something similar. Machine learning is somehow the sexiest concept, but that doesn't mean the other ideas don't work. To build a successful product, you have to know what options you have. And here the technology works, sometimes even unbelievably well. Scientific papers have appeared with titles like “The Unreasonable Effectiveness of Machine Learning in…”
Machine learning has quite a few branches and is achieving serious success. Still, we should know that
the human factor is the most essential.
It is also the most problematic and raises the most questions. The interaction between people and what they have created, and what they use. That is precisely what most distinguishes machine learning from the other parts of AI. Otherwise it is just a fancy little word some people use to make more money.
– Which directions have surprised you with some huge breakthrough? For me, lately, it's definitely image generation—it's a whole paradigm when it comes to the way we create…
– For me personally, the most exciting thing has always been image recognition, computer vision. So the moment it turned out that advances in computing power could help it exist, around 2014, it developed enormously. Until then it had been practically unthinkable.
For me that was the biggest progress—can you tell whether there's a terrorist in this picture or not; whether there's a cat, it's the same task. Suddenly we got an explosion of algorithms in just months. Yes, there are always limitations, but they suddenly shoot up exponentially and you have exponentially more freedom.
There are many fields where things that were unthinkable are now possible. In astronomy, for example, there's predicting the trajectories of asteroids. In recent months, image generators really are a good example. There's a lot in the latest AI trends to impress us. For me personally, one of the most interesting algorithms recently was one for solving three-dimensional puzzles.
I think that if more people took an interest in AI, everyone would find something that interests and impresses them.
– As long as we learn to talk about the subject intelligently and look beneath the surface. To try to understand what this kind of technology gives us, rather than stopping at the Facebook-style “The Terminator is coming!”
– Those comments aren't groundless; there is scientific research dealing with exactly that part. And it matters, because
something your life depends on shouldn't be a black box.
On the other hand… really, let's not build a Terminator by accident, okay?
But yes, many topics would benefit from a little more intelligence being put into them. A little more thought and a little less opinion.
– For me the most dangerous part is still information, getting locked in bubbles. The fact that we've made recommendation algorithms too good. They dramatically distort the way we get information about the world and the way we communicate. Are there any applications of algorithms that could fix this?
– Is it the algorithms that create the bubble? Well, before them it was the newspapers. Our brain is built to choose lots of shortcuts. And it needs them, as an organ that works under an extremely heavy load. Its main job is not to work—to save as much as possible on what it has to do.
Bubbles are created out of sheer necessity.
Can we eliminate them? Hardly. Can we provide more information? It's already out there. Can we get people, ourselves included, to seek out more information and be critical? Yes.
Can we do that for everyone? Absolutely not.
Is it possible to build an algorithm that deals with these problems? Yes, of course.
Ever since the old foundations of programming, a program does what you tell it to. That is still true in machine learning, and here comes the delicate part: what will you tell the algorithm?
Will you end up regretting what you wished for?
– The series of courses you teach has remained, for five years now, the only one in Bulgaria that leaves you ready to work in artificial intelligence. How did that happen?
– There are quite a few courses, but they are short, a month or two each, and cover only some of the topics. It's true that, so far, there is no other complete series that lasts a year.
How did I start? One day I told my direct boss at SoftUni, a Bulgarian software training academy, Angel Georgiev, that I had been working professionally with data for three years. “Do you want to try a course like that?” His reaction was instant: “I don't know how much interest there'll be. Let's do a seminar and see.” More than 250 people came to a hall that held 200. Within a month or two we managed to organize a course; that was at the end of 2017.
Not only was there interest, it was also badly needed—more and more people were starting to talk about data. We try to keep up with the trends, and
I myself wanted to teach something I found interesting and deeply inspiring.
Back then I was happy that we were the only ones; now, when I hear it's still the case, I get a little uneasy. Partly because of the responsibility that falls on me, partly out of surprise—why are there still no others? I'd like there to be competition, for more and more people to take an interest.
The small detail is that the material is complex and not for everyone.
– That's true, I can personally confirm it. But then you have an undeniable talent for presenting it in an attractive, accessible way. Maybe demand in this field in Bulgaria simply isn't big enough yet? And at the same time it requires more study than, say, becoming a programmer.
– Becoming a programmer also takes a lot of study. And you have to develop every day, or you fall behind somewhere in the past.
Programming isn't easier and doesn't require less study. But the entry level you need to work with data is higher. It's harder to break in, but it pays off—where there's great risk, there's great reward.
The thing is, programming today is a little computer science and a lot more pounding on the keyboard and solving specific engineering problems. And those are more or less fixed; we have processes for them to which, by and large, we know the answers. In machine learning we also have processes, but also scientific questions whose answers are unknown. Things where it's unclear how long they'll take. How much data do you need to build a self-driving car? How much of it do you need to keep it from hitting someone on the road? It turns out it's never enough; it will always find a way to crash.
That's the more complicated part—the very essence is often more abstract, more theoretical. I'm not saying it's harder than traditional programming. It's good for more people to give it a try, because it's interesting. And because it's worth it!
