What do Machine Learning Scientists do?
What do Machine Learning Scientists do?
  • Hyokun Yun (Machine Learning Scientist)
  • 승인 2021.02.27 23:42
  • 댓글 0
이 기사를 공유합니다

(Disclaimer: This article does not reflect the opinion of my employers nor my company.)

I am a Machine Learning Scientist at Amazon. When I was an undergraduate student at POSTECH, I was imaginative about what my future would be: I considered becoming not only a college professor or a software engineer but also a business consultant and even a sound engineer. Yet I didn’t really imagine much of what I have actually become today, for better or for worse. To the best of my knowledge, my job title—“Machine Learning Scientist”—did not even exist ten years ago, so it could not have been my dream job back then for sure. I wanted to talk about this peculiar-named job I ended up taking, hoping this would broaden today’s Postechians’ considerations of their future careers.
I entered graduate school in 2009, and I chose machine learning as my field of research. At the time, it was quite an effort for me to explain to my friends what my research was about. I used to say “it is an area of artificial intelligence which focuses on leveraging the large amount of data we can easily collect in today’s world, rather than relying on input from human experts…” and I could already see it in their eyes that I had lost them. I readily accepted this as an inevitable challenge for any researcher though, since almost every academic friend of mine seemed to share the same difficulty. By the time I finished my Ph.D. defense in 2014, however, machine learning suddenly became a buzzword everyone knew and was excited about.
Many factors contribute to the high public interest in machine learning, but I believe the main driver is the surge in demand from businesses. The initial successes of machine learning applications to business problems came from tech companies like Google, Amazon, and Netflix. In applications such as digital advertising and recommender systems, more precise predictions of people’s preferences provided a more convenient customer experience, which helped with these companies’ development. Today, however, machine learning is becoming instrumental in almost all types of industry; it powers not only financial trading and robot manufacturing but also the discovery of new medicine and the creation of artworks. Even in our daily lives, many people utilize machine learning powered voice assistants to conveniently control lighting and appliances in their homes.
Even at this very moment, machine learning technology is still rapidly evolving, and new business opportunities are constantly opening up. As a result, a lot of business problems are at the state where people in the problem domain vaguely understand the potential opportunity to leverage machine learning, yet understanding exactly how machine learning can be used entails a high degree of scientific uncertainty. This is where Machine Learning Scientists like myself come in. The title of the job is not very well-established, and different companies put different names to it. Still, the expectation about the job is gaining consensus: we understand the business context, translate vague business requirements into scientific formulations for machine learning, and work with software engineers to build a software system that trains and deploys machine learning models, which automate certain decision makings within the business.
Technical problems we have to solve in this process are often scientifically challenging to the extent that we publish findings from our work to academic conferences and journals. For this reason, Machine Learning Scientists tend to hold a doctoral degree, although there are still plenty of successful people without the degree. However, I believe there is a crucial difference between “industrial” research we Machine Learning Scientists do and “academic” research done in research institutions. While both kinds of research never proceed as planned—otherwise, we wouldn’t call it “research”—academic research enjoys a greater degree of flexibility in setting up objectives. When I was a graduate student, I often found myself in situations where the technology I developed for a certain problem would not work so well on that problem but well on other important problems. This was usually not bad news, because I could still publish the finding, and the science community gained useful knowledge from it. However, in industrial research, the problem I need to solve is anchored. At the end of a prosperous and exciting journey of scientific research, I need to tell my colleagues the solution to the problem they are waiting for.
The job is stressful. It is human nature to avoid risk, but every project of mine is inherently risky; if I already knew the right machine learning approach to the problem, software engineers would work more efficiently on implementing it instead of me. And as my career progresses, my colleagues expect me to deal with an increasing level of uncertainty. Nonetheless, I like my job. It is quite rewarding to see my work being built into large-scale systems and making sizable differences in the world. Since machine learning technology is rapidly evolving, I need to keep myself up-to-date and constantly rethink my approach, but I take this as a refreshing challenge. Furthermore, the job is relatively high-paying for a salaried job, which I did not appreciate enough before I became a parent.
Therefore, it was quite an opportune decision for me to start machine learning research in 2009. My generation of machine learning researchers has been amply rewarded from the sudden, unexpected increase in demand from the job market. I could have attributed this to my foresight and brag about it, but, frankly speaking, I didn’t see this coming. My choice of research area was just a consequence of a series of capricious, ill-informed decisions, each of which looked just reasonable. I also want to tell you that none of the renowned researchers I have met told me they saw this coming either, or at least when exactly it was going to happen. We were all just lucky. I want to stress this, because we often give too much credit to lucky people’s judgments in hindsight and belittle our own judgements. We need to constantly make career decisions in this fast-moving world, so let us not fret too much about what is not under our control, just as we do not worry about our faculty to pick good lottery numbers.

 

Hyokun YunMachine Learning Scientist
Hyokun Yun
Machine Learning Scientist