– What kind of people have surprised you as part of the course?
– People like you! But I have detailed statistics, especially for those who have taken exams. There are two main groups, roughly equal in size—people with no idea about math who are relatively good programmers, and the reverse: people who know math, statisticians with five, ten, fifteen years of experience, touching programming for the first time.
Of course, a large share are complete beginners, and my program is aimed mainly at them. But it turned out that everyone can take something away. The vast majority are interested in one way or another in what machine learning can do. They've heard of it, they know
how wonderful the idea is of training a program without telling it exactly what to do.
They want to try it. And not for the money—I know how dangerous it is to set off down that road. The main thing that drives people is curiosity about what machine learning actually is.
– Do you have impressive examples of people who've gone far after starting with the course?
– It depends on what impresses you. Some are already working at Tesla, at Meta, at Microsoft Research. Big company or small, it doesn't matter; whether you start at Microsoft or at a startup is more a question of desire than of opportunity. But I find it interesting that quite a few people have gone on to master's and doctoral programs and do more scientific research, at top universities at that.
It's very pleasant to see that the people who enroll in the program have all sorts of interests, and that I can somehow help most of them go through their own adventure, do what will interest them later on. I'm glad I manage not to push them into a rut, which often happens in beginner courses. They tell you, “This is the way,” and you copy it for the rest of your life. What if the lecturer was wrong?
My courses have two main goals. First, to get people to think critically and to doubt many things, above all themselves, however destructive that can sometimes be. And second:
to unleash the creativity within them.
Not to be afraid to take from the course what they like, what is their own adventure. I know it sounds very strange, philosophical, abstract, but for me it is very concrete—thinking and creativity.
– How did you end up in this field? Did your interest in it come from astronomy?
– As a kid I spent a lot of time in the municipal library in Kotel, a small town in the Balkan Mountains. I loved reading whatever I could get my hands on. I was interested in computers; I'd been a fan of games since I was four or five. I came across a book on programming. I didn't yet know what that meant, but I saw the basics of BASIC. In another book I read that
one of the most interesting things is neural networks, and that artificial intelligence will one day conquer the world.
It literally said so in the little book! I found it fascinating, it was very inspiring. I didn't yet know about The Matrix or the science fiction writers who dealt with artificial intelligence in some way and told stories of what might happen, the good sides and the bad.
Many years later, I'm finishing my first year at university, studying physics, and a friend asks me whether I can explain some math to him. I take a look, see some slide from a presentation, and answer: “That's really simple, I learned it in ninth grade, I'll explain it to you.”
Twenty minutes later he knew what linear regression was, and I got curious where that slide came from. He soon sent me the presentation; it turned out to be from Andrew Ng's course on machine learning. I took the course, got really interested, and that's how things started, almost by accident. Then I saw that my knowledge of working with scientific data, of the scientific method, was very valuable, and that I could combine my two passions—programming and astronomy.
Which came first? I don't know; somehow both have been with me my whole life. I don't remember when I got hooked on astronomy either. I know that at three or four I was already interested in the sky, and by third grade I was correcting my teacher in science class.
– What are the most exciting ways machine learning can help astronomy?
– There are many, starting with the most standard application—automating tasks that are very hard and slow for humans.
One direction is searching for new objects that haven't been discovered yet. Knowing about every single thing in the sky brings us some kind of concrete benefit. Some objects help us determine distances, some stars let us figure out whether the planets around them could be habitable, nebulae tell us the history of the star that came before them.
We take pictures. Just like in CSI, we look for ways to make them sharper, to see better. There's software there too, and artificial intelligence that can help us enormously.
That's how, for example, we see images of a black hole—something almost impossible to capture directly, which now happens through the magic of data processing and machine learning.
There are many benefits—from predicting solar activity to all kinds of protection for Earth, from asteroids, for instance. The applications are so numerous that I find it hard to list them.
– You're now a PhD student in astronomy. What direction is your research taking?
– It's about machine learning in astronomy. I'm working on the search for objects not yet discovered—mainly periodic variable stars. The main type I deal with, Cepheids, are very good distance markers. Based on them, we can calculate quite precisely how far away various galaxies are. The main goal of my doctorate is to understand where and how stars are positioned relative to the cosmic dust that absorbs light.
Each new discovery brings many benefits, and the idea is to automate this as much as possible. It would take an astronomer a long time; it's complex, specialized work that also depends on the astronomer's own knowledge.
– What are the most exciting breakthroughs you expect astronomy and astrophysics to achieve with the help of machine learning?
– What would excite me most is if one day
I saw a headline saying they've discovered extraterrestrial life.
And I'm absolutely sure it will look very dry and boring. The scientific paper won't be “We Found Life on Planet X” with lots of exclamation marks at the end, but something along the lines of “Indications of the Presence of Biologically Synthesized DNA on Planet X at Confidence Interval Y.” But yes, finding traces of an existing alien civilization would be the culmination of a great, great deal of research.
